Subscribe now

The Token Refinery Value Capture

I was sitting outside La Boulangerie in Hayes Valley two Fridays ago, by the park benches, after a long week of back-to-back meetings, demos, and chats about AI-everything, and the semi-stressful life of walking around downtown and looking over my shoulder before sunrise.

Six days of intense AI-ing and no care for any jet lag, I was finally relaxed, looking around. Everyone was sunbathing and seemingly enjoying one of those perfect California afternoons — clear sky, warm sun, the kind of day that finally makes you understand why people still put up with the rent and chaos.

Just before that, I was having coffee with a VC who has been investing in ‘national defense’ tech, a hot sector that seemingly popped up in recent years amid the back-and-forth of the trade war and the rise of China’s soft influence globally. The conversation kept circling back to one word: winning. “So who’s going to win?” “Do you think they’ll win?” “We must win.”

Who’s winning the model race? Whether Anthropic is winning against OpenAI. Whether open-source is winning. Who’s winning at inference cost? Is China winning? Who’s winning in China? Will China win? In what way and why will they win (or not)?

The word kept coming up, and it threw me off a little. Not because it was entirely wrong, but because it felt a bit narrow. On a business level, the AI market has moved past the point where “who has the smartest model” is the most useful question. The more interesting question now is where value is actually being captured, and where it is already being competed away. That is what brought me back to an analogy I heard earlier that week from another VC investor: AI is starting to look a lot like the oil business. And the ‘four-year Albertan’ in me could not resist that one.

Tokens are crude oil

Models produce raw intelligence. They generate tokens. But tokens by themselves are not the end product that any (most) customer actually wants. What customers pay for is legal work completed, code shipped, claims processed, research synthesized, and decisions supported. They pay for refined output.

A lot of investor attention has naturally gone to the infrastructure layer. But once the model is understood as an intermediate good rather than the final product, the analytical center of gravity shifts. The question is no longer just who can produce intelligence, but who can turn it into something usable, trusted, repeatable, and economically legible. In other words, who can refine it?

At the bottom of the chain are the token producers: the frontier labs and open-weight model builders – think OpenAI, Anthropic, Google DeepMind, Meta, DeepSeek, Qwen. They produce the raw capability. This layer is expensive to build, technically impressive, and still moving fast. But the model is not the final product, any more than crude oil is gasoline. And what most enterprises/consumers are paying for is gasoline.

Of course, the first form of refinement occurs within the labs themselves, where raw model capabilities are turned into products like ChatGPT, Claude, and Gemini. This is already a real business. It may not yet fully fund the drilling, but it is clearly one viable business model. The labs are packaging intelligence into interfaces, workflows, and, increasingly, work products that users consume directly. This is why the “labs will be commoditized immediately” argument has always felt too simplistic. The best labs are not only producing crude; they are trying to capture the downstream markup themselves.

External third party refinement occurs when AI is embedded in specific workflows, industries, and customer relationships. Harvey is doing that in legal. We see AI inside defense software. AI inside clinical workflows, compliance, finance back offices, customer support, and vertical SaaS. This layer takes cheap model output and turns it into something more valuable by leveraging better context, process, service, trust, industry know-how, and distribution. That’s why software cannot be just all “vibe coded”.

This model also most closely resembles where software value was historically captured. AWS built many important first-party products, but the cloud era’s great value creators were often third-party companies, such as Datadog, ServiceNow, and Workday, that sat atop shared infrastructure and added enough domain value to justify an independent, premium-priced existence. That pattern now looks increasingly relevant to AI as well.

A model can answer a question. A product can route the answer, compare it with prior cases, attach it to a record, push it to the right person, maintain an audit trail, and make it fit into an institution’s existing way of working. In the coming year, the focus is shifting away from AGI, at least for businesses; it is the conversion of raw intelligence into finished economic output where much of the real business value lies.

The spread is the business

The easiest way to understand any AI company is to ask one question: how wide is the spread between its token input cost and the value of the output it delivers?

In much of today's knowledge work, service professionals still charge based on labor scarcity, credentialing, and process complexity. AI changes the cost structure underneath that. If a model can perform meaningful parts of that work at very low marginal cost, the battle shifts to who captures the markup between cheap intelligence and high-value delivered outcome.

But what protects the spread from collapsing? So I think there are a few factors to examine that can help separate genuinely durable businesses from products that merely look exciting in the current moment.

The first is workflow ownership. If the AI product is deeply embedded in a complex, vertical specific, high-consequence workflow, it becomes much harder to replace with a generic model plus a thin wrapper. Customers don't just buy the output; they buy the reliability that it appears in the right place, in the right format, connected to the right systems, with the right approvals and audit trails. So products have to have that extra moat. Once a product becomes part of the enterprise's operating system or process, then the competitive question changes. It is no longer simply whether another model is cheaper or smarter. It is whether the entire surrounding workflow can be rewired without cost, risk, or disruption, which usually it cannot.

The second is accumulated context. Clever prompting at this point gets absorbed into the base model's capabilities surprisingly quick (they suck that data in like Kirby). Real accumulated context: customer records, prior work product, compliance history, integration into adjacent systems, operational memory. The more useful the product becomes because it sits inside a growing body of context, the less interchangeable it becomes. From closely following the Chinese AI market, one pattern is clear: anything engineering-related — tooling, prompts, harnesses — is usually not sustainable for very long. Non-engineering advantages like domain expertise, user data, distribution, and user habits are more durable. The things people can demo most easily are often the things that travel fastest across the market. The things that stay sticky are usually quieter. They sit in behavior, habit, data exhaust, and institutional memory.

The third is captive customers and limited competition. So this is why the viral X post is relevant. If 50 startups are all refining tokens into generic customer support chatbots, the spread collapses quickly. If you are one of a small number of companies trusted to operate in medicine, national defense, or other regulated environments, the spread can hold for longer because access itself is scarce. The opportunity to arbitrage the gap between low token input cost and high output value is simply too attractive — which is why so many refineries are gravitating toward the high-value sectors first. Over time, competition narrows the gap, but if there is structural scarcity of competitors — regulatory barriers, security clearances, deep institutional trust — the spread persists. In those cases, what is scarce is not intelligence itself, but goes back to permission, trust, and workflow access.

This is why the word “wrapper” has become too dismissive to be analytically useful. There is a massive difference between a generic chatbot product and a company embedded in a regulated workflow with years of context and trust built into the product. This is also why the language has started to shift from “wrapper” to “harness.” The market is already trying to distinguish between a thin interface and something more deeply embedded.

The next phase of the market will probably focus on categorizing thin refinement versus deep refinement. One is easier to copy because it is mostly a presentation layer. The second kind is harder to copy because it is intertwined with the base infrastructure and processes.

China already shows what happens when crude oil gets cheap

In the U.S., the value chain still appears to have two visibly distinct profit models. There is one at the model layer because frontier labs still command attention, capital, and, in many cases, a real commercial premium. And then there is one at the application layer, where companies try to build products on top of that intelligence, whether it is in the form of a wrapper or a harness.

In China, that picture already looks slightly different. Strong open-weight models have become widely available, and the economics of access have moved lower much faster. That makes it harder to build the entire investment case around proprietary intelligence alone - models are the most competitive if anything. This does not mean the model layer disappears, nor that model quality ceases to matter. It means the center of gravity shifts. More of the competition, more of the product differentiation, and more of the monetization pressure have moved downstream.

My point is that China is not just another market. It may be an early preview of what happens when model access becomes cheaper, more abundant, and less differentiating. The logic of competition becomes easier to see there because the layers have already compressed. To borrow from my earlier writing on LLM business models, I once argued that the dish is only as good as the quality of the fish. That may still be true for sashimi, maybe ceviche too. But here I have to counter my own analogy a bit: sometimes the quality of the cooking matters more than the fish’s rarity.

Because this in turn forces companies to differentiate elsewhere: distribution, product design, speed of iteration, vertical depth, ecosystem leverage, and the ability to build a product that feels native to how users actually behave. It forces them, in other words, to refine better. And we’ve written much about this- see AI Proem’s previous coverage.

And one thing that strikes me, each time I speak to investors in the Bay Area, is how poorly understood this still is. Not because people are unintelligent. Far from it. But because too much of the U.S. discourse still frames Chinese AI through a geopolitical or ideological lens rather than considering market structure and technological development. So you get questions about whether open-source adoption in China is mainly state-led, or whether local edge deployment is mostly driven by paranoia about cloud privacy. Those questions themselves show how far the framing can be from the actual reality on the ground. China started embracing open-weight AI for the same reasons markets adopt cheaper, workable inputs everywhere: it is effective, it reduces costs, and fierce downstream competition forces product companies to move fast. As we have written repeatedly here at AI Proem, it started as an economic one.

Why it matters for global AI

So sharing three observations from the Chinese market that I feel are worth internalizing for investors:

• First, commoditized crude does not kill the refinery business. If anything, it can expand it. Lower input costs and lower barriers to entry create more experimentation, more verticalization, and more attempts to package intelligence into different end uses.

• Second, the winning refineries increasingly differentiate on distribution, product sense, and contextual fit rather than on raw model intelligence. That is not some uniquely Chinese curiosity. It is a likely preview of what happens in any market once the question “who has the smartest model?” becomes less decisive than the question “who has the most economically useful product?”

• Third, spreads become thinner. Intense competition compresses the gap between token input cost and final selling price. That means Chinese refineries often have to run leaner, move faster, and iterate with more urgency. If open-source continues closing the gap globally, more of the AI world may converge toward that structure. The standalone profit pool at the model layer may not disappear, but it does get pressured. And once that happens, the central investment question shifts quite dramatically. One stops asking primarily who owns the best model and starts asking who owns the workflow in which the model becomes economically indispensable.

Some (especially public) investors I spoke to in San Francisco were not unaware of this risk. But many of them felt oddly complacent about it, as if the direction of travel was obvious but somehow still dismissible. The defenses all sounded familiar: the frontier labs will maintain a lead; the market is still so early that commoditization does not matter yet; if the major labs eventually monetize at a huge scale, who cares what the end state looks like; the U.S. government will ultimately step in and slow distillation. Perhaps some of those things will prove partly true?

In commodity markets, analysts obsess over the marginal producer because that is what sets the price. Something similar is happening here. China deserves close study, not as an exotic side case, but because it may be showing what happens when the marginal cost of intelligence falls faster than the narrative can comfortably absorb.

Connecting Thompson and Evans

This also connects, at least loosely, to frameworks Ben Thompson and Benedict Evans (my OGs) have been writing about recently. Seen this way, the current debate among AI commentators looks less contradictory than incomplete.

In Stratechery, Ben Thompson has been right to emphasize that the model makers may be better positioned than many expected because they increasingly own first-party products, not just raw model access. If the labs can package intelligence directly into useful products, then more value can flow back to them than the “all labs become commodity utilities” camp assumes. That is directionally consistent with his recent writing on software and AI, even if the exact end state remains open.

The framework captures why first-party refinement matters, but not fully why third-party refinement can still become enormously valuable. The history of technology stacks suggests that foundational platforms often push downward and upward at the same time, yet still leave large spaces in which third-party companies create durable value by owning particular workflows, categories, and customer relationships. The existence of a powerful platform does not eliminate the need for downstream specialists; it merely raises the bar for which specialists/verticals get to survive.

Benedict Evans, by contrast, has made one of the cleanest bear cases on OpenAI specifically. His argument is that labs like OpenAI do not clearly have a unique technology or product moat, that user engagement appears shallower and more fragile than current perception suggests, and that incumbents have broadly matched capabilities while bringing stronger product and distribution advantages to the fight. And we’ve kind of seen that disruption happen so quickly with Claude.

This relates to my point that the internal-lab-refinery may not have enough context, workflow depth, or defensibility to justify the strategic centrality people have projected onto it.

Thus, putting the two analyses together, I was inspired. Thompson is right that the model-maker advantage is real, especially in first-party refinement. Evans is right that this does not automatically create an impregnable moat at the product level. The synthesis, then, is that model advantage matters, but only up to a point. Beyond that point, what determines value capture is whether the product in question has sufficient workflow, context, and distribution to sustain the spread. So it goes back to my point about third-party refineries.

The frontier labs remain the hardest call. They are clearly extraordinary businesses in many respects, and their pace of productization has already surpassed many skeptics' expectations. But the question at very large valuations is not whether they can continue on this impressive ARR growth trajectory.

Can the economics they enjoy today remain durable as model capabilities diffuse further and downstream competition matures? To underwrite those valuations, one has to believe not only that crude production remains meaningfully profitable, but that first-party refinement becomes sufficiently central to capture a large and lasting share of the downstream economics. Perhaps that happens. Perhaps the strongest labs evolve into more like full-stack platforms, fostering ecosystems atop their intelligence while also capturing direct end-user demand. That would likely be the most durable version of the story. But API access alone will likely be insufficient, and first-party chat products may be insufficient as well if the rest of the market becomes more efficient at refinement than current enthusiasm assumes.

One broader thought underlies all of this: the pattern of upstream capability commoditizing while downstream refinement captures value is not unique to AI. One sees versions of it across many waves of technology. In the cloud, raw infrastructure became foundational, but much of the most durable value was captured by the software companies that turned that infrastructure into category-defining systems of work. In mobile, the platforms and devices mattered enormously, but so did the companies that transformed mobile distribution into new end markets and behaviors.

In AI, tokens are the crude. But the more enduring question is the same as it has always been in these technology shifts: where does the markup actually live once the raw input becomes cheaper, and what keeps that markup from being competed away?

So, let me ask you again, ‘who’s winning?’ or ‘who will win?’

We all know this saying, which originated in the ad industry. But now, the DeepSeek team has made it the brain prompt of the month for AI developers.

DeepSeek dropped an OCR paper today, and what jumped out to me is their extreme creativity and ability to think outside the box. DeepSeek’s team first showed its engineering ingenuity in its R1 model and has now turned “a picture is worth a thousand words” into a literal technological breakthrough.

From Moonshot’s team’s Pink Floyd rock band devotees to DeepSeek’s philosophers. Who said they’re all just a bunch of nerds?

How it works

In plain English: DeepSeek-OCR [Optical Character Recognition system] treats a page image like a super-compressed “zip file” for text. Instead of keeping thousands of text tokens in memory, the model turns the image into a small set of vision tokens, and when you need the content back, it “reads” those tokens into words. The paper reports roughly ~97% recovery at ~10× compression, and still around ~60% at ~20×, which is a big deal for cost and speed.

Think of a memory from twenty years ago. You have memories of you enjoying that holiday by the beach with your parents, but you’re not sure what the flavor of your ice cream cone was anymore. Details are fuzzy, but with focus, you can also reconstruct a lot based on some reverse logical thinking, and things become much more vivid. This behaves similarly to what we’re talking about here; you can keep a compact “picture memory,” work out 97% of what it was, and then reconstruct the details with some tricks and prompting, maybe. It’s a clever, very practical way to preserve brain real estate.

How it’s built

So their innovation is two parts: a DeepEncoder that turns a page into a tiny number of vision tokens (64/100/256/400, plus a tiled “Gundam” mode for dense pages), and a lightweight Mixture-of-Experts decoder that turns those tokens back into text (only ~570M parameters “wake up” each step, so it runs like a small model). The encoder is like a smart zipper: it looks locally, shrinks tokens ~16× before doing the global pass, and lets you pick the cheapest setting that still reads cleanly. When you need the content later, the decoder reads the tokens back into words—simple.

Why this matters

This unlocks much cheaper memory for long contexts such as old chat history or big PDFs. Store images/vision tokens instead of mountains of text tokens, then recover the text when you actually need it. They even show a “gradual forgetting” trick: keep older pages at lower resolution to save more tokens. And the throughput isn’t hand-wavy, hundreds of thousands of pages per day on a single A100-40G—so this feels deployable, not just academic. If this can scale, then it could redefine how LLMs handle context and, more importantly, cost.

Bigger picture

As Andrej Karpathy has pointed out, pure text can potentially be a wasteful interface for models. Multimodal approaches like this hint at a future where images (and other modalities) could carry context more efficiently, and even Elon Musk chimed in on this thought. So maybe it’s only a matter of time before LLM inputs expand far beyond text.

Related material:

• DeepSeek’s Ascent Reshapes China’s Burgeoning AI App Landscape

• DeepSeek’s Open Source Week: Sharing the Future of AI Efficiency

• DeepSeek V3 puts China AI on the global map: consumer use and capital expenditure implications

• Everybody Losing Sleep Over DeepSeek: Industry Implications to LLMs and AI Infrastructure

• The Jevons Paradox in AI Infrastructure: DeepSeek Efficiency Breakthroughs to Drive Energy Demand

Well, today, I want to loop back to the relationship between AI, Data Centers, and Energy and provide an update on that front.

Subscribe now

See some relevant readings here:

• AI Arms Race Far From Over: Chips is Only Half the Game, Infrastructure is the Other

• Trump 2.0: Infrastructure Investment, Traditional Energy Resurgence, and Foreign Capital

• Converting Crypto Mining Centers for AI Data Centers?

• A Whole (New) Nuclear World

• Why Data Centers Can’t Go Full Renewable—Yet

• The Jevons Paradox in AI Infrastructure: DeepSeek Efficiency Breakthroughs to Drive Energy Demand

• Johor, Malaysia: AI Data Centers Drive Investment, But Infrastructure Challenges Persist

• Canada Should Prioritize Growth and Position Itself as an AI Infrastructure Leader

Power Demand Continues to Increase

As we know, a big part of AI development right now is dependent on electricity. The race to build ever-larger data centers is reshaping national power grids, driving record energy demand in both the U.S. and China.

Reuters reported today that U.S. power consumption will hit record highs in 2025 and 2026, the Energy Information Administration said in its short-term energy outlook. And we’ve seen the Trump administration greenlighting power plants and encouraging energy capacity buildup at a faster rate than the country has seen in decades.

The EIA projected power demand will rise to 4,187 TWh in 2025 and 4,305 TWh in 2026, up from a record 4,097 TWh in 2024. This surge is largely driven by cryptocurrency and AI demand.

As written before, in the energy race, China has been far ahead with decades of strategic planning in building out renewable energy plants, from hydro to solar to wind farms, across the nation. But now American corporations are catching on and are also trying to capitalize on the AI boom as well as provide the necessary power to supercharge the data centers filled with GPUs.

Goldman Sachs Research forecasts that data centers will increase their share of global power demand from 1–2% in 2023 to 3–4% by 2030. In the U.S., that share could more than double, from 4% in 2023 to ~9% by 2030.

“If global data center growth in 2030 vs. 2023 levels were its own country, it would be a top 10 global power consumer,” according to Goldman Sachs Research analysts Jim Schneider, Carly Davenport, and Brian Singer.

Furthermore, based on Goldman Sachs’ research, the current demand for energy in the U.S. is 62 gigawatts (GW), split across cloud workloads (~58%), traditional workloads (~31%), and AI (~13%). AI is projected to grow to 28% of the total by 2027, while cloud drops to 50% and traditional workloads fall to 21%.

Meanwhile, the IEA projects global data-center electricity use to ~945 TWh by 2030, doubling from 2024, with AI-optimized facilities as the dominant driver. That’s nearly 3% of global electricity in their base case. And the geographic breakdown of electricity used by data centers in 2024, with the U.S. accounting for 45% of the usage, China 25%, and Europe 15%.

Remaining at the lead are the U.S. and China. In the U.S., demand is clustered in Northern Virginia and Arizona, driving up bills and environmental concerns. In China, it’s concentrated around Beijing, Shanghai, and satellite economic zones, where the government has steered businesses over the last decade.

, a China energy expert at the Lantau Group, noted on X: “According to China's renewable consumption quota policy, all new data centers in these hub regions must buy at least 80% of their power from renewable sources.”

He highlighted that this should be feasible in “blue hubs” rich in renewables, but more difficult in “red hubs.” [See graph below]

Adding that, outlooks vary by projections/ province, but forecasts suggest 60–70% of China’s new compute will land in these hub regions, subject to renewables quotas. Data centers currently account for ~2.5% of national power demand and could double by 2030.

Goldman Sachs estimates that 40% of incremental power demand from data centers will be met with renewables, and the rest will be met with natural gas. Still, the IEA cautions that coal, gas, and nuclear will remain in the mix and, more likely than not, renewables alone won’t be enough.

China, with its East Data, West Compute strategy, is notably optimizing its land resources, thus accessing lower electricity prices in the western regions to power the data demands on the eastern coast. While this all sounds like the optimal solution, the IEA says that realistically, the energy mix to power more AI data centers will also include coal, gas, and nuclear. See here why renewables cannot replace all energy sources.

New Kind of Data Center. More Capital Required

The thing is, the AI data centers are not the same as what data centers used to be. Modern AI workloads require multiple GPUs working concurrently. In 2022, a cutting-edge AI system only integrated eight GPUs into a single server. But by 2027, the leading system will likely have nearly 580 GPUs stacked up in racks and working together, which requires 600 kilowatts, and for context, that is equivalent to delivering enough power for 500 US homes.

And capex continues to increase across hyperscalers, seen in each big tech’s earnings filings and announcements. Goldman Sachs projects that the five highest-spending U.S. hyperscale enterprises will have a combined spend of USD 736 billion in 2025 and 2026 for AI data center build-out.

The Love-Hate Relationship: Electricity, Data Centers, and AI

While talks about how AI will actually help optimize more efficient electricity generation, that has yet to be proven on a mass scale. The hurdles right now for more energy are 1) legitimate environmental concerns, 2) policy constraints, and 3) the ability for companies to ramp up.

On a human-race level, the question should be: if we cannot avoid the increasing need for data centers with the progression of AI development, is covering our earth with data centers the solution to our future?

On a nation-to-nation level, the U.S. and China are racing not just to build better models, but to secure the energy infrastructure that sustains them. In the U.S., hyperscaler capex will cross $700 billion in just two years, largely to supercharge AI data centers. In China, renewable energy mandates will tie AI expansion to the country’s green transition, although coal and gas will remain in the mix.

What’s clear is that AI has turned good old boring electricity into a new strategic resource. Whoever can deliver abundant, affordable, and sustainable power will not just run the best models, but they will set the pace for the global AI economy.

Welcome back to AI Proem. As mentioned, I apologize in advance that updates will be less regular in the next three months as I need some time off to welcome a baby. In the meantime, I’m also working on a fascinating audio/video project that I will release here in Q4. So please stay tuned, and thank you for your patience.

Subscribe now

CLSA’s latest China Internet: AI Superpowers 2.0 report makes a bold call: it says that with the explosion of AI applications, spanning everything from gaming to digital agents, China’s public cloud market could grow at more than 20% CAGR through 2027, reaching USD 80 billion. If true, that would be a huge win for the country’s leading providers—Alibaba, Tencent, Baidu, and ByteDance—who are building not just raw infrastructure, but “cloud-plus” services that bundle proprietary models, technology, and talent.

And my personal view is that it will benefit Alibaba Cloud the most, as this is exactly the trend Alibaba has been betting on. Among the BBAT / big four in China, it has leaned most aggressively into an enterprise-focused AI strategy, hoping to cement its position as the country’s number one cloud provider.

Beyond Infrastructure: What Clients Really Want

As we argued in earlier deep dives on Huawei and Alibaba, the next stage of the cloud race won’t be won on infrastructure alone. Enterprises don’t just need servers and compute, they need help training vertical models, building AI-driven applications, and customizing solutions for their industries. That’s where better technology and service become decisive.

As put it in plain language: “IaaS is primarily designed for IT operators, while PaaS is built for software developers.” And it’s really in the PaaS + SaaS layers where the battle for AI cloud differentiation will play out, so in that sense, the SOE players like China Mobile and China Telecom might not be able to offer competitive service.

The AI App Race: Competition or Consolidation?

On the consumer front, the picture is more mixed. Doubao, DeepSeek, and Baidu’s Ernie are gaining traction, but they still lag far behind ChatGPT in terms of training scale and user adoption. The question is whether this market plays out like China’s earlier internet battles. We saw that with payments, ride-hailing, and food delivery, usually two players come out on top after multiple players fought bitterly before carving up market share, so that could be a scenario between the LLM players. Another scenario that could happen is that we may see big tech eventually swallow the most promising startups.

If is right, as in the LLM game is only for those with serious capital, then consolidation seems more likely. Especially, as MiniMax, Moonshot, and Zhipu already have financial or strategic ties to the BBATs.

Falling Costs, More Apps

At the same time, costs are collapsing. The mass adoption of Mixture of Experts (MoE) architectures since DeepSeek V3’s rollout has sharply reduced inference costs in recent months. Combine that with the flood of open-source models, and you get exactly what we’re seeing now: an AI application explosion.

Lower costs make it easier for startups and enterprises alike to experiment, and that experimentation almost always runs on the cloud.

China’s Catch-Up Moment?

Despite their head start in infrastructure build-out, Chinese cloud providers have long lagged U.S. hyperscalers, mainly because their businesses remained domestically focused/ constraint. As JS Tan has argued, this has limited their abitility to monetize abroad.

But this may be an inflection point. IDC projects global public cloud demand will grow at 26% a year to USD 626 billion by 2027, with penetration reaching 50%. China continues to court the global south to join its ecosystem, as seen in the recent announcement at WAIC. Then it can finally find a way to leverage its cloud technology for international growth, as well as export cloud services along with its growing AI influence.

Why This Matters

Well, in the report, CLSA’s thesis is simple but powerful. With lower inference costs and the widespread adoption of open-source models, China could end up producing the largest number of AI applications in the world. If that happens, the country’s public cloud providers will be the biggest beneficiaries.

For now, the U.S. remains ahead in both cloud and AI application adoption. But if China can find a way to scale its advantages and create the most demand for cloud through AI, as well as export its infrastructure, then there is a big chance these companies can capture untapped market share.

Btw, a recommendation for the week is this Podcast interview with Kendra Schaefer, Partner and Head of Tech Policy Research at Trivium China, and Basilinna Senior Fellow, titled What is the Promise of AI?

If you have found value in my work, I would appreciate you sharing it with more people. It is word-of-mouth referrals like yours that help me grow.

Email: aiproem@substack.com

New to AI Proem? Start here

Almost two weeks ago, NVIDIA confirmed it will resume sales of its H20 AI chips to China, marking a sharp reversal of U.S. export restrictions. This abrupt policy shift ends a three-month freeze that had left China’s AI ecosystem scrambling—and signals a pivotal moment in the U.S.-China tech rivalry.

If you've been following my previous pieces on China’s AI push—from the DeepSeek frenzy to the robotics surge—you know this isn’t just chip news. This marks a pivotal moment for the entire AI value chain. In this post, I’ll unpack the implications from three angles: NVIDIA’s rebound, Big Tech’s relief, and domestic players’ recalibration.

With that…

Subscribe now

What Did the H20 Ban Mean?

The H20 is a scaled-down version of NVIDIA’s Hopper chips—purpose-built to meet U.S. export controls, yet still powerful enough for training and inference work inside China.

Washington initially blocked the H20 over national security concerns, fearing the chips would turbocharge China's AI advancements. That decision cost NVIDIA an estimated $10–15 billion in potential annual sales and forced it to write down $4.5 billion in inventory that was stuck in limbo.

Let’s Talk Capex

Morgan Stanley’s most recent report still projects China will spend around RMB 380 billion (~USD 52 billion) on AI infrastructure in 2025, with upside if supply bottlenecks ease. This H20 unban clears the biggest roadblock to accelerating mass-market GenAI rollouts, especially with apps like ByteDance’s Doubao and Tencent’s Yuanbao, AI-native interfaces, and cloud-based upgrades.

According to Morgan Stanley’s latest report, China AI: Reviving a Giant with H2O Chip Availability, China's capital expenditure (capex) trajectory is now aligned with NVIDIA’s own $50 billion revenue opportunity from AI accelerators in the region.

In short, as long as Chinese AI demand remains deployment-focused, then embedding LLMs into apps, services, and use cases will only increase the appetite for GPUs, and that hunger will not just persist but likely grow.

Big Tech’s Relief

For China’s tech giants, the BBAT—ByteDance, Alibaba, Tencent, and Baidu—the H20’s return is a godsend. These four accounted for roughly 80% of H20 shipments in 2024 and reportedly had over 1.1 million chips on order for 2025 before the ban hit. Many reports stated that insiders were concerned about the GPU constraints, which were considered the biggest bottleneck for further LLM development.

Now that the freeze is lifted, those plans are back in motion—along with cloud deployments, LLM training, and product timelines that had been stalled or derailed.

Semi Supply Chain Winners

This is also a tailwind for Asia’s hardware ecosystem.

During TSMC’s Q2 2025 earnings call, Chairman Dr. C.C. Wei called the H20 resumption “positive news,” and Morgan Stanley expects TSMC to beat its 45% CAGR target for AI semiconductor revenues over the next five years.

Samsung, too, stands to benefit: it had previously taken a ₩1.5 trillion (~$1.1 billion) write-off in Q2 due to the H20 ban. With sales back online, this loss may partially reverse. Samsung also remains a key GDDR7 memory supplier for NVIDIA’s RTX Pro series.

Component and server makers like Wistron, FII, KYEC, Auras, and Unimicron should also see demand rebound as GPU-related projects resume.

Domestic Hardware’s Short-Term Setback

However, it’s not all good news, at least probably not for Huawei, as we’ve written about how Huawei was initially poised to benefit the most from an H20 ban, given that the BBATs had no other viable option.

Ironically, while Jensen Huang recently called Huawei “formidable,” Huawei founder Ren Zhengfei has repeatedly admitted that their chips are still one to two generations behind. This unusual PR tactic, where NVIDIA hyped up Huawei while Huawei downplayed itself, was likely a combination of lobbying and expectation management.

The short H20 ban window gave Huawei and GPU foundry SMIC a temporary boost. However, domestic GPU vendors may now lose that incremental share to NVIDIA again.

That said, this is not a reason for Chinese firms to abandon self-reliance. What it will likely do is simply buy them time. As Huawei’s own leadership has said, the long game is still domestic.

NVIDIA’s Revenue Snapback

NVIDIA pegs China as a $50 billion market opportunity. With the H20 back, analysts expect it could recapture $10–15 billion in AI chip sales in 2025, including a backlog of $10.5 billion worth of demand per Bernstein estimates.

The catch? Every H20 unit still needs individual export licensing. There are over 1 million H20s in inventory or production, so timing will depend on how fast these licenses can be cleared.

NVIDIA’s CFO recently noted that shipments may resume in the company’s fiscal Q3, meaning revenue is expected to start trickling in then. But in the long term, China could again resume being a low-teens percentage of NVIDIA’s revenue mix.

(Oh, how Jensen’s probably so happy as all those 14-hour flights and press meetings under the scorching sun in Beijing wearing a leather jacket paid off.. And don’t forget, when he was asked what American firms have that others don’t? He said - President Donald Trump.)

Final Thoughts

The H20 ban lift gives NVIDIA a vital tailwind, re-energizes China’s AI giants, and temporarily relieves domestic pressure on Huawei.

However, this episode also highlights a broader truth: every GPU shipment now passes through a gauntlet of diplomacy and export controls. In this U.S.-China tango, nothing is guaranteed. (As there are a few unstable personalities involved, it seems)

And while NVIDIA may have regained access, the strategic plan in Beijing hasn’t changed. AI deployment comes first. And the push for indigenous innovation continues, just with a little more breathing room.

At , we’ve touched on the risks of AI going rogue or robot safety before, but only at a superficial level. Today, shares her research on WHO exactly is behind the trust and safety programs and the AI safety organizations globally, then zooming in on China.

Globally, AI governance and the growing number of Policy Fellowships, Trust and Safety managers, and shorter-term certificate-issuing programs are not only being offered by universities and government agencies, but also by private, wealth-backed think tanks, corporate branches, and social institutes.

Typically, this is how a leading AI policy research hub works –

Many of them have a university affiliation and blended funding options. The Center for the Governance of AI, commonly referred to as GovAI, is an Oxford-affiliated institute that focuses on frontier regulatory issues. It is backed by Open Philanthropy, the Leverhulme Centre of Cambridge, and formerly the Future of Life Institute. Its research found its way to UK parliamentary briefings and global tech diplomacy, while producing a robust pipeline for tech policy researchers active in the field. Similarly, Stanford’s Institute for Human-Centered AI relies on corporate partners, private donors (Phil Knight and Reid Hoffman), as well as federal grants.

There are, of course, corporate-backed institutes or branches of operations with a policy research mission that blends with a philanthropic angle. Prime examples include OpenAI’s governance team and VC-funded think tanks, such as Anthropic’s Long-Term Future Fund.

There are also official policy research branches; the main intergovernmental one is the OECD.AI Policy Observatory. Country-specific governments, such as those of the UK, South Korea, and the US, launch their versions of regulatory institutions specifically for AI.

AI governance has grown in a more decentralized way in the West. Startups like Anthropic or nonprofits like OpenAI’s governance team are funded by private capital and engage in open-ended conversations about AI’s long-term societal impact.

What about China?

By contrast, China’s approach to AI governance looks slightly different. Policy discussions are anchored in research institutes, national labs, and academic fellowships, with very few independent or investor-funded players in the mix.

China’s AI development is shaped by strategic national planning, including the New Generation Artificial Intelligence Development Plan and long-term technology roadmaps, as well as the Global AI Governance Initiative.

China’s AI governance framework is characterized by a top-down approach, where research is primarily driven by established institutions rather than independent or market-driven actors.

Funding for AI policy work primarily comes from public sources, including government ministries (e.g., the Ministry of Science and Technology, Cyberspace Administration of China) and regional innovation funds, rather than private foundations or venture capital.

Corporate involvement, while significant, remains pragmatically aligned with applied ethics and operational issues. Tech giants like Alibaba and Tencent engage in AI ethics and governance research. Still, their focus is on applied challenges such as algorithmic fairness, AI safety, and industrial integration, rather than broader ideological debates or charitable angles.

Venture Capital Follows Clear Signals

In China, venture capital plays a crucial role in scaling AI technologies. However, in policy-related areas, investments are shaped by strategic clarity rather than ideological experimentation.

Public-private coordination is common. Investment often comes from guidance funds tied to regional innovation clusters (e.g., in Shenzhen or Hangzhou) or from national-level funds focused on emerging technologies.

Private VCs adapt to clear priorities. Top venture firms like Sequoia China, while having their own professional taste and judgement, tend to align their portfolios with sectors that have clear regulatory direction and strong infrastructure support, such as industrial AI or medical tech.

This helps explain why there are fewer independent startups focused exclusively on AI policy. The opportunity space for governance-oriented ventures tends to be shaped by broader institutional frameworks, rather than by contrarian or disruptive goals.

One of the most visible parts of the AI governance ecosystem is the large number of fellowships and training programs embedded within universities, national academies, and tech consortia.

By comparison, fellowships in the U.S. or U.K. (e.g., at Stanford HAI or the Oxford Internet Institute) often serve as incubators for both academic and policy entrepreneurship. The Chinese model tends to emphasize institutional integration and tangible contributions.

Private-Sector Governance Efforts Do Exist: They’re Just Subtle

While independent think tanks are rare, that doesn’t mean the private sector is absent from AI governance in China.

University-industry collaborations are, in fact, growing. Notable joint labs are increasingly common, but they are still not as prevalent as in the US. These partnerships explore critical areas like medical imaging, urban smart systems, and financial technology, bridging academic research with industry needs while aligning with national AI priorities.

Hong Kong plays a hybrid role. Leveraging its unique position as a global financial hub, the region’s research institutions frequently partner with multinational firms and international foundations on digital governance topics. These collaborations are particularly active in FinTech regulation, cross-border data privacy frameworks, and responsible AI deployment, blending China’s policy-driven approach with global best practices.

Rather than acting as challengers to mainstream frameworks, these initiatives tend to work toward improving implementation, compliance, and real-world effectiveness. The system emphasizes coordinated planning, institutional research, and talent development over venture-backed experimentation. That means public funding is a key driver of research, and private actors contribute through technical solutions, not policy entrepreneurship.

If you’re trying to understand where experimental AI governance thinking might emerge in China, the answer may lie in interdisciplinary academic work, cross-border collaborations, or hybrid research initiatives—especially in cities like Hong Kong.

Check out these institutions:

• Tsinghua University – Institute for AI International Governance (I-AIIG), which focuses on Global AI governance, standards, and ethics. It also shoulders many of the “international cooperation” tasks, including collaborations with UN agencies participating in the annual "AI for Good" events. It will continue to be one of the most internationally visible AI governance hubs in China.

• Chinese Academy of Sciences – Institute of Automation, which focuses on ethical AI frameworks, algorithmic transparency, and technical safety research. The Research Center for Brain-Inspired Intelligence also explores governance implications of AGI. It works closely with government ministries to inform national standards.

• Zhejiang University – Interdisciplinary Center for AI and Ethics, which has a strong regional tie with Hangzhou’s tech scene, including Alibaba, Ant Group, and the latest DeepSeek.

• Concordia AI is an independent institution and social enterprise that has been producing numerous publications and seminars in China. It has both its finger on the pulse of the latest news in China and has been providing safety training to technical communities.

• Hong Kong University of Science and Technology – Center for AI Research (CAiRE) & Division of Public Policy. The focus is AI for social good, policy experimentation, responsible FinTech, and governance. Notably, it collaborated with Goldman Sachs on the development of ethical AI for the finance sector. This is one of the most outward-facing and globally networked AI centers in the region.

• The University of Hong Kong's Ethics Lab, which appears to have a strong legal and philosophical orientation applied to regulatory questions.

  

• Looking at Tencent’s AI Lab, Dr. Kai-Fu Lee’s Sinovation Venture, and Alibaba’s DAMO Academy, all these prominent AI research labs and ventures do not tend to have an ethics or regulatory lens, or safety research, while they have an ESG focus. Their leadership does, however, in general support ethics research by meeting attendance and other affiliations. This, so far, has been the most different thing compared to their Western counterparts.

• Corporate labs prioritize standards and compliance. Labs at Tencent, Huawei, and Baidu are involved in technical governance, such as fairness metrics, risk evaluation, and system reliability, often in collaboration with universities.

Feel free to DM or reach out for a one-on-one chat.

Share

A Few Updates on China AI Startups

Back in February, I wrote about China’s leading AI startups and how many of them are emerging from Hangzhou. So I was glad to see The New York Times pick up the thread this week with a story on Liangzhu, a Hangzhou suburb that’s become an unlikely hub for the country’s AI boom: NYT, July 6, 2025.

I still remember visiting Liangzhu a few years ago, during the COVID-19 pandemic (which is only about a 40-minute drive away from Alibaba’s main Xixi campus), when the indie art center had just opened and the area was still mostly empty, with streets and half-built apartment blocks. The speed of infrastructure buildout and how quickly talent concentrates around opportunity continue to amaze me in China.

Anyway, let’s take a look at how some of these startups are doing. Not all of the “tigers” or “dragons” of China’s LLM race are roaring ahead, so let’s see how a few of them are faring…

Baichuan LLM

In December 2024, Baichuan co-founder Hong Tao left the company citing personal reasons. And just today, news broke that another co-founder, Xie Jian, is also on his way out, although no details have been provided or reported by the media at this time.

For many, DeepSeek’s breakout moment rekindled global attention on China’s AI and internet sector. But for others, it marked the beginning of a fierce market competition they couldn’t keep up with. As I’ve written before, DeepSeek R1’s release forced China’s so-called “four tigers” to pivot one by one. And now, by the looks of it, for some, just pivoting wasn’t enough—this may be the beginning of the end for Baichuan as a serious contender in the LLM space.

MiniMax

On June 18, MiniMax launched its first reasoning model, MiniMax M1, which is reportedly the most capable open-weight model in China (Github), following DeepSeek R1. It also boasts a context window of up to 1 million tokens.

See video on X of a person using MiniMax’s Agent to clone Netflix in one click.

Anyway, a few friends and I have tested its new Agent mode and found it surprisingly capable, with research and PowerPoint presentation quality equivalent to what you’d expect from a fresh college graduate. Its video capabilities are also drawing praise, with many comparing it to Midjourney in terms of visual quality.

Speaking of Midjourney, they also released a new batch of videos two weeks ago, which are also worth checking out.

According to media reports, MiniMax is reportedly preparing for a Hong Kong IPO as early as this year. Backed by Tencent and Alibaba, it’s the first of the top-tier Chinese LLM startups to move toward public markets.

MiniMax’s open-source model is also now ranking near the top of several public LLM benchmarks, adding further momentum.

See X: https://x.com/ArtificialAnlys/status/1935311012137402678

However, this raises a question: if model developers can create truly capable and user-friendly agents, do we still need standalone agent companies? Manus may say otherwise. Will research labs see agents as lightweight consumer-facing apps, or will vertically specialized agent startups still carve out defensible ground?

Zhipu LLM

In April, Zhipu also confirmed plans for an IPO, a move broadly welcomed by the Hong Kong capital market. Hong Kong’s IPO scene is showing renewed signs of life, with The Financial Times reporting that the city is on track for its most active listings year since 2021. According to HKEX, 208 companies applied for primary or secondary listings in the first half of 2025, the most vibrant it’s been in a while.

So hooray for us in HK! Friends in IBD are saying they haven't been this busy since pre-COVID times, as complaints of working late hours are resurfacing, but energy and optimism seem to have returned as well.

And companies along the AI and robotics value chain are the biggest beneficiaries. Case in point, a recent example is the warehouse robotics company Geek+ (02590.HK), which just successfully listed on HKEX yesterday, is making a splash across the city and media. See their IPO coverage here.

It’s not just capital confidence that’s returning. I recently spoke with a senior executive at Alibaba, who said the mood within China’s internet sector shifted dramatically after the DeepSeek moment. Before that, most AI research had been quietly ongoing within Big Tech, overshadowed by broader geopolitical headwinds and pessimism around the Chinese internet sector.

Suddenly, breakthroughs were gaining recognition, and the conversation shifted back to research and product, rather than just high-level geopolitical narratives. Media coverage changed 180.

Manus

Oh, the new AI darling. Many Chinese startups are trying to study the Manus playbook as it received a shocking Benchmark capital injection in April. The thing is, most companies wouldn’t be able to replicate Manus’ marketing and PR genius. First, what I heard is that the company heavily relied on one of its investors’ PR and marketing teams and took a non-conventional approach. It sold scarcity and exclusivity, making it intentionally difficult for traditional media journalists to obtain access to the trial codes, and shared its trial codes with various KOLs, which then flooded English and Chinese social media. It brought out the ultimate FOMO in everyone. Part of the hype was driven by its legitimate quality product, but the mystique definitely added to the interest.

Even when it started to actively PR, it did not push for a strong corporate branding message, unlike many of the big tech companies, which are all about their flywheel stories. It simply let KOLS rave about them, and then released a founder-direct message, which made it seem much more approachable and “cool.”

Amid the last four months of media frenzy, it’s actually been cutting headcount in its Chinese office and quietly (but not so quietly) hiring aggressively in Singapore. Many of their hires are from Singapore-based former ByteDance employees or other leading Chinese tech firms. While loudly and proudly rolling out a new campaign across the Singapore MRT (subway) system, as well as actively hosting offline developer and founder events.

ByteDance

Then we’ve seen ByteDance recently release a new suite of models that are increasingly seen as challengers to Kling’s dominance in text-to-image and video generation. You can find a detailed coverage of ByteDance’s AI strategy here.

In general, it seems like the big tech players are charging ahead with LLM development, consumer app user conversion, talent recruitment, and capital deployment. Since DeepSeek’s R1 launch, Chinese big tech is going 100000% ahead in this global AI competition.

Sentiment Regarding Chinese LLMs

Interestingly, during a webinar hosted by The Information in May, Jui Tan of Lanchi Ventures stated that app developers are now testing a wider range of LLMs—including many Chinese models —instead of sticking to a single one. A phenomenon we’ve also heard from Altos Ventures in Korea, regarding startups being model-agnostic in building AI apps.

Additionally, I’ve heard from a few journalists in Silicon Valley that some U.S.-based startups are quietly experimenting with DeepSeek and other Chinese models. The issue isn’t technical performance—it’s PR optics. Like what the Alibaba executive said, business is actually doing well, and extreme negativity has dissipated.

But in a pragmatic sense, most non-Chinese application companies are still trying to steer clear of being seen as using Chinese LLMs, which doesn’t mean they’re not using them. For Chinese companies going global, some are still adopting the “China shed strategy” by setting up shop in locations like Singapore, which, frankly, I’ve had many inbound inquiries on how to do better.

Final Food 4 Thought: Are We in an AI Bubble?

Lastly, a discussion around whether we’re hitting an AI bubble has been widely discussed within the investment space in China. While most still believe that infrastructure needs to be continuously built out to meet the demands of future AI compute use, and many are still holding onto the idea that vertical use case is the future of AI SaaS.

Some have expressed a bit of skepticism and offered this framework (see the framework provided by 阿尔法工场, a WeChat blog backed by A-share focused hedge fund Keywise) that has caught my attention:

1. 「第一波：技术突破」 - DeepSeek证明了技术可行性，Manus展示了产品可能性

[First Wave: Technological Breakthrough] DeepSeek proved the tech is viable; Manus showed what the product could look like.

2. 「第二波：资本狂欢」 - 概念股涨停，融资估值飙升

[Second Wave: Capital Euphoria] ADRs soared, funding rounds went wild.

3. 「第三波：现实打击」 - 商业化困难，成本压力巨大

[Third Wave: Reality Bites] Commercialization remains challenging, and cost pressure increases

4. 「第四波：市场洗牌」 - 只有真正解决问题的产品能活下来

[Fourth Wave: Market Shakeout] Only products that solve real problems will survive.

Questions for you all: Where are we now, considering the framework above? A recent conversation with a UK-based industry leader made me reflect on three broader questions:

1. Does China’s AI sector take a more pragmatic, product-driven approach, and is that a strength or a limitation?

2. Does Chinese capital only follow proof of concept, whereas Silicon Valley is more willing to back bold ideas and founding teams?

3. Are industrial policies in China truly “top-down” drivers of innovation, or do they follow innovation once its usefulness has been proven?

Would love to hear your thoughts!

Stay tuned. A deep dive into RedNote and its AI strategy is underway, in collaboration with China consumer expert of Following the Yuan.

The Substack community really is incredible. When I messaged Mary Clare MacMahon that her recent Huawei piece on CUDA vs. CAAN was very informative, she said she referenced my old piece on Huawei for hers. It all goes full circle.

See original primer piece published in Jan 2025: Is Huawei like China’s Nvidia plus Google?

I updated some details throughout the piece since its May publication, as so much has happened in the last few weeks, most notably:

1. WSJ reported in late April that “Huawei has approached some Chinese tech companies about testing the technical feasibility of the new chip, called the Ascend 910D, people familiar with the matter said.” (WSJ)

2. Trump looks to scrap Biden-era rule that aimed to limit exports of AI chips, exploring semi tariffs (FT)

3. Nvidia looking to “downgrade” H20 again, Jensen Huang says next chip after H20 for China won't be from Hopper series (Reuters)

You’ve likely heard of Huawei during the infamous Meng Wanzhou trials in Canada, where she was put under house arrest due to U.S. charges against her for allegedly violating sanctions on Iran. I still recall the founder Ren Zhengfei, the father of Meng, went on an exclusive with CNBC while I was working there and was trying to pull all the cards to have the public court of opinion work in his favor and release his eldest daughter from detainment despite having kept a pretty low profile until then. And despite being on the U.S. trade blacklist for nearly six years, it appears to have continued innovating and releasing more high-end smartphones and tablets over the years.

However, since her return to the HQ, the company has remained relatively low-key and has primarily focused on its Harmony OS capabilities and R&D efforts. Until being thrust back into the spotlight again recently with Nvidia’s H20 chip ban on China, which was first suggested by President Joe Biden and now formally imposed by President Donald Trump. Prior to the trade war and the battle of semiconductor chips, it was just another tech firm in China.

Huawei is no mysterious company in China. Many know it for its sprawling castle/villa-designed campus and often bidding out top talent with crazy money. It also runs its own scholarship program to promote STEM and hires many domestically trained talents to go straight into its almost-like management trainee programs.

However, in the West, it has been deemed a mysterious and untouchable black box that just churns out products that seemingly shock Western innovators every time it does. Today, let’s try to demystify it a little.

So, fast forward to the present day. The U.S. government imposed new curbs on Huawei. Under the new guidance issued by the U.S. Commerce Department, the use of Ascend chips “anywhere in the world” by any company could be interpreted as a violation of American export controls last month.

Chips - Ascend Series

In early May, Nvidia CEO Jensen Huang raised concerns to U.S. lawmakers about Huawei’s growing AI capabilities, according to media reports. The H20 ban has led to speculations and (confirmations) that demand for Huawei chips have shot up as there is no better option left for Chinese AI companies then and after the ban was announced, Huawei timely announced its new Ascend 920 chip that will be rolled out later this year as it steps up its game to close the gap in technological capabilities. And shortly after the series of news, Jensen Huang said that the U.S. semiconductor export controls on China have been “a failure,” and are causing more harm to American businesses than to China.

After its most recent earnings report, Jensen Huang went on to Bloomberg Technology and was asked by the reporter whether the increase in demand for the reasoning service Blackwell has made up for the loss of the China market for the current period outlook.

Jensen said, “I guess so, but you cannot underestimate the importance of the China market. This is the second-largest AI market. This is the home of the world’s largest population of AI researchers and we want the all of the world’s researchers and the world’s developers building on American stacks.”

However, the Chinese might think differently, as the risk of being cut off keeps becoming a reality in advanced technology. And as we know, within the hopper architecture, H20s are the lowest specs they possibly can be.

“The limitations (from the U.S. government) are quite stringent, quite limited, if you will. H20 is as far down as we can take a hopper. We don’t know how to make it even less, and that’s really the limit,” said Huang.

“Whatever we make, (it) ultimately has to add value to the market. It’s a really tight rope, because the Chinese competitors have evolved and advanced greatly over the last year. Like everybody else, they’re doubling, quadrupling their capabilities every year, and the volume is increasing substantially. And remember, these are data center chips. They don’t have to be small and can be quite large.”

And highlighting his biggest concern: “Without American technology, the availability of Chinese technology will fill the market. Whatever we offer has to be competitive and add value to the market.”

To better understand the nuances, I spoke to Ray Wang, Washington D.C.-based semi analyst who broke down the technicalities in the most accessible way for me. [See our 20-minute interview here, recorded on May 9, 9 pm HKT]

Topics we covered:

1. What are the actual capabilities of Huawei's Ascend 910C and 920 chips, and how do they compare to Nvidia's H20s?

2. How may the lack of access to H20s affect Chinese AI firms?

3. What does it mean when people say Huawei has advanced systems that, despite their chips being less advanced, they can still work very well for AI training and inference?

4. Explanation of the software and hardware integration process? How do CANN and CUDA compare?

5. Who are the major players in China in chip design and manufacturing? And what is the relationship between Huawei and SMIC?

Subscribe now

What is the Ascend 920?

It is a chip created at the advanced 6-nanometre node level and expected to begin mass production later this year, which can make it a proper, fair alternative to Nvidia’s H20 chips.

The last generation of Huawei AI chips, the Ascend 910C GPUs, has already proven to have been used in inference training use cases by Chinese AI labs, such as in DeepSeek’s R1. According to Chinese media outlet AGI Hunt, Ascend 910C achieves 60% of NVIDIA H100’s performance, delivering strong inference results. They are built on SMIC (China’s answer to TSMC, albeit lagging much behind) 7-nanometer N+2 process with 53 billion transistors. It is said that on a per-GPU basis, its performance matches Nvidia’s H100s (see more detailed specs comparison here). However, some say even though the Ascend 910B and 910C continue to improve, they still lag about ~5 years behind in terms of performance, per-watt, and interconnect density compared to Nvidia’s H100s. But as of now, they’re good enough, or as good a replacement as Chinese AI companies can get their hands on.

wrote that this gap won’t be closed in one or two product cycles. It’s a function of fab capacity, export enforcement, and ecosystem readiness that makes the U.S. remain in the lead in chips in the long run, even as open-weight parity inches closer.

How did Huawei get here?

In many ways, industry experts are saying the ban on Huawei under U.S. trade restrictions since 2019 has pushed the private company to focus on R&D and laid out the foundation of their capabilities today for AI chips. It originally relied on chip imports from American vendors such as Intel and Qualcomm, but the ban effectively severed this supply chain, potentially in hopes of hindering Huawei's ability to manufacture new laptops and phones at scale. But similarly to what happened with DeepSeek, “necessity is the mother of invention.”

Refer to the Deep dive into Huawei piece that looked into the history of Huawei’s cloud build up and its six-year grind.

But as I have written in the article highlighted above, Huawei did not take on the approach like other big tech did, whether in China or the US. It didn't go cloud-first to try to sell to enterprises. Instead, it honed in on privacy and security, which then often won the hearts of Chinese SOEs and government entities.

It largely embraced what wrote as “AI in a box” approach, which essentially served its clients the 一条龙服务 full stack service, starting from the foundations of hardware—including “packaging GPUS, networking equipment, server racks, databases, container services.” Then, on top of that, almost like a cherry on top service, is the software side of the AI model—Pangu, which then you can tweak and modify. At its core, the company has always been a hardware equipment business.

In an exclusive conversation between AI Proem and Nuo Jiang, Founder of GroundAI and former Chief Digital Officer at Huawei in Singapore, Huawei’s unique strengths lie within its R&D and hardware capacity.

“First, the company is the world’s second-largest company by its investment in R&D, after Alphabet, which gives strategic investment and endurance to long-term chip design. Second, the company’s DNA is in hardware design and manufacturing. Third, owning the leading Chinese semiconductor company HiSilicon, gives it a massive advantage over its peers. Lastly, its end-to-end capability, from material science, chip design, manufacture, and cloud stack to its mobile operating system, e.g., Harmony OS, is unparalleled globally. All together, they create a compelling ecosystem.”

But what will a Nvidia H20 ban truly mean for China’s semi sector? Ray Wang said Huawei is using this opportunity to build out their own semiconductor foundry which will be a positive push for China’s AI chip making capabilities.

1. Is that it will add the competition for SMIC, which currently monopolizes China’s chip-making space, despite not being able to match TSMC’s capabilities, it still remains the only fab for all chip orders in China.

2. This potentially adds a possibility for China to make advanced chips (< 7 nanometers).

3. It can potentially mean Huawei can be self-sufficient and increase overall production capacity.

[By the way, Kyle Chan did a great background piece on Huawei; see it here.]

CUDA vs. CANN

In my original Huawei piece, I included a section on the software but removed it in the end. So today, I want to spend some proper time delving deeper into this.

Nvidia has a CUDA system, and Huawei has a toolset abbreviated as CANN. Both CUDA and CANN are tools that help programmers make their applications run faster by using the power of graphics cards. CUDA is more general-purpose and widely used in various fields, while CANN is designed to focus more on making AI applications work better on Huawei's devices.

According to Nvidia’s data, over 2 million registered developers utilize CUDA, and on average, 438,000 downloads of CUDA occur each month. In comparison, there is no publicly available information on the adoption rate of CANN. However, some have said that the same size and type of training or inference run on Huawei’s CAAN is often slower than that of CUDA. Thus, the number of developers utilizing CANN is probably 1/10 of the CUDA developers in the ecosystem.

Now, CUDA has largely been popular due to its free access and developer friendliness. Huawei has taken note of that and launched its own two platforms: MindSpore and CANN. To get a more technical breakdown, check out Mary Clare McMahon’s coverage here. What she accurately points out is that Huawei has realised that to be truly competitive and considered as an alternative to Nvidia for AI clients, it has to undermine Nvidia’s software moat, which means it’s more than just reaching parity in hardware GPU performance but upping its own software capabilities. But first, what is Nvidia’s CUDA?

CUDA stands for “Compute Unified Device Architecture.” It is a technology created by Nvidia that allows programmers to use the power of Nvidia graphics cards (GPUs) to perform complex calculations much faster than regular computers (CPUs).

• Parallel Computing: CUDA lets computers use many small processing units in the GPU to work on different parts of a problem simultaneously。

• Programming: With CUDA, programmers write code that tells the GPU what to do. They can write this code in languages like C or Python, which are easier to understand than machine language.

• Applications: CUDA is used in many fields, such as video games, scientific research, and artificial intelligence (AI), because it can handle large amounts of data quickly.

Subscribe now

What is Huaweis’ CANN? And MindSpore?

CANN stands for Compute Architecture for Neural Networks, which Huawei developed. It’s similar to CUDA but specifically designed for AI and machine learning tasks.

• Focus on AI: While CUDA can be used for many calculations, CANN is tailored for neural network tasks. Neural networks are models that help computers learn from data (like how we know from experience).

• Efficiency: CANN allows developers to create AI applications that run efficiently on Huawei’s chips. This means they can build intelligent applications that recognize images, understand speech, or make predictions based on data.

• Compatibility: Just like CUDA works with Nvidia GPUs, CANN works with Huawei's hardware, helping developers take advantage of the specific features of Huawei’s chips.

Together, CANN and MindSpore work as a comprehensive AI software stack designed to optimize AI model development and deployment on Huawei’s Ascend AI processors.

CANN provides the tools and runtime needed to execute high-performance machine learning models on Huawei hardware. Meanwhile, MindSpore is designed for ease of development, automatic differentiation, and seamless model training and deployment. It seamlessly integrates with CANN to leverage Ascend hardware efficiently.

Beyond enhancing the user experience and deploying engineers to China’s leading AI firms to better support CAAN integration, Huawei is focused on prioritizing its compatibility with open-source machine learning frameworks such as PyTorch. As Mary wrote, “the challenge for Huawei, then, is not only to make PyTorch run on Ascend hardware but also to make it run well enough that developers don’t notice they’ve switched ecosystems.”

CloudMatrix System

SemiAnalysis, the leading semiconductor research firm, recently wrote that the CloudMatrix 384 is “China’s newest and most powerful domestic solution,” directly competing with the GB200 NVL72, and “in some metrics is more advanced than Nvidia’s rack scale solution.”

The importance of the system is that although Huawei’s chips are still technically a generation behind the leading Nvidia chips, its scale-up solution is said to be a generation ahead of Nvidia and AMD’s current products on the market. What does that mean?

Huawei is compensating for its less powerful individual processors by deploying many more of them and leveraging a highly scalable, high-bandwidth optical interconnect architecture. And though it requires more space and energy to power the compute, this has not been a constraint for Huawei. This approach has temporarily now made CloudMatrix a viable alternative in China (likely domestic demand is so high that exporting this will not be a priority in the short-term) where Nvidia's cutting-edge GPUs are restricted or unavailable, despite higher costs and energy use.

• Lower performance chips: Huawei's CloudMatrix 384 uses 384 Ascend 910C processors, which individually have lower performance and efficiency compared to Nvidia's more advanced Blackwell GPUs, but collectively deliver about 300 petaFLOPS of BF16 compute-roughly 1.7 times Nvidia GB200 NVL72's ~180 petaFLOPS, according to SemiAnalysis.

• The CloudMatrix system achieves this through sheer scale and a sophisticated all-optical mesh network that provides much higher memory capacity (3.6× more HBM memory) and bandwidth (over twice Nvidia's).

• Power efficiency trade-off: CloudMatrix consumes nearly four times more power and is about 2.3 times less efficient per watt than Nvidia's system.

• Huawei's approach reflects a "brute force" strategy to compensate for less advanced chip manufacturing technology by maximizing parallelism, memory, and interconnect speed at the system level rather than per-chip performance.

And its capabilities were recognized by Jensen Huang. In the same Bloomberg interview mentioned above, he added that though Huawei’s best offer is currently only “comparable to a H200,” its CloudMatrix system is able to scale up to an even larger system than Nvidia’s latest grace Blackwell.

“And so you know, Huawei is a formidable technology company,” said Huang.

As for the Alibaba and Tencent of the world, they have “no choice” but to buy from Huawei in the face of U.S. policies in place, said Jensen. At the same time, adding that it makes sense and is “prudent” for the Chinese developers to develop their stack on Huawei. As we can see, he’s caught between a rock and a hard place.

Subscribe now

Parting Words

This answers many of the questions around what Huawei is doing, and it can really replace Nvidia for Chinese AI firms. But as I wrote in this Fortune Op-ed, we’re still in such an infancy stage of AI adoption, and in many ways it is still too early to say who has “won”. The decoupling and bifurcation in the industry have and will continue to push China to be more self-reliant.

Whether it's in LLM development or semiconductor advancement - DeepSeek or Huawei, now willingly or not, being touted as national champions will rally the support of national-level capital, policies, and support in China. And if the commercial choice is removed from companies (no more Nvidia chips), then capitalists (businesses) will also lean towards them. Hence, Jensen Huang’s plea to not close off the Chinese market to his business.

Speaking to Ray Wang, the semi-analyst, also reconfirmed my belief that export control will only hinder China’s AI development in the short term, but in the long term, it is actually pushing China to find more self-reliant and innovative solutions. Although U.S. companies like OpenAI and Meta currently still have the huge upper hand in access to cheap compute and higher quality hardware, that gap could be shrinking exponentially day by day.

Share

Before we dive in, at a high level, I want to quickly sum up a few key themes/ topics we’ve written about, and I was pleasantly surprised to see that it all largely aligns with Morgan Stanley’s findings from its May 13th China AI Blue Paper, which I will reference for this write up. (If anything, I think we’re ahead of the curve ;p)

1. Physical AI: China is projected to hold ~30 % of the global humanoid installed base by 2050 (302 m units), with commercial sales >50 million per by then, according to MS, and we’re seeing areas of continued AI integration. [See Electric Dreams, Robot Reality: China's EV-to-AI Evolution]

2. Industrial robots + AI = cross-disciplinary fusion, or what I’ve been calling a convergence of technology moment. There are over 1.7 million industrial robots in Chinese factories (51 % of global demand). Couple that with AI software. Think of the future of manufacturing and labor. [See Rise of China's Robotics Industry: from Manufacturing Arms to Embodied AI]

3. Super apps powering the next wave of Agentic AI: large proprietary data sets and a large base of DAUs set consumer internet big tech up for catching this opportunity. [See China's AI Application Super Powered: Open-Source Momentum Meets Super-App DNA]

4. Proprietary data strengthens multimodals: Google (YouTube), Alibaba (Youku), Kuaishou, Meitu, and Tencent (Tencent Video) lead the way in images, videos, and text models. We’ve examined Bytedance and Kuaishou before, which generate around 80 million new videos per day and have over a billion users combined daily. [See The Ghibli Hype: Next Frontier is in Video & Introducing China’s Leading AI Video App, Kling] & [Why Tencent's Integration of DeepSeek Into its AI App is a BIG DEAL]

5. Inference- price deflation continues, and model commoditization sets up the space for better consumer-facing use-case adoption and a free-to-use business model. Monetization on the consumer end will come from ad sales, value-added services, and new innovative ways. [The Great AI Power Shift - From AI Models to AI Applications

6. Open-source tilt and model quality parity: China supplies 40 % of the top-10 open-weight models on HuggingFace, with Qwen-72B on top; DeepSeek-R1 & peers close the quality gap on Chatbot-Arena scores. [See this article]

7. China continues to push forward green energy solutions to power the AI boom. AI tasks draw up to 33× the energy of task-specific software, and China already has 246 green DCs and targets 100 % renewable DC power by the early 2030s. (This, let’s see, but there are notable efforts) [AI Arms Race Far From Over: Chips is Only Half the Game, Infrastructure is the Other]

Subscribe now

But like I said, I want to dive deeper into the open-source/ open weight topic today. I wrote about DeepSeek’s open-source week in March and the personal and mission drivers that led CEO Liang Weng Feng to choose to open-source DeepSeek here, FYI.

Closed vs Open

Morgan Stanley wrote that China’s open-source LLM ecosystem is gaining significant traction with the launch of DeepSeek, but we didn’t need the investment bank to tell us that. Hugging Face has ranked several iterations of Chinese AI startups’ open-source models highly for common sense reasoning, math, coding, and reading ability. And not just the startups ranking highly, but also the big tech models. Chinese companies currently own four of Hugging Face’s top-ten open-weight slots, with Alibaba’s Qwen-2.5-72 B-Instruct leading at number one.

In fact, in Morgan Stanley’s report - Exhibit 42, four out of ten red bars are of Chinese AI labs. Compare that to a snapshot from late 2023, where none were Chinese. China’s open-source models are catching up and leading the way.

No One Wants To Pay

According to Morgan Stanley’s research, as the graph above shows, 83% of enterprises prefer open-source models for the flexibility to tweak and run them privately. And only 21% say they would pay a premium for closed-model software. Budgets are moving accordingly: hardware’s share of next-year AI spend rises to 34%, software slips to 40%. So if you sell SaaS wrappers in China, prepare for thinner margins, or you may need to bundle unique data or tools that users can’t replicate behind the firewall.

Before, I thought only consumers would put free-to-use as top priority, but shockingly, AlphaWise China CIO Survey found that enterprises are found to prefer open-source models - and not paying for the (premium) models - as well.

And this may challenge what we’ve previously written about, that Chinese AI wrapper companies were looking to innovate within vertical industry use-cases. So, if even enterprises don’t want to pay (or pay little for APIs) for AI wrappers and instead make tweaks and adjustments based on customized needs on top of open-source models. These vertical AI applications may need to reconsider their monetization strategy or offer something SO ready to use.

Thus, price remains an advantage for many of these Chinese open-source/open-weight models, leading by example, is DeepSeek, which has its pricing undercutting every Western frontier model today ( ~ USD 0.55/2.19), and its API price is expected to continue to drop. Cheaper inference and cheap tokens will collapse the moat that closed weights have enjoyed until now. The growth of users and adopters will likely come from value-added services and sheer volume of users, which, in a way, will make even the enterprise playbook look more and more similar to the consumer-facing strategy.

And will open-source become the new norm? Venture Capitalist Marc Andreessen said that American AI labs also need to seriously consider open-source AI or “risk ceding control to China.” And as Business Insider wrote, as hardware bifurcation deepens, more attention is being shifted to software. Andreessen added that it is "plausible" and "entirely feasible" that open-source AI could become the global standard.

Subscribe now

Rumors: DeepSeek R2 Leak Reveals Cost Efficiency and Performance Metrics

Over the past weekend, news of an R2 leak has circulated on X and Chinese social media. The R2 model was reportedly scheduled for release in early May, but that may now be accelerated. According to SCMP’s report, Chinese stock-trading social-media platform Jiuyangongshe wrote that R2 is developed with a so-called hybrid mixture-of-experts (MoE) architecture, with a total of 1.2 trillion parameters, making it 97.3 per cent cheaper to build than OpenAI’s GPT-4o.

I would still approach this news with caution, as we’re not sure how reliable these online rumors are. But so far, the consistency of details across multiple tech media (including SCMP, I linked to) and the timing of the leaks (ahead of a likely early May launch) have fueled expectations.

Key Rumored Features (Reddit)

• Massive Scale:
  DeepSeek R2 is said to feature a 1.2 trillion parameter architecture, with 78 billion parameters active per token, leveraging a hybrid Mixture-of-Experts (MoE) design. This approach allows for high performance while keeping computational costs low.

• Cost Efficiency:
  The most widely cited rumor is that R2 is 97.3% cheaper to train and run than OpenAI's GPT-4 or GPT-4o. API pricing is rumored at $0.07 per million input tokens and $0.27 per million output tokens. This would make R2 dramatically more affordable for enterprise and developer use.

• Hardware Independence:
  Unlike most leading AI models, DeepSeek R2 is reportedly trained entirely on Huawei's Ascend 910B AI chips, not Nvidia hardware. This marks a strategic shift, signaling China's growing self-sufficiency in AI infrastructure. Reports claim DeepSeek achieved 82% utilization of its Huawei GPU cluster, reaching 512 PetaFLOPS of FP16 compute.

• Training Data and Benchmarks:
  R2 is rumored to have been trained on 5.2 petabytes of high-quality data, spanning domains like finance, law, and patents. Early leaks suggest it scores 89.7% on C-Eval 2.0 (a Chinese language benchmark) and 92.4% on the COCO vision benchmark, indicating strong reasoning and multimodal (text + vision) capabilities.

• Multimodal Capabilities:
  Multiple sources speculate that R2 will natively support both text and vision tasks, which is in line with the latest trends in state-of-the-art AI models. This would be a major step up from DeepSeek's previous text-only releases.

• Open-Source Release:
  DeepSeek is expected to continue its open-source strategy, potentially making R2 freely available and further challenging established players like Meta and OpenAI in the developer ecosystem.

The leak also noted what we’ve predicted: a Nvidia H20 ban will matter much less to Chinese LLM development now as they shift away from the U.S. supply chain and to domestically made chips. [See Fortune op-ed U.S. tariffs will hasten, not slow, China’s drive for tech self-sufficiency]

Ok, China's leading technology firms, Alibaba, and Baidu, are actively advancing their artificial intelligence capabilities. Over the past week, they’ve both rolled out new models and are exploring various applications and monetization strategies. These developments coincide with observations regarding China's Internet valuation closing the gap with U.S. peers.

Baidu's Strategic AI Roadmap and Monetization Efforts

Baidu launched a new AI model, Ernie 4.5 Turbo. At the Baidu Create 2025 event, the tech giant announced two new models, ERNIE 4.5 Turbo and X1 Turbo, which are described as enhancing multi-modal and deep reasoning capabilities at low cost.

The event also saw the introduction of several multi-modal and multi-agent applications. Baidu emphasized the importance of MCP (Model Context Protocol), positioning it as an industry standard addressing developers' pain points.

Baidu has outlined its AI strategy and recent achievements, highlighting both short- and long-term opportunities in AI, including digital humans, coding agents, and autonomous driving. Chatbots are noted as still evolving, currently showing low engagement.

The company’s short- and long-term priority in AI development remains in lowering training and inference costs of large models, creating opportunities in digital humans (streaming/customer service) and coding agents (facilitating AI agent development).

Although it’s been leading in autonomous driving, many competitors in China are catching up, and the next phase of focus will be achieving AGI, but that is expected to take over 10 years. Its robotaxi (RT-6) is deemed highly competitive and cost-effective compared to many overseas peers. So there’s more to anticipate in this space.

Frankly speaking, Baidu’s 2C chatbot and offerings have lagged behind competitors and have not offered any major breakthrough or anything exciting in a while.

Alibaba Debuts Qwen3 with Enhanced Performance

Alibaba has introduced Qwen3 on Tuesday, described as setting a new benchmark for open-source AI. It is based on a hybrid reasoning model. wrote a very insightful and technical piece here.

Just to highlight a few key points from the Qwen3 debut:

• Hybrid Reasoning Model: Uses a 4-stage training process .

• Model Architecture: Released six dense models (0.5B, 1.8B, 4B, 14B, 32B parameters) and 2 MoE models (30B with 3B active and 235B with 22B active).

• Training Data: Qwen3 is trained on 36 trillion tokens, double the number used for Qwen2.5 .

• Enhancements: Key enhancements in multilingual tasks, stronger agent integration, superior reasoning, and human alignment.

• Performance: Achieved top results in different areas, and performance comparable with other top-tier models.

The hybrid reasoning model allows users to switch between thinking modes for complex, multi-step tasks such as mathematics, coding, logical deduction, and a non-thinking mode for fast, general-purpose responses is available for global users to download. This is achieved via the 4-stage training process, incorporating long chain-of-thought (COT), cold start, reinforcement learning, thinking mode fusion, and general RL.

Alibaba also highlighted that it lowered deployment costs for developers using Qwen3. The Qwen3-235B+22B model is mentioned explicitly as having significantly lower deployment costs compared to other state-of-the-art models.

Comparing Qwen3 to its predecessor, Qwen2.5, the new model demonstrates superior reasoning capabilities. Qwen3's capabilities across benchmarks such as AIMEMaths (mathematical reasoning), LiveCodeBench (coding proficiency), BFCL (tool and function calling), and Arena-hard (instruction-tuned LLM) are described as comparable to other top models, including DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro.

GitHub：https://github.com/QwenLM/Qwen3

Hugging Face：https://huggingface.co/collections/Qwen/qwen3-67dd247413f0e2e4f653967f

On the Alibaba Cloud site, the company wrote: “Qwen3 represents a significant milestone on our journey towards Artificial General Intelligence (AGI) and Artificial Superintelligence (ASI). By scaling up pretraining and reinforcement learning, we have achieved a higher level of intelligence. We have seamlessly integrated thinking and non-thinking modes, providing users with the ability to flexibly control their thinking budget. In addition, we have expanded support for multiple languages, helping more users around the world.

Looking ahead, we plan to enhance our model from multiple dimensions. This includes optimizing model architecture and training methods to achieve several key goals: expanding the scale of data, increasing model size, extending context length, broadening the range of modalities, and leveraging environmental feedback to advance reinforcement learning for long-term reasoning. We believe that we are transitioning from an era focused on training models to one centered on training Agents. Our next generation of iterations will bring meaningful progress to everyone's work and life.“

This is a pretty big deal as it reinforced Alibaba’s position as a leading player in China’s AI space with the advantage of also having infrastructure support and distribution reach.

said to CNBC today, “Alibaba’s release of the Qwen 3 series further underscores the strong capabilities of Chinese labs to develop highly competitive, innovative, and open-source models, despite mounting pressure from tightened U.S. export controls.” Adding that, “in the broader context of the U.S.-China AI race, the gap between American and Chinese labs has narrowed—likely to a few months, and some might argue, even to just weeks.”

“With the latest release of Qwen 3 and the upcoming launch of DeepSeek’s R2, this gap is unlikely to widen—and may even continue to shrink,” said the Washington, DC-based semi/ tech analyst.

Conclusion

As written in my Fortune op-ed, “the U.S. may hope that the right mix of tariffs, subsidies, and export controls can preserve its tech leadership. But instead, the continued push to cut off China’s access to advanced technology is going to make it more self-sufficient out of necessity. The trade war, even if it leads to a deal, will push China to invest in its tech sector even more. The next time the U.S. tries something like the H20 chip ban, it may mean very little to the Chinese AI ecosystem.

Competition can be healthy, but it doesn’t need to mean collapse. The challenge for both the U.S. and China is to draw clear guardrails to support national security without shutting down collaboration entirely. Climate tech, healthcare, AI safety, academic research, and open-source development should still be seen as areas for cooperative leadership.”

On top of that, Jefferies wrote in a research note that China’s internet valuation has narrowed the gap with U.S. peers in recent months. And I think that could partly be due to some volatile adjustments for the Mag 7 in the U.S. we saw over the last few weeks and a reinstated confidence in China’s tech sector boosted by DeepSeek.

Jefferies added that, recent earnings results for most large and mid-cap companies across sub-sectors were either in line with or beat expectations, setting a solid foundation for expected growth. Things are catching up and China’s AI space seems to be getting more and more self-reliant.

Ok guys, I’ll (try to) be offline for a week. Taking my little one to Japan for some major yumz and nomz.

Subscribe now

What happened?

Goodbye to globalization as we know it.

U.S. President Donald Trump announced a series of tariffs on April 2, on what he called “Liberation Day” and what will be known as the reinvigoration of American industrialization.

Trump’s slew of tariffs hit countries around the world—including China, obviously. Then, the world panicked, and stock markets everywhere crashed. Headlines like this plagued the whole internet. Global stocks are seeing a sell-off. CNBC: Dow futures fell 900 points as Trump's tariff market collapse worsened. Nasdaq futures had their worst day since 2022. This is the biggest market capital erosion of big tech since forever, essentially.

At a high level, you can see it as three buckets of tariffs: 1) China tariffs; 2) Rest of the world tariffs; 3) Auto tariffs mostly hurting Germany, Canada, and Mexico.

Key updates from the U.S.:

1. Trump claimed a 10% tariff on all nations and a much higher rate of up to 50% on individual countries starting April 5.

2. The U.S. put an additional 34% tariffs on ALL Chinese goods, in addition to the 20% tariffs previously announced.

3. New tariffs include 49% on Cambodian products, 46% on Vietnamese imports, and 20% on EU goods.

4. NAFTA got smashed with 25% on Mexico and Canada, 10% on Canadian energy, 25% on all steel and aluminum, 25% on all foreign made cars.

Key points from China:

5. China retaliated with a blanket 34% tariff on all U.S. goods imports, effective April. 10.

6. Sixteen U.S. companies were added to China’s Control List for dual-use exports and 11 firms to the Unreliable Entities List.

7. China launched an anti-monopoly investigation into U.S. biotech giant DuPont and a series of anti-dumping proves into various medical companies.

   ---

Now…This is the first time Trump has expanded his tariff strategy so aggressively beyond China. The ripple effects will touch many sectors we’ve been following—renewable energy, EVs, batteries, and even parts of big tech’s supply chain—all of which could impact AI investments down the line.

For the U.S., the risks are now mostly twofold. Retaliatory tariff risks and inflationary risks may lead to a domestic US recession.

China’s response: tit-for-tat

From China’s perspective, their move was a tit-for-tat, not an intended escalation. However, the back and forth has sent shock waves and reverberations across sectors. Markets crashed like they really went to sh*t the last few days- in the U.S. and across Asia. The thing is, most of China was on break for Tomb Sweeping Festival, a holiday dedicated to cleaning up the tombs of ancestors and loved ones, paying respect to the graves, and a little respite for the 996ers since the Lunar New Year. But that didn’t really go as expected for many…

I happened to be in Beijing and Shanghai the past five days and as my husband and I were frantically checking the news and stock market every 10 mins. As we were strolling around Xintiandi and The Bund, the most commercial/ touristy/ vibrant areas of Shanghai city, we looked around us and asked - how come everyone else look so calm?

For most of China’s domestic tourists and young aspiring “网红” (influencers), the trade war seems like background noise at this point. Lines outside D2C makeup brands, pastry shops, and even Cartier were as long as ever. It feels like the trade war has been normalized—just another “it is what it is” situation.

At dinners with friends and family, though, it was a different story. Those over 30 couldn’t stop talking about the trade war—whether it was businesses relocating manufacturing to Vietnam or concerns about how a U.S. stock market downturn might hit A-shares. Someone even asked if this could escalate into actual warfare.

Global Shockwaves - especially for America’s Western allies

But honestly, the Europeans, Canadians and ASEAN countries were much more blindsided probably. In Canada, anti-American sentiment is at an all-time high. I remember back in high school when someone in Social Studies 11 class said they hated being mistaken for Americans while traveling—that was about as far as anti-Americanism went back then.

Over the last few days, my social media feed has been full of boycotts against U.S. goods like California oranges— and it’s not just from the BC lefties but across the board (including the Albertans). Bloomberg’s report on rising Canadian hate towards the U.S. is the highest we’ve seen in modern history.

Bloomberg reporter puts it exactly how frankly we all low-key thought, “what happened in Canada and the U.S. was kind of one in the same” during its Bloomberg Technology podcast. But that has changed, car parts have been heavily hit by tariffs and that retaliatory fear is driving even “America-made” Tesla’s stock prices to drop.

Canadian leader Mark Carney called the tariffs “unjustified, unwarranted, and in our judgment misguided.” Slapping a 25% tariff on U.S.-made autos. [For someone who grew up 40 minutes away from the US-Canada border where we saw people from the two nations drive across the border for things even as small as groceries and gas refills. It is crazy to see that the two are no longer BFFs!!]

Across the Pacific Ocean, ASEAN countries were also smacked in the face with a shocker. The Singaporean Prime Minister Lawrence Wong said in a speech warning his fellow citizens that “we are entering a world that is more dangerous and unpredictable.” Adding that, the tariffs that were just announced by the U.S. will have an effect on the global economy and wider global order.

Singapore has long balanced a peculiar relationship with the two superpowers, as many in the APAC region have over the years, and tried not to piss off either side, tbh. But largely leaning towards U.S., hosting military bases and serving as a hub for American tech firms and financial firms. It’s quite shocking to see the leader of the Lion City make such claims, to rethink strategic alliance and positioning, to be cautious of what was taken for granted.

And on the other side of the world, you’ve got French President Emmanuel Macron asking French companies to stop investing in the U.S. signaling broader shifts in global trade dynamics. This attempt to reorganize the global industrial and trade order may be more than just the short-term stock market pain. It may lead to longer-term reputational challenges for the U.S.

American exceptionalism, no more?

This brings me to the Trump administration’s move and its PR implications…

This brings me to a piece we wrote about quite a long time ago which is about Trump’s undiplomatic approach to international relations. And unpredictability. In many ways, Nico and my assessment back then was that it may turn allies to rethink US partnerships and US reliance and could give Chinese companies and innovation an opportunity to export… as friends unite over a new found foe. But now, we’re not sure what will happen, everything just seems so chaotic, as this degree of “America First” is beyond what any of us expected.

From a PR stance, Trump’s efforts are really hurting America’s soft power and reputation.

Whether this 54% tariff will hold, we are not sure but one the thing we know with the current Chinese leadership is that they really care about “face.” For some nations, kowtowing to the Trump administration or kissing the ring may simply be enough to feed his ego in this “deal-making” process and encourage him to soften the tariff blows. But what Trivium’s Andrew Polk also wrote here, “the Trump team genuinely seems to have expected that China would sit back and take this broadside — something that is completely incongruous with how China has acted over the past seven years, since the trade and tech war started.”

Now, China doesn’t bow down to anyone… that’s their take. At least not publicly. So the retaliatory tariffs were only expected…

Meanwhile, China is eating this moment up… Robert Wu from Big One Lab wrote,“I won’t expect China to create substantial troubles for U.S. businesses in China because now is the time to project an image of openness to contrast with Trump’s isolationist.”

How does it affect the tech and AI industry?

Now on the surface, the heavily tariffed industries do not touch on key industries for AI yet. But obviously overall sentiment is not good for any cross-border collaboration or the idea of open-source sharing. Confidence is low across…

One thing that President Xi and President Trump seemingly agree on is their dislike of big tech companies/ executives. The Mag Seven got crushed this week. Apple saw a wipe out of $300 B. Meta is having a pretty crappy time… all these big tech are getting hit. It’s said to be a USD 3 trillion+ wipe out in market cap YTD. Many of them shifted their manufacturing out of China to Vietnam and Bangladesh over the last few years since the trade war started to de-risk their businesses but that obviously is irrelevant now.

These U.S. companies, big and small, now basically have two options ahead of them 1) absorbing the costs that will hurt profitability or 2) raising prices that will hurt demand. A rock and a hard place.

China Tariffs

Despite having all the major U.S. tech executives at his inauguration ceremony and in China, Xi’s recent meeting with all of them to bolster confidence in the economy. The two world leaders clearly do not like how much power, influence, and fame these executives have. Protecting big tech’s interests just doesn’t seem top of mind. (For more detailed China updates, check out ’s what’s happening in China.)

Now, one of the voices that has come out is saying that the Trump administration has pushed China to become self-sufficient or at least try to be starting in 2018, and thus the Chinese supply chain for advanced tech, albeit not all world number one, has all become China-made.

And as Chinese companies continue to figure out how to make “good enough” chips, this tariff battle can actually help China advance in its technological aspirations faster, since Trump’s plan to reshore manufacturing will take time. That’s not just moving the supply chain back to the U.S., but finding all the necessary raw material, training up skilled workers (finding enough laborers), building up infrastructure, and so on will take time. The argument expands and says if the U.S. continues to push boundaries, it may even impact their access to Taiwan’s TSMC (now that Taiwan has been slapped with tariffs, with semiconductors exempt - ummm transactional much) and EU being hit - which the Dutch company ASML being the world's leading supplier of lithography systems, will they remain loyal as they have been? Even America’s most loyal Asia allies - Japan and South Korea are being hit by tariffs and are maybe maybeeee will cuddle up to China to sort out a new tri-relationship agreement (despite having had a pretty icy relationship over the last few years since the THAAD for SK and pretty much since WWII for Sino-Japan).

China Tech

The impact on China tech will be more on the e-commerce players, especially the cheaper goods platforms like PDD’s Temu and SHEIN where they were able to get around the tariff loopholes for a long time. That may end soon. Even the one dollar-worth goods may be subject to customs and tariffs going forward.

And then there’s the ongoing TikTok deal that now might be grouped into the negotiations behind closed doors. A bit of a side tangent, the rumors are now that the agreement is that the TikTok U.S. business will spin off and new U.S. investors will inject capital into this entity, according to Brad Gerstner, Founder of Altimeter, an investor in ByteDance. 60% of the business is currently held by U.S. investors, and then ByteDance, the Chinese parent company, will be able to keep <20% of the new entity. For context, the estimated valuation of the 600-800 billion dollar business of ByteDance, of which ~90% is mostly the valuation of Douyin - the Chinese TikTok, Doubao - its LLM, and Rest of World TikTok. In many ways, the U.S. business is actually quite small, so that’s why ByteDance wanted to shut down its U.S. business at one point, given the geopolitical uphill battle.

As of now, we can only wait and watch how the wider tech and AI industry implications will unfold. In a few days, I’m introducing a guest writer who has written a piece on AI x Tariffs.

For now, just all go back to sleep. Like Mel Robbins preaches, “Let them, let me”. There isn’t much we can do, so let them tariff on and let me just keep calm and carry on.

There has been a lot of scattered information that needs to be organized and dissected together to be made sense of, so today’s post is not a deep dive or deep analysis but rather a quick news update and syndication of some current happenings in the AI world and a call for action from you - readers.

Subscribe now

Shift in Mentality around LLMs in the U.S.

There seems to be a real shift in mentality about LLMs in the U.S. After Grok 3’ release two weeks ago, the conversation has shifted to as long as your model is “good enough,” then the real competition is now in actual adoption, and can you reach the right audience? It seems like there is some discourse between Meta’s grand ambition and its actual implementation. So far, it still hasn’t rolled out any functions that have brought on a significant amount of users for its AI agent. After all that I've written extensively about the U.S. leaning into enterprise and China leaning into consumer apps, maybeeeee we’re going to see them meeting each other in the middle.

On the BG2Pod podcast, the two avid investors in AI compared consumer AI adoption to what Search was in the internet era. There are 5 to 6 players who are clear leaders already, but eventually, it’s not just about better technology (when everyone reaches “good enough”) but more about distribution strength. Google’s self-cannabilism of its search engine seems like its smartest AI strategy yet. Still, the one who can better integrate the experience into their existing app will likely capture more users. Similarly to what I’ve been writing about, it might not be a winner-takes-all scenario; rather, a winner takes most, but then specialized vertical adoption of AI will really have its space to grow, too. And as much as the frontier model is a key competitive advantage, reach is key to capturing consumer AI. This phenomenon and advantage have been shown by Tencent’s adoption of DeepSeek in the Chinese market. See here. I’ve also written about how Tencent’s WeChat has its advantages, as X does, compared to some of its peers due to functional adjacency.

Even at the TMT conference held by Morgan Stanley this past week in San Francisco, Sam Altman of OpenAI himself has admitted that LLMs are becoming commoditized, and he’s exploring upstream and downstream - in applications and infra. When asked where value will accrue? He said at the top, it’s in applications, and at the bottom, it’ll be in chips, power, datacenters. Currently, in the middle layer, which includes models, it’s becoming hard to capture value in the long run since it's being commoditized, according to my industry source on-site at the event.

Manus - China’s new AI superstar?

Manus broke the Chinese internet yesterday. See RecodeAI Tony Peng’s quick update on this on Substack and Manu’s launch video here on X.

So far, I haven’t been able to test it out since it’s an “invitation-only web preview “ (neither has any of my mainstream media friends so far). So this information is all second-hand. The SCMP described Manus capabilities as follows: “can execute practical tasks such as creating a custom website through a step-by-step process, according to a demonstration video on its website manus.im.”

It has been hailed as a groundbreaking AI agent capable of independently executing complex tasks, marking a significant shift from traditional AI assistants that primarily provide suggestions or answers. In its launch video, it shows that it can screen CVs, provide stock analysis, and write a program to calculate budgets for your business, which seems very impressive.

Manus has been described as having a"DeepSeek moment" for AI agents, marking a transition from single-task execution (mostly prompt-to-answer chat) to intricate decision-making processes. Its ability to reason and then deliver complete task results positions it as a significant player in the AI agent space.

But while many people are raving about its capabilities and user experience, some are questioning why it feels more like a PR win for Manus than a technical breakthrough.

Subscribe now

The launch video shows the founder, a young Chinese man, Yichao ‘Peak’ Ji (recipient of Forbes 30 under 30), speaking perfect English and walking you through the product in English. This itself has led to some netizens asking why it was not done in Chinese if the target audience is for the Chinese audience. Alternatively, many netizens in China are saying he embodies the internationalism of China’s young entrepreneurs. So we can see that, in many ways, he’s getting more attention for who he is and his background than the product itself. The whole “国运1” narrative of China’s AI is on the rise and is fueling this frenzy somewhat - which I think is going a bit out of control…

As seasoned China tech analyst Rui Ma puts it, “Don't underestimate the power of fortuitous timing,” hinting that Manus is also riding on the success of DeepSeek to garner so much public attention.

https://x.com/ruima/status/1897760289145471073

According to China Daily’s report, the invitation codes were being resold for nearly 100,000 yuan（$13,797 on the second-hand trading platform Xianyu - which is a bit wild. People are saying that the people who have had trial codes are mostly all “self-media,” meaning frankly, they could have been paid for PR, but technical people are not getting access to it. The company’s website itself is only in English as well - a bit odd. A lot of the articles are really emphasizing that Manus is a “Chinese team,” another example of China’s AI prowess. This nationalistic angle is also quite in contrast to DeepSeek’s initial coverage, as most of the conversation was about efficiency and model capabilities. I don’t have further takes on Manus, but I will be watching the development here.

Share

A Multipolar World

These are things that are on my mind/ radar, and I hope you can contribute to them. Please leave a comment or DM me if you have any thoughts on these.

1. How has China’s “real economy” prioritization shaped its AI strategy? How will AI affect 实体经济 - tangible economy - manufacturing, real estate, etc -

2. What are some innovative ways you’ve seen vertical experts adopt AI, such as Style3D, in fashion designing

3. Can China and U.S. AI supremacy really not coexist? The AI arms race is being framed as a national security issue in the U.S., but it’s truly tough to “win” when it isn’t actually a binary thing. And it seems like more and more people are realizing that the “there could only be one winner” mentality is not conducive to or productive for anyone. We’re only at the infancy stage (actually, Brad Gerstner called it the “embryonic” stage) of AI development, and the world will only develop and adopt more. The doomers can keep going at it, and we need them for checks and balances, but I really don’t think somehow the world will just not continue in AI innovation.

   China’s official governmental response to how it views America’s efforts in ensuring China doesn’t catch up in the AI race was rather non-aggressive. Foreign Minister Wang Yi said that high walls cannot stifle innovative thinking and decoupling or cutting off chains will ultimately isolate oneself...adding that science and technology should not be used as tools to weave an iron curtain… See the full transcript provided by . (btw put together AI takeaways from China’s Two Sessions here).

   To be honest, I think the discussion should really shift to the following: Can we reach a point of international collaboration in regulating new technology or finding more efficient ways to utilize it so we don't completely destroy our earth? Wouldn’t that be much more efficient of a conversation to focus on? Can we have a multipolar world? The BG2 guys seem to think it’s quite “naive” for anyone to think that China will not achieve frontier AI simply by blocking chip exports. Mhmm, this topic deserves more time and thought. I’ll come back to it another time.

   ---

Last but not least, a little shameless self-promotional plug: check out my commentaries on China’s AI startup space and the reasons for the keen embrace of open source in the community in this SCMP article.

Hi everyone, as promised, this is Part Two of the collaboration series with on the divergence of AI monetization strategies adopted by Chinese and U.S. big tech companies. In Part One of this series, we examined how AI commercialization by big tech in the U.S. is driven by enterprise applications, leveraging the country’s strong cloud computing industry. In contrast, China's leading tech companies’ approach prioritizes embedding AI into existing consumer platforms, resulting in a distinctly different path for AI’s diffusion. In Part Two, we’ll look at potential explanations behind this divergence.

Subscribe now

TL;DR

China and the U.S. are taking divergent paths in AI commercialization, reflecting fundamental differences in their digital ecosystems and even broader economic structures:

1. U.S. Focus: Enterprise software and cloud-based AI solutions

• Leverages strong cloud computing industry

• Caters to a developed knowledge economy with a high demand for productivity tools

2. China's Approach: Consumer-facing AI applications

• Embeds AI into existing popular platforms (e.g., WeChat, Meituan)

• Driven by intense competition for user engagement in the consumer internet sector

In this article, we focus on understanding the key factors behind China's weakness in enterprise software adoption and how that has affected China’s AI monetization strategy:

• Industrial composition favoring labor-intensive sectors

• Prevalence of state-owned enterprises with unique constraints

• Availability of cheap labor, which reduces incentives for automation

• Informal work processes and blurred work-life boundaries

The divergence may impact AI's broader economic effects:

• In China, AI benefits may remain concentrated among tech giants

• In the U.S., wider adoption through enterprise services, but productivity gains are still uncertain

  Share

The divergence in AI commercialization between the U.S. and China is, in many ways, unsurprising. China’s digital economy is heavily consumer-driven, dominated by e-commerce, online payments, and social media, while enterprise software—particularly Software-as-a-Service (SaaS)—remains sorely underdeveloped. In contrast, the U.S. has a more balanced mix of enterprise and consumer-facing firms, with globally dominant players on both sides of the enterprise-consumer divide. America’s mature cloud computing sector also means that it is better positioned to commercialize AI through the many enterprise customers already using the cloud.

If AI enhances existing processes, then it makes sense that tech firms in each country would commercialize AI in ways that align with their most profitable business lines. Thus, explaining the divergence in AI commercialization ultimately comes down to understanding how the U.S. and China developed such fundamentally different digital ecosystems in the first place.

Let's start by unpacking why China does not have a robust enterprise software ecosystem.

China’s inherent weakness in enterprise software

A. Industrial Composition

The most effective way to understand this is to look at the demand-side story for enterprise software in the two countries.

What kind of firms benefit from enterprise software? If enterprise software is about driving productivity gains for white-collar workers, then it is most valuable to knowledge-intensive sectors or, more broadly, firms with a predominantly white-collar workforce. This distinction between knowledge-intensive and labor-intensive firms is critical for enterprise software because the economic composition of firms in the U.S. and China differs significantly along this axis.

The U.S. has a highly developed knowledge economy, with industries like finance, insurance, and professional services making up a large share of its GDP. Even in traditionally less knowledge-intensive sectors—manufacturing, textiles, hospitality, and food & beverage—U.S. firms tend to dominate the higher-value portions of the industry, focusing on branding, product design, and marketing rather than physical production. They operate at the two high-value ends of the “smile curve” (see graph below), where demand for enterprise software is naturally greater.

The result is that knowledge workers represent a significant portion of the U.S. labor force, with estimates indicating that there are approximately 100 million knowledge workers in the U.S. as of 2023—roughly 60–76% of the total workforce.

In contrast, Chinese firms tend to sit at the bottom of the smile curve, focused on labor-intensive fabrication. Outside of manufacturing, construction and agriculture—also low IT spenders—are top contributors to China’s economy. In these segments, white-collar productivity software has far less impact, resulting in overall resistance to enterprise software adoption.

Another key challenge for less knowledge-intensive sectors in China is the lack of high-quality datasets, which limits business’s ability to use sophisticated enterprise software. Because non-knowledge economy industries have historically underinvested in IT infrastructure and enterprise software, they operate with mostly analog, fragmented, low-quality, or even non-existent datasets.

This lack of digitization creates a major barrier when using enterprise software, let alone the customization of AI models for industry-specific applications. So, even if AI tools were available, firms in these sectors would struggle to adapt them to their workflows, let alone find ways to commercialize them.

B. State-Owned Enterprises

Some estimates also suggest that up to 28 percent of China’s GDP comes from state-owned enterprises (SOEs). SOEs are often subject to stricter data and security regulations that may restrict the use of private-sector technology. And only certain companies have the capacity (or are willing to adjust their business models) to meet these requirements.

As one-third of the economy is driven by the state, it can often move slower, too. One of the key reasons is that there are added bureaucratic practices such as chain reporting and cross-departmental oversight. In theory, the goal is to implement checks and balances to make sure individuals or teams are not misusing governmental resources. In practice, however, this often results in slower processes within these organizations. And as with many government entities, they are slower to introduce new technology, such as software, which is often seen as non-essential.

Finally, jobs are often created by China’s SOEs not for efficiency gains but for social stability purposes. In government-funded firms, it wouldn’t be hard to find people whose jobs are to literally sit on a stool and press the elevator buttons for you. Or patrols sitting in a toll office at an automatic payment parking lot. The point is not to increase efficiency but to create jobs and, thus, enhance social stability.

C. Cheap labor

China’s comparatively cheap labor costs have also created an unwillingness to pay for enterprise software. In sectors with low wages, businesses have less incentive to adopt software solutions that enhance worker productivity, as hiring additional labor is often cheaper than investing in automation or AI-driven tools.

This is particularly evident in industries like retail and customer service, where tasks that might be automated in high-wage economies are still performed manually in China. For example, rather than deploying AI-powered chatbots or automated workflow systems, many companies simply hire more staff to handle routine tasks.

The same logic applies to most back-office functions. While enterprise software in the U.S. streamlines operations and reduces white-collar labor costs, Chinese firms often rely on cheap office labor instead of software solutions. As long as labor remains relatively inexpensive, the financial justification for investing in enterprise software—and, by extension, AI-powered enterprise solutions—remains weak.

Subscribe now

D. Irregular, informal workflows

Another major challenge for enterprise software and SaaS adoption in China is the prevalence of informal work processes within many firms. To this day, many professionals still use consumer-facing apps for work, such as WeChat, to communicate with colleagues, transfer files, and run meetings, even though there are existing Slack-like alternatives such as Alibaba’s DingTalk or ByteDance’s Lark that have rich collaborative features.

Part of the reason for this is that the boundaries between work and life are often blurred in Chinese workplaces.1 In the U.S., it would be uncommon to add your colleagues on Instagram or get their phone numbers unless you're actually friends. But in China, everyone is defacto connected on WeChat. This blurring of lines means that employees never quite “log off.” In many workplaces, bosses expect to be able to reach their employees 24/7 (something Grace experienced firsthand when working in China). And not promptly responding to your boss is practically unthinkable for most Chinese workers - given that they know you are always on WeChat and not checking your emails is not an excuse.

From a management perspective, this access to employees can be a competitive advantage since it allows employers to squeeze workers of greater productivity. However, the blurring of work and life also makes it extremely challenging for other software to displace WeChat as the defacto productivity platform.

Unlike in the U.S., where standardized workflows make it easier to integrate software solutions for automation and efficiency, many Chinese companies operate with highly customized or ad-hoc procedures that are difficult to codify into software. Decision-making can often be relationship-driven rather than process-driven, and informal workarounds frequently replace structured workflows. This makes enterprise software far more challenging to adopt, as rigid software solutions struggle to accommodate the fluid, and often informal nature of many business processes in China.

In addition, during the early 2000s in China, pirated software (as well as DVDs and CDs) was so rampant that it became a kind of pseudo-legal operation in the grey area. Small shops blatantly sold pirated software like Word, Excel, and Photoshop. Even legitimate companies would sometimes choose this much more cost-effective option. And where a company employer wasn’t willing to spend on software, individual employees might go out on their own and purchase pirated versions.

Competition in China’s consumer internet

Putting aside China’s lack of competitiveness in enterprise software, Chinese tech firms have largely leaned towards consumer-driven applications—even when it comes to AI—because of the highly competitive structure of its tech industry.

Compared to U.S.-made apps, which focus more on providing specific functions or utility to users as their core mission, Chinese tech giants prioritize capturing and controlling the entirety of a user’s online activity because more activity equals more engagement, which equals more opportunities for monetization.

While their entry point is typically through a single service—WeChat with messaging, Meituan with food delivery, Trip.com with travel—their long-term strategy is to expand their ecosystem and lock users within their digital walled gardens. This is the driving force behind China’s super apps. Where U.S. firms tend to stay in their lane—Uber in transportation, Airbnb in tourism, etc.—Chinese companies continuously integrate payments, social media, live streaming, and now AI chatbots to deepen user engagement and prevent churn.

This intensely competitive environment means that China’s tech giants are constantly vying for market share. Each giant does whatever it can to keep users locked into its ecosystems. When new technology emerges, this urgency is amplified. Each player hops on the bandwagon, fearing that inaction will result in losing users to the competition. [Grace wrote about this phenomenon in her Tencent integrating DeepSeek article.]

In this way, AI is seen as an immediate competitive advantage in the battle for 流量 (traffic/engagement), which remains the central currency of China’s consumer-centric internet economy.

Conclusion

As we’ve shown, China’s inclination to commercialize AI through consumer platforms rather than enterprise software is driven by deeper structural factors, not just a matter of preference. The socioeconomic fabric of the two economies is vastly different. While there has been overlap in innovation and productivity drivers over the last three decades of the consumer internet era, the age of AI is pushing them down distinctly divergent paths.

Why does it matter whether AI is commercialized through existing consumer platforms or enterprise software and SaaS? The answer lies in productivity gains. As with any general-purpose technology—if AI can truly be considered one—the key to its economic impact depends on how widely it is diffused. For AI to drive meaningful productivity growth, it must be adopted broadly across industries, not just concentrated in the hands of a few dominant firms.

In China, where AI diffusion is happening primarily through existing digital platforms, its benefits may remain highly concentrated among the tech giants that control the internet economy. If AI adoption is limited to enhancing consumer engagement rather than improving business operations across sectors, its impact on overall productivity may be minimal.

Even in the U.S., where AI is being commercialized through cloud-based enterprise services, it is still unclear whether adoption will translate into actual productivity gains. The fact that companies are buying AI services from Microsoft, Amazon, and Google signals that they have the capital to invest and are willing to experiment, but whether these investments will drive efficiency improvements or sustained economic transformation remains an open question.

OK, SO…

Last week, DeepSeek held an “open source week” to disclose details about its low-cost, high-performing models. However, the event largely went unnoticed by the media—at least not discussed in terms of its implications. It did, however, make waves in the open-source community.

Over the course of six days, the company released tools that aim to redefine how artificial intelligence is developed and deployed. These tools promise to make AI faster, cheaper, and more accessible, not just for tech giants but also for smaller players.

Subscribe now

Many people have asked me why DeepSeek is embracing Open Source or why Chinese tech seems to be more keen on the open-source model. I think there might be a few reasons.

1. For DeepSeek, it really is a personal philosophical choice for CEO Liang Wenfeng.

2. For China, in some ways, it could be a bit more personal - a chip on their shoulders. For decades, the world has pointed fingers at China for “copying” innovation, for better or worse, often offering cheaper and better-copied goods. However, for the generation of entrepreneurs born in the 80s and 90s, there has been a desire to really prove to the world that Chinese companies can, in fact, innovate. And we’ve seen that especially come through the internet era with TikTok, Shein, DeepSeek, Unitree, etc. And Kevin Xu coined it quite eloquently - the rise of the “开元情怀 open-source zeal” - where it is more exciting to be quoted/ used by developers and businesses outside of China than making money off of the project.

3. On top of that, the overly heated fear around China’s tech innovation is partially driven by real concerns of state security in the U.S., as obviously, whichever country can have a grasp on the most frontier technology will also mean having a global influence on economies, societal expectations, and military prowess. But that’s only half of the argument, I think. The other half that feeds into the fearmongering narrative is also partially fueled by the fear of competition and, again, that, in a way, drives firms to want to show off their capabilities, maybe. Nonetheless, these have been the dual drivers of fearmongering about Chinese tech in the U.S. around consumer apps like TikTok, Shein, and RedNote.

4. I think the open source model is widely accepted and preferred because of the ecosystem’s monetization strategy. and I wrote about the divergent commercialization strategies the U.S. and China are adopting—specifically, why many in China expect free LLM and commercialization is at the application end.

5. [added after initial publishing] Adding to the initial thoughts, I am including another hot take by Kevin Xu that I think is super interesting. He wrote, “open source is the equivalent of soft power in tech,” in a way, an ultimate flex.

Anyway, it’s interesting to see the Chinese tech and AI companies really getting a boost to the front row again after a dull few years for the Chinese internet sector.

What Is DeepSeek Open Infra Week?

Day 1: FlashMLA

• What It Is: FlashMLA is like upgrading your car engine to go faster without burning extra fuel. It’s a tool that helps AI systems process information more efficiently on GPUs (the hardware behind most AI models).

• What It Does: It optimizes how GPUs handle variable-length data (like sentences of different lengths), making them faster and more memory-efficient.

• Why It Matters: This means companies can run AI models more quickly without needing expensive hardware upgrades. For users, this could translate into faster responses from AI systems like chatbots or recommendation engines.

Day 2: DeepEP

• What It Is: DeepEP improves how computers “talk” to each other when training massive AI models. Think of it as upgrading from a bumpy dirt road to a smooth highway for data transfer.

• What It Does: It speeds up communication between GPUs and uses a lightweight data format (FP8) to make everything run more smoothly.

• Why It Matters: Training large AI models often takes weeks or months. By reducing delays, DeepEP makes this process faster and cheaper, helping smaller teams compete with industry giants.

Day 3: DeepGEMM

• What It Is: DeepGEMM is like installing a super calculator in your computer. It handles the heavy math that powers AI systems with record-breaking speed.

• What It Does: It processes calculations 10 times faster than older tools while using less energy.

• Why It Matters: Faster math means quicker training for AI models and real-time applications like voice assistants or translation tools that respond instantly.

Day 4: Optimized Parallelism Strategies

• What They Are: Two new strategies—DualPipe and EPLB—designed to help GPUs work together more efficiently.

• What They Do: DualPipe ensures GPUs don’t waste time waiting for instructions, much like an assembly line where each worker knows exactly what to do next. EPLB balances workloads so no GPU is overwhelmed or underused.

• Why They Matter: These strategies allow companies to train massive AI models faster and at lower costs, opening the door for even more complex and capable systems.

Day 5: Fire-Flyer File System (3FS)

• What It Is: 3FS is like upgrading your computer’s storage system to handle huge files at lightning speed.

• What It Does: It allows multiple parts of an AI system to access data simultaneously without slowing down.

• Why It Matters: Faster data access means shorter training times for AI models. This makes it easier for researchers to test new ideas quickly.

Day 6: DeepSeek-V3/R1 Inference System

• What It Is: A system designed to make using AI models (inference) cheaper and faster while maintaining high quality.

• What It Does: Handles millions of users with minimal hardware by dividing work across GPUs smartly. Achieves incredible efficiency—serving tens of millions of DAU’s inference demand (600 billion input tokens and 150 billion output tokens per day) with just ~2,200 GPUs.

• Why It Matters: This system drastically reduces the cost of deploying advanced AI models, making them accessible even for smaller businesses. For example, DeepSeek outperformed expectations by serving massive user bases with far fewer resources than competitors.

Industry Implications

DeepSeek’s Open Source Week went largely under-covered by Western media this time. Still, it actually is quite important as it could redefine the economics of AI, the competitive landscape, and the pace of innovation across industries. Since the release of R1, we have known that we can actually cut costs in AI inference.

Now, with additional tools like FlashMLA and DeepEP, running large models will be much more accessible for startups and non-big tech companies. These tools are designed to optimize the performance of existing hardware, reducing costs while speeding up operations. And ultimately, consumers are the biggest beneficiaries, as this means cheaper and faster AI-powered services—like smarter chatbots, better recommendation systems, or even real-time language translation tools. This is a huge step toward democratizing AI infrastructure adoption acceleration across industries and potentially drawing in more AI users.

Economies of Scale

Due to economies of scale, DeepSeek’s innovations will actually amplify big tech’s reach even further. As these companies already serve massive user bases across various touchpoints, it means that with steady high demand for compute usage, you do not need to fear any spikes to maximize efficiency in their compute clusters.

For example, DeepSeek demonstrated that with a large enough user base, you may only need to provide 30% extra capacity instead of 3-5x the computing power that smaller companies would require. This “economies of scale” effect means that as these giants grow larger, their cost per user shrinks further, which further brings down the price of consumers using AI, cementing their dominance in consumer-facing AI services. This will probably be especially useful for Chinese big tech companies as they try to find ways to monetize AI in consumer-facing application use cases. This creates a virtuous cycle: lower costs lead to more users, which leads to even greater efficiency and cost savings.

However, as mentioned, this will help further democratize AI adoption. While big companies benefit from economies of scale, DeepSeek’s commitment to open-source its technologies does level the playing field for smaller players. Startups and mid-sized companies will be able to access cutting-edge tools without needing to invest millions in proprietary infrastructure or expensive GPUs. Hopefully, this will spark a wave of innovation from smaller teams who now have access to the same level of efficiency as industry leaders.

Final Thoughts

By serving tens of millions of daily active users (DAUs) with just 2,200 GPUs while processing 750 billion tokens daily, this is a fraction of the hardware competitors like OpenAI or Google are believed to use for similar workloads. What it means for the broader industry is that companies can no longer afford to be inefficient with their computing resources when DeepSeek has demonstrated that it can be done with much less. Value per GPU is the new key indicator added to measuring how good a model is.

As I’ve written before about - Jevons Paradox - as compute become cheaper and more efficient, demand will inevitably grow. So that brings us to one of the most provocative questions raised by DeepSeek’s innovations, which is whether we’re overestimating how much compute power is actually required today. Let’s break it down:

1. DeepSeek used just 2,200 GPUs to process 750 billion tokens daily while serving tens of millions of DAUs.

2. Extrapolating from this efficiency: 200,000 GPUs (roughly 5% of NVIDIA’s estimated 4 million Hopper GPUs sold in 2024) could theoretically handle 75 trillion tokens per day.

3. For context: The Information reported that OpenAI’s total daily API token usage as of January 2025 was around 2 trillion tokens per day—a fraction of the 75 trillion number above.

This past week seemed to further the narrative DeepSeek started with its R1 release in January. By proving that frontier performance can be achieved with far fewer resources than previously thought possible, DeepSeek has set new expectations for what efficient AI infrastructure looks like. We’ve heard this so many times now that necessity is the mother of innovation, but truly, this might not have happened if DeepSeek had the resources the big techs did—that is, capital and access to GPUs (in the U.S.).

To be honest, I don’t think the most pressing question is really about whether we’ll need more computing power right now—it’s how to use energy more efficiently to reach AGI goals while being mindful of our environmental and climate challenges. Jevons Paradox ensures that as we optimize today’s resources, tomorrow’s demand will only grow larger. But thanks to DeepSeek’s innovations, we now know how far those resources can stretch—and how much more is possible when efficiency takes center stage.

DeepSeek’s R1 release last week sent shockwaves through the U.S. power and technology sectors, causing significant declines in stock prices and raising doubts about future energy demands for AI infrastructure and the direction of AI capex.

Specifically, many power companies heavily linked to the surge in data centers within the tech industry experienced substantial drops in their stock prices. Despite being among the top performers in the S&P 500 earlier this year:

• Constellation Energy: Down over 16%

• Vistra Corp: Fell by over 16%

• GE Vernova: Dropped approximately 18%

• Talen Energy: Decreased by more than 15%

However, Microsoft CEO Satya Nadella was the first to say that this might be the wrong reaction to increased efficiency.

This was first published as a guest post for . Please don’t forget to like, share, and subscribe if you find this valuable.

Subscribe now

Market Myopia: Misreading the Efficiency-Demand Cycle

Following the release of DeepSeek R1, energy stocks plunged, revealing a potential short-term misreading of energy economics or, more accurately, knee-jerk panic. Like 19th-century observers who thought coal demand would fall with the steam engine (by 100x, based on the JPM Asia DeepSeek report), sometimes markets conflate efficiency with sufficiency.

In a recent report, UBS analysis indeed confirmed this oversight: “While DeepSeek’s 97% cost reduction in AI training (vs. GPT-4) appears deflationary, it mirrors the internal combustion engine’s 100x efficiency leap, which drove 223x global oil demand growth post-1900.”

Here's why it's not as simple as "more efficient AI = less energy use":

1. Cheaper AI: First off, AI just got a whole lot more affordable. We're talking about a 97% price drop for using AI services, with costs now at $2.20 per million tokens compared to GPT-4's $81, according to JP Morgan's AI Adoption Index. This means tons of medium-sized businesses that couldn't afford AI before are now jumping on board - about 73% of mid-market firms, to be precise. It's like when smartphones suddenly became affordable - everyone wanted one.

2. Video: Next, consider how much of the internet now is in video format. According to a Bank of America Report, it's a lot - about 80%. The thing is, AI needs way more juice to handle video compared to just text - 3 to 5 times more compute power, as noted in JP Morgan's Compute Multiplier Report. So, as more companies use AI for video (think TikTok filters or YouTube recommendations), the energy demand will skyrocket. We're looking at a potential $47 billion latent demand pool here.

3. Edge AI: Lastly, we're putting AI everywhere—not just in big data centers but into our phones and street lamps. This "edge AI" is excellent for speed, but it means we need to upgrade a lot of equipment. JPMorgan reports that on-device AI processing boosts 5G base station power draw by 40–60%, necessitating 22 million new global nodes by 2030. Each of these uses more power than before.

So, while AI is becoming more efficient, we're using it for so many more things in so many places that total energy use is still set to climb. It's a bit like how fuel-efficient cars led to more driving, not less gas use overall.

For additional context, Scott Chamberlin, who builds tools to examine the environmental costs of certain digital activities, tested how much energy a GPU uses for a DeepSeek query and found that a 1,000-word response from the DeepSeek model took 17,800 joules to generate. As reported by MIT Technology Review, that is about what it takes to stream a 10-minute YouTube video. That’s because DeepSeek’s answers tended to be longer than those of other chatbots. So, although the model is more efficient, it didn’t decrease the overall amount of energy used, which is not like what people assumed.

Efficiency, in energy systems, is demand’s accelerant.

Nadella’s economics lesson wasn’t just academic. It foreshadowed Microsoft’s $13 billion AI revenue run rate, which was disclosed days later during its FY25 Q2 call — up 175% year over year. His framing mirrors Meta’s playbook: Cheaper inference costs don’t reduce energy demand because AI can be embedded into every platform (which is cool and scary).

Suppose the R1 efficiencies are realized in all new models. The hyperscalers' AI Capex and energy usage budgets will likely remain strong in the long run. Because…

“When AI gets cheap, you know what’s going to happen? There’s going to be a lot more AI,” said Travis Kalanick, founder of Uber, on the All In Podcast. “I think the price elasticity on this one is actually positive. So, as the price goes down, the revenue and usage will go up (through the roof).”

AI isn’t inventing the wheel here. In a 2010 New Yorker article titled “The Efficiency Dilemma,” David Owen wrote, “The first fuel-economy regulations for U.S. cars—which were enacted in 1975, in response to the Arab oil embargo—were followed not by a steady decline in total U.S. motor-fuel consumption but by a long-term rise, as well as by increases in horsepower, curb weight, vehicle miles traveled (up a hundred percent since 1980), and car ownership (America has about fifty million more registered vehicles than licensed drivers).”

For the Econ nerds, I highly recommend the New Yorker article. It goes into the background of how the economists of that century were related and how John Maynard Keynes “found out about” William Stanley Jevons and thought of him as “one of the minds of the century.”

Jevons Paradox 2.0: The Three-Phase Demand Engine

So this leads to Jevons Paradox, which I first touched on in the piece about R1’s industry implications.

What is the Jevon Paradox? The Jevons paradox occurs when the effect of increased demand predominates, and the improved efficiency results in a faster rate of resource utilization. (Wikipedia)

In layman’s terms, when technology becomes cheaper, usage explodes—even if it feels just "good enough." For example, in the 1800s, burning coal became more efficient, but factories used more coal to power new machines instead of reducing demand.

Similarly, suppose AI becomes more affordable (DeepSeek R1 is currently much cheaper than other frontier models). In that case, we can expect startups, governments, consultants, hospitals, and even schools to use far more AI than they do today.

Just as Anthropic’s co-founder Dario Amodei wrote on his blog, “⁠because the value of having a more intelligent system is so high,” it will lead to “companies to spend more, not less, on training models.” So if companies get more for their buck, then they will have more reason to invest more in AI and thus energy. Adding that, “the gains in cost efficiency end up entirely devoted to training smarter models, limited only by the company’s financial resources.”

Share

The English economist and logician William Stanley Jevons’ 1865 insight—that efficiency gains paradoxically increase resource consumption—now basically governs AI’s infrastructure buildout through three evolutionary phases:

Phase 1: Overbuilding Precedes Overconsumption

• In 1999, telecoms laid 28 million miles of unused “dark fiber,” which was mocked as wasteful until streaming video exploded bandwidth demand by 10,000x. Today’s parallel: OpenAI’s $500 billion Stargate Project targets 100-trillion-parameter models that are not yet needed but essential for 2030’s autonomous supply chains, which require real-time global logistics simulation.

Phase 2: Compression Fuels Complexity

• DeepSeek R1’s 45:1 efficiency gain over GPT-4 follows this pattern: By making real-time 4K video translation commercially viable, it could triple AI’s inference workloads by 2028 (according to McKinsey’s The State of AI in 2024 report).

Phase 3: Infrastructure Demands Better Infrastructure

• Just as broadband required five times more network hubs than dial-up, Meta’s $65B AI capex surge funds edge devices (smart glasses, AR headsets) that will necessitate 22 million AI-optimized 5G nodes. These nodes consume 40–60% more power than legacy systems, creating a self-reinforcing cycle in which more ingenious devices demand smarter grids.

Hyperscalers’ $1 Trillion Commitment: Building the Intelligence Furnace

For months, we’ve examined the energy needed to power data centers. We've also examined China's and America’s competitive advantages in driving top-down supply build-up. We’ve also explored renewable options, especially nuclear energy (SMRs), as the U.S. currently faces an energy shortage.

And those explorations weren’t completely wrong. According to an industry expert, the AI industry is indeed facing a major energy challenge, and U.S. cloud service providers (CSPs) are turning to nuclear energy, with its impressive 90%+ efficiency compared to solar's 50%, to meet ambitious carbon reduction goals (especially under Biden’s administration).

There have also been suggestions that since in the U.S., CSPs' electricity demands have already been outpacing grid capacity, driven by the rapid growth of Edge AI and increasingly complex models. This trend points to a future where AI compute and energy are bundled as a utility service, with customers paying based on their AI usage.

For consumers, the future of spending on AI could be like paying for water or electricity in our homes—the more you use, the more you pay. New technology often obviates the old way of doing things entirely.

And for hyperscalers, despite initial market reactions, after all this DeepSeek fuss, we aren’t seeing any of them lowering their AI Infrastructure spending in the near term, nor are the two superpowers lowering energy demand expectations.

1. The Stargate Project

OpenAI and SoftBank’s $500 billion infrastructure initiative, signed last month, symbolizes the most significant private-sector tech investment in history. The deal isn’t about raw compute but power grid dominance. It requires 12GW of sustained electricity (equating to Denmark’s grid capacity), and the demand size is insane. And Stargate’s architecture includes modular nuclear reactors to bypass overloaded utilities. Confirming what Microsoft’s CTO Kevin Scott said: “Training clusters are becoming national infrastructure.”

2. Meta’s Unwavered Capex Plan

Just days after DeepSeek’s release, Meta’s Zuckerberg raised 2025 AI spending to $60–65 billion (+50% year-on-year) during the latest earnings call, declaring, “I believe scaling up infrastructure is still an important long-term advantage.” This timing underscores a critical insight: far from curbing investments, efficiency breakthroughs like DeepSeek validate hyperscalers’ infrastructure arms race, unlike what many are panicking about.

Share AI Proem

Conclusion: The Intelligence Inflation Era

Last week’s energy and power sectors sell-off mirrors skeptics in 1865 in Jevons’ book The Coal Question, who thought more efficient coal would reduce coal demand (spoiler alert, it increased massively). DeepSeek’s efficiency leap isn’t a singularity—it’s the spark for AI’s future use-case explosion across three frontiers:

1. Energy Arbitrage: At $0.02/kWh parity (Goldman’s 2026 forecast in its AI’s Energy Tipping Point report), AI inference becomes cheaper than human labor for 23% of service jobs.

2. Systemic Complexity: According to Reuters, Walmart’s AI supply chain processes 6x more data than its 2023 system but uses 9x more energy—a Jevons tradeoff executives accept for 40% fewer stockouts.

3. Infrastructure Inertia: Once hyperscalers deploy 22 million AI nodes, their operational costs create irreversible momentum, similar to gas stations cementing fossil fuel dominance after the 1900s.

Given the short-term panic caused by energy stock sell-offs, some investors might forget that efficiency gains have always been the demand’s accelerant, not its limit.

The questions to ask ourselves are: 1) How can we be more mindful of preserving our earth while we KNOW that we need more energy for increased AI demand? 2) AI demand will only increase, as internet demand did. It will be seamlessly integrated into our daily lives and eventually normalized. How do we regulate it better?

The defining resource of the 21st century may not be about who controls artificial intelligence or computer programs but who has access to and can control the electricity needed to power these technologies.

Subscribe now

Additional thoughts:

In the book Electrifying America: Social Meanings of a New Technology, 1880-1940, which many AI industry analysts have recommended recently, author David E. Nye highlighted key points about electricity, many of which can be applied to AI and its impact on society. He wrote that:

1. Electricity touched every part of American life, becoming an extension of political ideologies and shaping the image of the modern city.

2. Electrification was expressed in utopian ideas and became a focal point of various art forms and moral debate.

3. Electricity completely redefined domestic work, agricultural produce, and what it meant for efficiency for humanity.

4. Despite social/ safety/ ethical concerns, we humans only wanted more and more of it. In many ways, our discussions around AI have been similar.

Innovation that increases productivity isn’t new; we have seen it throughout history. When electricity became more accessible, our expectations for productivity only increased. The rationale is that if you increase the productivity of anything, you effectively reduce its implicit price because you get more return for the same money—which by default means that the demand will go up.

Nonetheless, given that the newsletter is called AI Proem, I feel I need to provide some analysis of what this breakthrough means for the LLM business model.

Don’t forget to like, share, and follow if you enjoy this piece. Thank you.

For the founder Liang Wenfeng’s background and its V3 innovation, see here.

[For context and other perspectives, go to the very bottom. I’ve listed out the various opinions on DeepSeek the interweb has to offer]

Subscribe now

DeepSeek's AI Breakthrough Sends Shockwaves Through Tech Markets

Headlines have mostly zoomed in on the fact that it only cost DeepSeek $6 million to train its latest R1 model, but it’s a little bit deceptive (which I’ll explain further below). What is undeniable is the shockwave R1 has sent through Big Tech. If better models can be built smarter and cheaper, what happens to the trillion-dollar AI infrastructure we just wrote about, the GPU arms race, and Silicon Valley’s “bigger is better” dogma?

Chamath Palihapitiya is one of the first to call it out as it is, “ AI model building is a money trap.”

In a seismic shift that's rattled Silicon Valley, DeepSeek's cost-efficient AI breakthrough has triggered a massive tech sector selloff. Nvidia alone shed a staggering $600 billion in market value on Monday alone.

The Chinese startup that put Chinese AI on the global map with its V3 model, released in December 2024, has done it again. This time, it has released an R1 model that can achieve state-of-the-art AI performance at just 1/20th to 1/40th the cost of U.S. models. This has made global investors question the narrative that Big Tech has been telling: more money = better infra = better training = better AI.

Microsoft CEO Satya Nadella acknowledged DeepSeek's "super impressive" efficiency gains and urged the industry to take Chinese developments "very, very seriously." This revelation is forcing cloud giants like AWS and Microsoft Azure to reconsider their data center expansion plans, potentially spelling trouble for high-end GPU makers.

Many say DeepSeek's success using restricted H800 chips (rather than top-tier H100s) suggests a future where premium GPU demand might soften considerably (uh-oh, Jensen). While this spells uncertainty for traditional semiconductor leaders, it's creating opportunities in adjacent sectors like edge AI and automotive chips. For the AI industry, the message is clear: the AI gold rush is entering a new phase where efficiency, not just raw computing power, will determine market winners.

As one Wall Street analyst put it: "DeepSeek didn't just build a better AI model - they've forced everyone to rewrite their playbook for the next decade of tech investment."

Share

Evolution of My Thinking & Context

About five months ago, we started with the thesis that the more AI we need, the more GPUs we will purchase, and the more power the hyperscalers will consume. The other half of the arms race was the ability to ramp up AI infrastructure quickly. Then, we re-assessed the AI Capex when o3 came out with its inference-time compute workings, especially when DeepSeek released its V3.

What R1 is doing now is that it’s “undo(ing) the o1 mythology,” wrote Stratechery. The tech analyst puts it eloquently: now, with the open weights, it means that instead of paying OpenAI (hundreds of dollars) to get the reasoning, you can run R1 on the server of your choice or even locally for free.

Various industry leaders have acknowledged the technological breakthrough that has led to a systematic paradigm shift in the AI value chain.

What were the Breakthroughs?

In particular, the R1 had three major innovation breakthroughs:

1. Mixture of Experts (MoE) architecture: The breakthrough that was introduced in V2 allows the model to activate the “experts” that are necessary, but this implication wasn’t apparent until V3 (released during Christmas break). You can understand this as DeepSeek optimizing their model structure and infrastructure since they’re limited to using H800s and not H100s.

Think of MoE as a Michelin-starred kitchen where each chef specializes in one dish. When you order pasta, only the pasta chef works; the sushi and dessert experts stay idle. Similarly, R1’s AI activates only the “expert” neurons needed for a specific task, slashing compute costs by 70% or more.

Why it matters: Traditional AI models are like gas-guzzling SUVs. MoE turns them into electric scooters—nimble, efficient, and perfect for narrow tasks.

2. AI that teaches itself: Training most AI models requires armies of human annotators to label data. DeepSeek let “AI figure it out.” They used pure reinforcement learning, which lets the model learn through trial and error, like toddlers learning how to build Lego or stack blocks. The machine generates answers, gets feedback (e.g., “This math solution is wrong”), and iterates until it nails the problem.

Why it matters: This cuts costs and avoids biases often shown by humans. The caveat is that it can be riskier and more prone to the domino effect of mistakes. Letting AI self-teach could lead to really whacky “answers.” But DeepSeek seems to have worked on this front and tamed the chaotic hallucination results.

3. Model distillation: DeepSeek trained a massive “teacher” model, borrowing tricks from OpenAI’s models, and then distilled its knowledge into smaller, cheaper “apprentice” models (R1). Imagine breaking down a 1000-page textbook into 10-page cheat sheets by keeping only the core key formulas and concepts and removing the anecdotes and nice-to-haves.

Why it matters: R1 delivers ~ 90% of the performance of models 10 times its size at a fraction of their costs. The reduced model size allows you to run the DeepSeek model locally already. In a few years, it is very likely that models can be run on consumer devices like our laptops or smartphones.

Subscribe now

What is This $6 Million Myth?

I think it was a very smart PR move. Some people are saying it was a big F-U to Stargate’s announcement, but honestly, the founder doesn’t seem to be dabbling in geopolitical play. Liang is just a nerd with grit. He probably just released the coolest new model whenever it was ready. In fact, the model was released one day before the Stargate announcement; it just wasn’t picked up by Western media until later.

The technical paper published by DeepSeek says that the $6 million does not include “costs associated with prior research and ablation experiments on architectures, algorithms, and data.” The $6 million covers ONLY the final training run on Nvidia’s H800 chips. At the same time, total cost, including prior research, has been estimated to come to “hundreds of millions” based on GPU infrastructure cost.

Why it matters: Let’s say the total cost is ~$500 million. Even at that, DeepSeek’s spending is still a fraction of what OpenAI has spent. OpenAI has raised a total of $16.6 billion in funding to date and has burned through most of that on the training and inference of all its models.

What’s mind-blowing is that despite not having access to the most advanced chips due to the U.S. chip ban, DeepSeek was able to innovate and reach the frontier model performance with H800s. Perplexity CEO Aravind Srinivas, on CNBC, said, "Necessity is the mother of invention. Because they had to figure out workarounds, they actually ended up building something a lot more efficient."

What Gavin Baker, CIO of Atreides Management, LP, pointed out also made sense to me: "What’s the point of restricting access to cutting-edge GPUs but not doing anything about China’s ability to distill cutting-edge American models? Why buy the cow when you can get the milk for free?”

So, I don't know. Maybe it’s better to work together or at least start competing on innovation.

All we can know for definite is that the previous AI industry playbook—” spend hundreds of billions to stay ahead”—is crumbling.

Industry Implications: Jevons Paradox?

What is Jevon’s Paradox? The Jevons paradox occurs when the effect of increased demand predominates, and the improved efficiency results in a faster rate of resource utilization. (Wikipedia)

In layman’s terms, when technology becomes cheaper, usage explodes—even if it feels just "good enough." For example, in the 1800s, burning coal became more efficient, but instead of reducing demand, factories used more coal to power new machines.

Similarly, suppose AI becomes more affordable (DeepSeek R1 is currently much cheaper than other frontier models). In that case, we can probably expect startups, governments, consultants, and even schools to all use way more AI than they are using today.

But for Jevon’s Paradox to kick in, we need real AI applications, killer apps that will reshape the way we live and work. Just like in the Internet era, we had Google, Facebook, Amazon, etc. Today, we are still waiting for AI’s Google moment (Microsoft’s Co-pilot clearly didn’t cut it…). In terms of applications, I think Chinese tech companies may actually have an edge in consumer-facing application innovation — see why here. Also, check out Bytedance’s Doubao (China’s ChatGPT)’s exploding usage already.

Model Commoditization and Cheaper Inference: Big Tech and AI Infrastructure

I’ve previously written about hyperscalers feeling FOMO and just chucking money at AI, thinking it’ll help increase their competitive advantage. But the AI spending spree might have just hit a wall. While it might not be bad news for big tech in the long run, they definitely need to reconsider their approach to AI investment in the near term.

In the near term, DeepSeek’s R1 has triggered an existential crisis for frontier AI companies (OpenAI, Anthropics) and hyperscalers (Microsoft, Google, Meta, Amazon). These firms may all need to rethink their spending on AI and may be receiving questions on whether they have wasted hundreds of billions of dollars.

DeepSeek rewrote the scaling law, which was simply more money => more compute => bigger model => better model. Suddenly, the game isn’t about who has the deepest pockets but who has the cleverest or savviest innovations. The new scaling law now includes an added element of human innovation and ingenuity.

For chipmakers and energy companies — the assumption has been that better models require more chips and thus require more energy, so chips and energy demand would just keep increasing as models become bigger and bigger. However, as the industry rethinks near-term training capex, hyperscalers may delay or shrink data center projects, hitting chipmakers like Nvidia and power providers like Vistra.

In the long run, model commoditization and cheaper inference could be good news for Big Tech and start-ups. This will significantly decrease the spending required by hyperscalers on data centers, energy, GPUs, and all the AI infrastructure surrounding them. So although on the surface, Meta (Llama), Microsoft, Amazon, and Google are getting whipped in the public sphere on why they were not able to produce something as advanced as DeepSeek’s model with a much bigger budget, in the long-term, the big boys might not be big losers after all. Lowering the cost of training and inference will increase the ROI on AI.

On the old scaling law, one interesting question is, was this all a big capitalist scam? Was the trade-off of energy/resources/water/environment for better AI always just a way to hoard more capital? See veteran AI journalist Karen Hao’s X thread, which is very insightful.

Ultimately, DeepSeek’s R1 is now competitive with OpenAI’s o1, and we see Sam taking to X and going into full combat mode.

Additional Thoughts

1. Another reason this DeepSeek release has received so much media frenzy and international attention is because of the bigger shocker that legitimate innovation is coming out of China when the headline or stereotypes have always been “China is good at copying.” This is a wake-up call for the U.S. AI industry, that a real competitor is quickly catching up.

2. Next is the big debate that this is a win not for China but rather should be seen as a win for open-sourced vs. closed. The open-source nature of DeepSeek, which the founder has said was a decision based on his personal philosophy but also for collaborations and talent attraction. Whatever his reasoning is, DeepSeek benefits small and medium-sized businesses and consumers the most – any software, internet company, or business globally that wants to integrate AI into its workflow. Basically, people who previously wouldn’t or couldn’t pay to try out the latest frontier model can now have a crack at it.

I’ve quickly listed out the various angles/ opinions on DeepSeek’s R1 release below. The most common fearmongering theme by mainstream media and some China hawks has been, “OMG. This is when China’s AI takes over the US, and the CPC will take all our data.” Sure, there are geopolitical implications, as we have also covered, but this really goes beyond that. I guess it’s fine to call it the “Sputnik moment.”

1. Yann Lecan led the more nuanced comparison, saying that this “win” reflects open-source's victory over closed-source. (echoed by Chamath, see X)

2. Veteran AI journalist Karen Hao’s take is that this shows the billions and billions “needed” by LLM companies were, in a way, a capitalist scam, maybe not a scam, but for personal gain reasons. The tradeoffs of energy consumption/ climate damage/ resource waste for better AGI were a narrative, not the truth. (which I touched on as well here)

3. See ’s piece that answers some of the FAQs on this DeepSeek R1 release.

4. And to learn more about DeepSeek R1’s specific specs, check out Geopolitics’ post.

5. I wrote about the founding background here previously.

6. There have been talks about this reflecting China’s innovation prowess, but writes an interesting angle here. DeepSeek is the anomaly, unlike other Chinese market leaders, such as EVs or renewable energy, which were driven by top-down policy and capital and talent encouragement.

7. This interview with Lennart Heim and Sihao Huang is also very good as it provides context on the issues that the headlines have missed, such as: a) DeepSeek trained on H800, which were designed to circumvent the October 2022 controls but are still very high performing chips. b) The real export restrictions on advanced chips started in October 2023. c) Overall, machine learning algorithms have become cheaper over time. d) Check them out. I cannot judge the technicalities well, but I thought it provided some extra color.

8. Perplexity’s co-founder, Aravind Srinivas, went on CNBC to talk about the meaning of DeepSeek’s innovation. Various other industry leaders expressed that the U.S. chip ban seems to have forced Chinese scientists to be more creative and build something more efficient. This milestone success has allowed Deepseek to be embedded into the global AI infrastructure.

9. There has also been some interest in the talented engineer Zizheng, who interned at NVIDIA in the summer of 2023 and decided to join DeepSeek.

Last but not least, for Funsies

We also played around with DeepSeek last night. See the thread and how its brain works behind the scenes.

I will work on a piece on what I think the potential business and market implications are. But first, a quick TL;DR: 1) Will AI capex go down? Siemens Energy (which sells gas turbines for power generation) sank as much as 20% today. 2) ASML’s stock prices crashed as much as 10%. 3) The Nasdaq 100 futures dropped, and Nvidia dropped close to 10% pre-market. This is probably only the beginning of a correction to the hype.

In addition, the DeepSeek R1 release has made people question whether it is necessary to have advanced chips for better models and the effectiveness of the U.S. chip export bans on China—which the founder of DeepSeek has explicitly said before. They would prefer they had more advanced chips; it’s just that they were forced to figure out a more cost-effective and technically efficient way using less advanced chips. Anyway, I will look into this more and share a few thoughts. (although there really are enough thoughts floating around the internet already)

Again, please hit like, share, or subscribe if you enjoyed reading this/ found it amusing/valuable. Or DM me and tell me it’s absolutely crap and how I can make it better. Honesty is always appreciated.

Subscribe now

There has been a heated debate about whether investors and hyperscalers should continue to chuck billions at further training computing or focus more on inference computing. While OpenAI does not disclose its inference demand directly (it does disclose that inference compute is roughly ⅓ of its total compute cost; OpenAI’s inference demand includes APIs, which means, if, say, Deepseek is using OpenAI inference to train its own model, that counts as inference for OpenAI; hence, real inference demand of OpenAI is not disclosed).,

China’s ChatGPT, Doubao, developed by Bytedance, just disclosed its inference demand trendline last week. And look at that wild growth trend below.

Translation: the purple curve tracks Doubao’s increasing inference token consumption from May 2024 to January 2025:

• Beginning at 120 billion inference tokens per day (May 15, 2024) with the launch of their foundation large language model

• Expanding to 500 billion inference tokens per day (July 25, 2024) as they deployed their model with photo-generating capability

• Growing to 1.3 trillion inference tokens per day (September 24, 2024) with the introduction of specialized models for video content, music generation, and voice synthesis

• Reaching 4 trillion inference tokens per day (December 18, 2024) as they rolled out more advanced capabilities, including visual understanding, 3D modeling, and their large model version 1.5

The chart here demonstrates ByteDance Doubaos’ exponential growth in AI inference computational scale, measured by the number of tokens processed for actual model usage rather than training. The progression shows wider adoption and practical implementation of their AI products, as each new model release contributes to higher inference demands from actual usage.

So, who’s to say OpenAI’s inference chart does not look like Doubao’s? Who’s to say that inference demand won't increase by 1 billion times in the long run, as Jensen Huang had predicted? The thing is, OpenAI has been so secretive about its pre-training and inference compute figures that we can only reference this as it is one of the first somewhat comparable LLM firms openly disclosing its token usage.

So it brings us back to what we’ve tried to understand before: does the pre-training scaling law still hold, and how will it impact GPU demand? (check this piece out on scaling law; I actually think I did a good job here)

For Nvidia’s Jensen Huang, he’s been saying that the world will need his GPUs whether it’s pre-training or inference demand. However, inference “is about to go up by a billion times,” Huang said during an interview with venture capitalist Brad Gerstner on the B2G podcast. In October last year, around 40% of Nvidia’s business was for A.I. training, and another 40% was for A.I. inference. (The remaining 20% of revenue comes from its traditional video, gaming, and automobile products.)

He said we need to shift how we look at it. Training models is like “going to school,” but the ultimate goal is to train models to perform real-world tasks like “inference.” This is like us getting a ton of degrees, and they’re useless unless you can perform in a job (in most jobs). And delivering results on the job is probably the most important. Thus, he believes that inference demand will become the real demand as we move ahead in this AI journey.

Subscribe now

What is the training vs. inference demand argument (o1’s chain of thoughts & training’s scaling law)? See here.

What is the argument that lower investment into training can still = a good model, so investment into training will likely go down (DeepSeek primer)? See here.

What does 4 trillion inference tokens per day mean?

Looking at this top-down. First the rumors are, OpenAI used ~11-13 trillion tokens to train GPT4. We can see that Doubao’s current DAILY token use is already one-third of that, and it would only take ~3 days for Doubao’s inference demand to consume as many raw tokens as it would take to train a GPT4! Obviously, this comparison is not apples to apples.

Model training would require tokens to be reused multi-times vs. inference is just one time without repetition. Also, for the same amount of tokens, training will consume more compute due to high parameter count whereas inference for Doubao is handled by smaller models with lower parameter count. All this is to say that training is still far more compute intensive than inference (see chart below), even if headline token numbers look similar. But you finally get my point on how big this 4 trillion is and why it is relevant!

Now, let’s look at this bottom-up to conceptualize what 4 trillion tokens can do. Today, if you ask Perplexity a normal complex question, on average, it will consume ~500 input/output tokens (based on an assumption of 500 tokens = 375 words, just enough for planning a travel itnerary or 3-4 back-and-forths on a topic). Doubao has roughly 9 million DAUs as of Dec 2024 (daily active users). If we divide the 4 trillion daily token consumption by 9 million, that would mean each DAU is consuming 4.4 million tokens per day, which would mean each DAU is asking Doubao nearly 9,000 questions (4.4 million divided by 500) every day!

But of course, that’s not really logical, actually quite absurd of a number. No one is asking Doubao that many questions, at least not yet. So there are better explanations of why daily token consumption is so high.

1. First, inference time compute means sometimes Doubao will break down one question into multiple questions when answering, so the token consumed can go up one order of magnitude larger than just answering one question. With the hugely reduced cost of inference shown by the DeepSeek R1 model that has been widely talked about recently, we can be certain again that inference can be scalable at an affordable price (relatively).

2. Second, Doubao disclosed that it currently powers industry-specific applications in more than 30 industries. Hence, we can assume a large amount of daily consumption is used by enterprises inference demand, which makes more sense for the scale.

3. Third, token consumption can increase greatly when the model is multimodal. If you ask AI to generate one photo, on average, it will consume 500 tokens. This is similar to the token intensity of a text question and answer. However, if you ask AI to generate an one-minute video, depending on resolution, a one second video consists of 24-60 pictures, so you can do the math.

• If we assume 24 pictures per second, then one minute of video is 60 x 24 = 1,440 pictures. Tokens consumed for generation one minute of video is 1,440 x 500 = 720,000 tokens.

• Inference demand increases by three orders of magnitude when the user starts generating videos (720,000) vs. pictures/texts (500-1,000).

• With video generation, 4.4 million tokens per day just means merely 6 minutes of video generated using AI.

btw if you think there is something severely off about my assumptions/ calculations, please do let me know, legit grateful.

And this is just video usage; now, imagine the token density when AI starts doing 3D. This aligns with Doubao’s disclosure—it launched a photo generator in July, video in September, and 3D in December. Token density is exploding as Doubao becomes multimodal.

Still, 4.4 million tokens per day per DAU is a lot of tokens, meaning Doubao users are using the platform extensively. More importantly, the fact that Doubao ‘s total daily token count grew so fast really shows the potential for explosive inference growth. Let’s do some quick math on theortically how much this can grow.

Inference token = number of DAU x token density

Douyin, the Chinese version of TikTok, has a DAU of 750 million. Thus, Doubao’s DAU could theoretically grow >50x (from currently ~9 million).

Today, token density is 4.4 million tokens per day per DAU (which we calculated above as just 6 minutes of video generated per day). Users can certainly generate far more AI content than just 6 minutes of video per day, and token density can also grow one order of magnitude.

This means Doubao’s total inference token can still grow multiple hundreds of times off the current level. So let’s wait and see what Doubao disclosed in January - their token consumption is growing at a 50% month-over-month rate to roughly 6 trillion tokens per day already.

Pivotal Moment for Microsoft & OpenAI’s Relationship - Stargate

OpenAI has partnered with Japanese investment firm SoftBank and an Emirati sovereign wealth fund to build $500 billion of artificial intelligence infrastructure in the United States.

Just last week, President Donald Trump announced the new company, The Stargate Project. He billed it as "the largest AI infrastructure project by far in history" and said it would help keep "the future of technology" in the U.S.

How does this inference ah-ha moment all link back to the most recent Stargate deal?

Gavin Baker, Managing Partner and Chief Investment Officer of Atreides Management, LP. Long-term tech and semi-investor took to X to share his hot take on the Stargate deal. See tweet [here].

1. He thinks that Microsoft definitely had its first dibs on the Stargate deal but turned it down strategically.

2. Satya is essentially the only “adult” here and rationally thinking that there is no point in wasting $500 billion to train the model.s

Baker wrote that Microsoft is “opting out of providing the pre-training compute capex for their JV past a certain scale and/or time. Instead, they will increasingly focus on providing inference compute.”

3. Microsoft will keep its $80 billion investment but that will be for multiple usage (and not just for pre-training)

Quoting Microsoft Satya Nadella’s interview, “I’m not particularly in the details of what *they* are investing. All I know is that I am good for my $80 billion. I will spend $80 billion building out Azure.”

Baker writes, “If this is the correct interpretation of his comments, Satya and Amy are likely to clarify that they are not investing in Stargate but are allowing some of Azure’s technology to be used in it hence they are a technology partner not an equity partner.”

Baker added that because “Microsoft is one of the more financially disciplined players in this game, so this is an interesting signal.” highlighting that the ROI on AI has shifted to focus on inference.

And given the context that recent models such as o3 and R1 models (see DeepSeek primer here) have “showed that the biggest future improvements in model capabilities are likely to come from RL during post-training, synthetically generating reasoning traces and test-time compute. These are all essentially inference (ish in the case of RL).”

Hence, Microsoft's move clearly signals that the hyperscalers are shifting from pre-training compute to inference compute. And this might also be the start of the decoupling or at least distancing between Microsoft and OpenAI.

and I guess Elon Musk agrees, as he gave Gavin a bullseye 🎯 See tweet [here]. Again, proving INFERENCE IS THE FUTURE ROI.

And this is the perfect segway to DeepSeek's R1 and what it means for the LLM business model. People are so worried about compute-demand going down as DeepSeek is so much more efficient BUT what if total demand for inference sky rockets and goes up by 1 billion times? Stay tuned.

Some relevant pieces:

1. AI Arms Race Far From Over: Chips is Only Half the Game, Infrastructure is the Other

2. Why Data Centers Can’t Go Full Renewable—Yet

3. A Whole (New) Nuclear World

4. Johor, Malaysia: AI Data Centers Drive Investment But Infrastructure Challenges Persists

5. Bitcoin Mining for AI Data Centers?

Today, we will go back to where we started with this newsletter. We take a closer look at data centers and energy consumption/ shortage and how Trump 2.0 will shake things up (for better or worse).

Subscribe now

In the evolving economic landscape, infrastructure has emerged as a key asset class, garnering increased attention from investors seeking stable returns amid rising inflation. This surge in interest has become more prevalent over the last decade, but some projects/ deals have been met with regulatory challenges over the years. As we see the return of Donald Trump to the U.S. presidency, we look at how his energy policies are poised to revive traditional energy sources and bolster fossil fuel production. As part of this shift, significant foreign investments are flowing into the U.S., particularly along the AI value chain, with some in data centers and technology sectors, exemplified by a landmark $20 billion commitment from Emirati billionaire Hussain Sajwani.

This article is a bit of a mish-mash of topics, but I try to tie it together by exploring how these interconnected trends are reshaping the future of the U.S. economy. It examines the implications of the trends in infrastructure investment, the energy mix, and foreign investment strategies in a rapidly changing political-economic relationship.

The Rise of the $1 Trillion Asset Class: Infrastructure Investment

Infrastructure has always been something private equity firms and public market investors looked at. Traditionally, infrastructure was grouped mainly into 1) utilities, 2) transportation, and 3) energy. Think about electricity grids, roads, bridges, and small nuclear reactor plants. However, as society has evolved and AI development has boosted data center demands, infrastructure investment for many investors has expanded into looking at fixed assets that would affect digital infrastructure.

Infrastructure has expanded beyond transportation or utilities. It now includes services critical to society's functioning, such as social welfare—healthcare and education—and digital facilities—optics, cloud computing, and data storage—making it a $1 trillion market. With the boom of capital injection across the value chain of artificial intelligence, this asset class has recently attracted billions in capital from major investment institutions such as BlackRock, Softbank, KKR and hyperscalers like Microsoft, Amazon, and Meta.

New-ish Asset Class: AI Infrastructure

With the boom of AI data centers, we see that “infrastructure” as an asset class evolving and becoming increasingly “sexy. ” I spoke to two investor friends who work at PE and mutual funds (one European and one Canadian) and mainly invest in private real estate deals. It was interesting to hear that their mandates have recently expanded to include data centers, which have become the most significant proportion of where their capital is being deployed currently. On the other hand, investors who traditionally covered the energy transition sector are also looking at how increasing AI demand will impact various energy sources. Hardware investors who traditionally looked at semiconductors sometimes also had a remit over connectivity technologies, and they are all being affected, too. As many traditional sectors are being impacted along the AI value chain, people are realizing that organically, a new asset class is being created - “AI infrastructure.”

In a report, the consulting firm McKinsey predicts that the U.S. data center demand will grow by at least ~4x to ~80-100 GW by 2030. And according to McKinsey's estimates, global data center demand will increase to at least 170 GW by 2030 at a 19% CABR. The speed and scale of investment in data centers have been staggering. And the cost is high. Building data centers will cost $900 billion over the next ten years, based on a report by Latitude Media, a digital media company focused on the energy transition, with others even forecasting that scaling AI and cloud infrastructure would cost $1 trillion by 2030 in the U.S.

Pankaj Sachdeva, a senior partner at Mckinsey, told Latitude Media that investors from all backgrounds are finding a way to get into data centers and AI infrastructure-related deals. “Whether you are a real estate investor, an infrastructure investor, a credit fund, or a private equity investor, you’re looking at data centers as an asset class and asking whether you should be investing in it,” Sachdeva told Latitude Media. “Even some of the investors that were not that interested [two years ago] are now starting to look at it from an opportunity standpoint.”

Energy Investments, Energy Bottlenecks

However, not all energy transition-related deals are for AI data centers. I’ve written about why renewable energy has not become a reliable and stable way to power data centers here and why the U.S. and global hyperscalers are trying to use nuclear energy despite safety concerns.

The key change factor is that the era of flat power demand increase has ended in the U.S. The country now faces an energy shortage, and Trump is ensuring that the issue is addressed (under the criticism of choosing brown energy sources over green). For many of the big tech, the core challenge impeding the deployment of data center capacity lies within power availability, special permits for site selection, long lead time for power equipment, and the shortage of specialized labor (which I’ve written about here.)

So, how are US hyperscalers navigating the challenges mentioned above?

1. Some hyperscalers are looking for alternative locations or systems that can provide the energy they need without the restrictions they face in their original markets. By moving to these secondary markets, they can access reliable power sources and better infrastructure to support their operations. [i.e., Indiana, Oklahoma for Google; Ohio, Iowa for Microsoft, Minnesota, Louisiana for Meta]

2. Hyperscalers are looking for new ways to obtain the power systems needed. Some are buying their own energy sources to avoid power shortages and using temporary power solutions until they can get regular electricity service. Significant money is also being invested to help them adopt these new energy sources and improve energy storage options. [i.e., switching to gas-powered generators, wind power, and so on]

3. Lastly, they're locking in renewable power contracts. [i.e., Microsoft’s $10 billion framework with Brookfield Renewable for 100% renewable mix]

Therefore, with Trump’s greenlight, McKinsey predicts that data center power demands will substantially increase (and supply will be met) in the next five years in the U.S., adding 450 Twh of new electricity demand at a CAGR of ~22%. This is not just led by AI but primarily by the usage of AI, data centers (some for AI), and manufacturing returning onshore (MAGA). Now, the context here is that power demand has not really increased in the US in the aggregate since 2007, and now, data centers make up about ~30% of the overall net new demand.

A bit of a tangent rant, this is very different from China’s case, where urbanization has led the country’s demand for electricity to grow steadily high single digital % for years, and due to the demand, the supply side has adjusted to keep up. Now, the surge in data center need for power is something the country can actually meet in supply easily, especially given its decade-long subsidy and policy pushing forward renewable options. For details on China’s Eastern Data, Western Computing policy, and energy policies that support AI infrastructure, see here. And to better understand the cloud infrastructure in the U.S. and China, see here.

Subscribe now

Trump 2.0 is Open to Business

At this point, hyperscalers are willing to pay a 10-30% premium for faster access to power, according to Mckiney’s survey. Now, they’ve not had an easy time as hyperscalers were shut down by regulators when they tried to build new nuclear energy plants last year. And you can read about FERC’s NO to Amazon here.

However, the Trump administration’s pro-business stance could mean an OK-GO to YOLO for American data center operators & energy providers.

Microsoft has vowed an $80 billion investment in data centers for AI by the end of the fiscal year of 2025, with more than half of that injected into the U.S. market. As I wrote about here, Microsoft’s Vice Chair & President of Microsoft, Brad Smith, op-ed where the company practically pleaded with the incoming administration to give big tech the go-ahead to invest and build up AI infrastructure and develop AI. The tech executive strategically positions all these motives as a national interest and a fundamental American thing to do.

President Donald Trump has said that he will consider “declar(ing) a national energy emergency” as he orders to unleash domestic energy production and undo many of Biden’s renewable energy support and policies that were once designed to combat climate change. And just yesterday, Trump pulled out of the Paris Climate Accord once again.

Some have said that Joe Biden is the least business-friendly president the U.S. has seen in decades. And in contrast, Trump is the extreme opposite, putting businesses first and foremost. The White House has said that his aim is to cut the red tape and regulations that have hindered free capital movement and consequently impacted Americans’ quality of life and access to parts of the economy.

Trump’s policies will take time to implement and come to life, but the weather wane is pointing to a business first and climate second future.

FDI into the U.S.

Just days ahead of his return to the White House, Trump announced a $20 billion deal to build new U.S. data centers, the caveat it’s led by Emirati billionaire Hussain Sajwan, founder and CEO of DAMAC Properties is a long-term business associate of Trump’s. This is only one of many foreign direct investments into the U.S. since Trump’s election victory.

In an interview with CNBC’s Dan Murphy, Sajwan was asked if he would eventually invest more in the U.S. to which he replied: “Well, I mean, the sky is the limit, we can invest more, as much as the market takes, then the limitation is not on.”

Sajwani seems to be enjoying his moment. The DAMAC Properties chairman went on Bloomberg’s podcast to talk about why they’re investing in the U.S. as well that day. He said that the company is seeing double-digit growth in data center growth and has invested in over 10 countries globally, the U.S. being one of them, the others in Asia and Europe (Malaysia, Indonesia, Greece, etc).

Adding that he’s been waiting for this moment for four years, that is, to invest in the U.S. when the foreign investment environment becomes more “friendly.”

At the end of his interview, he was asked if he had met Elon Musk and Sam Altman. He said he mumbled his words a little, saying something like, “I don’t (know) a Sam Altman,” but he confirmed that he met with Musk at Mar-a-Lago over New Year’s (I had a little chuckle here).

As the new administration has become so-called “friendlier” to FDI, foreign investment interest in the U.S. has increased. In December last year, the CEO of Japanese tech investment firm Softbank, Masayoshi Son, announced plans to invest $100 billion in the U.S. and create 100,000 jobs over the course of Trump’s four-year term.

[Jan 22 add: On Jan 21, the FT reported that OpenAI and SoftBank have jointly announced that they planned to launch a significant new US artificial intelligence infrastructure project, in a move Donald Trump lauded as a “declaration of confidence in America.”]

Will this be a trend in 2025? We’ll see. For all I know, given this crazy week of TikTok/ REDNOTE flip-flopping and chaos, that will likely not be CN $. And what about potential societal and environmental impacts? We’ll need to wait and see but it probably won’t be ideal.

Please check out his newsletter, which is full of wisdom. He’s someone who really takes the time to consider the societal implications of technical innovation and offers something unique.

Value Added

Value Added is a newsletter about the political economy of innovation.

By JS Tan

What is Cloud? And Why is it Important?

The cloud isn’t just a silent partner in the AI boom—it’s the engine driving it. Without the robust compute infrastructure made possible by the cloud giants, today’s most advanced AI models wouldn’t exist, let alone reach millions of users worldwide.

1. Cloud computing is dominated by the U.S. and China, making American and Chinese firms the only viable contenders in AI development. Yet cloud computing in the U.S. and China has long preceded the AI boom and has evolved in different ways. To understand the present and future of AI, we first need to unravel the story of cloud computing—how it evolved, how it differs across these two tech markets, and why it’s now the bedrock of the AI era.

2. Cloud computing is essential for AI because it provides the digital infrastructure to train advanced AI models. When researchers and engineers create AI models, they need access to vast amounts of computational power. This is where cloud computing comes in. It acts as the interface between high-performance GPU chips and the complex code that trains the models. For example, OpenAI trains its models on Microsoft’s cloud (Azure).

3. Cloud computing also plays a crucial role in deploying AI models for practical use by individuals and businesses. Once a model is trained, it needs to be hosted somewhere so users can access it. In some cases, models can live directly on “edge devices” such as cell phones (but are constrained by the computational power of the device). More commonly, AI models are deployed on the cloud and made available over the internet, enabling users and other businesses to interact with them, whether it’s through chatbots, image generation tools, or recommendation systems.

The public cloud existed long before generative AI became mainstream—as early as 2006, in the case of Amazon. Cloud firms are uniquely positioned to train and operationalize AI models because the cloud business model requires them to aggregate computational resources to rent out. This aggregation makes cloud providers uniquely positioned to dominate in the AI era.

But not all clouds are created equal. This post examines the cloud computing sector in the two countries from a comparative perspective, focusing on differences in corporate governance, industrial policy/state-business relations, and infrastructure.

Subscribe now

Leading Cloud Service Providers in the U.S. and China

The leading cloud providers in the U.S. are Amazon, Microsoft, and Google, in that order. Amazon started to develop its cloud services in the early 2000s, launching its first cloud storage (s3) in 2006. According to Wang Gufen, Huawei formed its cloud division as early as 2005, making it one of the earliest entrants in the space in China. Despite starting at similar times, Amazon—and U.S. firms—have soared ahead of its Chinese counterparts. Today, the U.S. dominates the cloud, with China trailing as a distant second and the U.K. following as an equally distant third.

The U.S. cloud ecosystem is dominated by software incumbents. By contrast, China’s cloud providers have more diverse competencies. The telecom SOEs are hardware-first companies with strengths in telecommunications infrastructure. Huawei is traditionally strong in hardware but has also become a formidable software giant in recent years. The company has created a comprehensive ecosystem and has long invested in cloud, hardware, and various touchpoints with enterprises and consumers (far before AI became hot).

In fact, among the Chinese players, the interesting breakdown is in their clientele. Alibaba, the once big-tech darling, mostly did things the “capitalist” way and provided services to those who could pay, in this case, the large corporations. While most of Huawei Cloud’s largest customers are state-owned enterprises in China (SOEs), many people say it's mostly a result of the buyer’s choice, given Huawei’s unique hardware-software integration and advanced safety features that the company often highlights.

Looking at the bigger picture, the U.S. cloud service providers have captured the majority of the Western market, and companies like Huawei and Alibaba have focused on growing their businesses domestically and in emerging markets. This has made them competitive providers across APAC, EMEA, and LATAM (essentially ex-Western markets).

Still, Amazon, Microsoft, and Google remain the preferred cloud providers in most regions worldwide (see the figure below) and have performed well in the Chinese market as well.

Ownership structure

In the U.S., the cloud is dominated by publicly traded giants like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud, which operate under traditional corporate governance models typical of publicly traded companies to optimize capital and shareholder return. In contrast, China's cloud computing landscape reflects a more diverse set of ownership models involving publicly traded firms, private firms, and state-owned enterprises (SOEs). The leading cloud service providers are still private firms, with Alibaba Cloud, Tencent Cloud, and Huawei Cloud being the biggest players. Now, the cloud businesses of Alibaba and Tencent are set up as business units under parent companies and are publicly listed in the U.S./ Hong Kong, while Huawei remains privately owned. Following the pack are China Mobile, China Telecom, and China Unicom, which are SOEs and China’s largest traditional telecommunications providers.

Since cloud computing is an extremely capital-intensive, low-margin endeavor and different ownership structures have varying abilities to raise and deploy capital. In the U.S., for instance, publicly traded cloud firms are bound to shareholder expectations. Still, their high valuations enhance their financial credibility and flexibility, allowing them to access capital more easily. Moreover, Amazon, Microsoft, and Google are all cash-rich firms and have funneled savings from their dominance in other sectors (e.g., online retail for Amazon) into their cloud business.

Privately held companies like Huawei allow it to focus on strategic or national goals without the immediate need to satisfy external investors. Similarly, SOEs are shielded from shareholder scrutiny. In China’s case, SOEs often also benefit from government backing and policy support, facilitating large-scale investment in critical infrastructure. And the relationships between state support and companies often go beyond financial funding and resource allocation support; they extend into providing the right connections, “guanxi,” and even client-sourcing support. In these ways, China cloud firms, which are less cash-rich than the U.S. tech giants, benefit from these ownership models that allow for long-term, worry-free strategic investment.

Share

Whether privately owned, publicly listed, or state-controlled, these different ownership structures also reflect the nature of state-business relations in each country-context. In the U.S., the state does not play an active role in directing the direction of cloud innovation nor the coordinating players within the cloud ecosystem. The extent of the state intervention is acting as a major client. Since 2016, tens of billions have been funneled into cloud provider’s coffers from various U.S. agencies. While being a client can hardly be characterized as state intervention, these large state contracts guarantee the demand for cloud service, thereby reducing the risk of investment by cloud providers.

By contrast, state influence in China’s cloud sector extends to a broader range of activities. The state provides financing (both as a customer and via subsidies) as mentioned, but it also plays a vital role in policy encouragement. It has actively planned the expansion of China’s cloud computing industry and directed the activities of private sector players. As early as 2010, the state launched its first initiative to encourage the development of cloud capabilities and to build out data centers for cloud computing. In the 13th Five-Year Plan for Economic and Social Development, issued in 2016, the state listed cloud computing as a key field encouraging more investment in the sector. In the 14th Five-Year Plan (2021), the party continued to emphasize cloud computing as a technological priority. Following the central government’s lead, local authorities have followed suit, launching new initiatives to attract investment in cloud and bolster its development. These policies showcase the state’s long-term commitment to supporting cloud computing.

Eastern Data Western Computing

This brings us to a key policy that was introduced in 2020, the “Eastern Data Western Computing” initiative. It is a national effort backed by five federal-level agencies, including the National Development and Reform Commission (the key agency that sets the agenda for the nation’s economic planning), which Sinocities lay out the details of here.

The ambitious planning strategy (which Grace has written about here) is perhaps the state’s most ambitious intervention into the cloud industry thus far. The project aims to reshape China’s digital infrastructure landscape, addressing the growing demand for computing power in the economically thriving eastern regions while leveraging the abundant renewable energy resources and uninhabited land found in the western regions.

Subscribe now

The plan's specifics involve constructing eight computing hubs and ten data center clusters to facilitate a balanced distribution of computing power across China. Its goal is to create an integrated national system of data centers that enhances the planning and intelligent scheduling of computing resources. The ultimate goal is to transfer data generated in eastern China—home to its major industries and dense population—to data centers in the resource-rich western provinces, and it came about because most of China’s data centers are currently located in the East, along the Eastern coastlines, where urbanization is much more developed. However, those data centers are becoming increasingly costly and difficult to maintain. Lack of access to renewable energy also means many of these data centers rely on fossil fuels, contributing further to pollution and climate change.

To break it down, the basic idea behind the concept is to leverage the resource-rich western province for computationally expensive tasks that do not require proximity to users. Consider rendering a feature-length animation. Rendering is a highly demanding process in terms of computational power, but it only needs to be done once. Once the animation is rendered, it can be distributed without requiring further heavy computation. In contrast, streaming that animation to viewers is an entirely different type of task. Streaming doesn’t require significant computational power—instead, it demands low latency and high data throughput to ensure a smooth viewing experience. Therefore, streaming services need to be handled by data centers that are geographically close to end users.

AI has a similar logic. Training an AI model requires massive computational resources that don’t need to be performed near the end users. Model training can take days or even weeks, but once completed, the model is ready for deployment. The idea behind the Eastern Data Western Computing initiative is to allow these training tasks to be offloaded to data centers in resource-rich, remote regions where land, energy, and cooling are more readily available and cost-effective. By contrast, inferencing an AI model, like streaming an animation, requires low latency to provide a seamless user experience. So, they must be handled by data centers close to users (mainly on the eastern coast) to minimize lag.

While it's too early to assess the plan's impact, it is clear that national-scale coordination across the various players is key for its success. And, to a large extent, what makes this coordination possible is the fact that China’s cloud ecosystem isn’t comprised solely of publicly traded private firms (as well as firms that have core competencies in hardware and not just software). Unsurprisingly, the two largest investors in the plan are SOEs—China Mobile and China Telecom—based on the research of Sinocities. Privately held national champion Huawei is a close third. Alibaba and Tencent—publicly traded and disciplined by short-term profits—are considerably smaller investors in the plan.

Source: Sinocities

China’s Energy Advantage

As mentioned, China’s ability to coordinate across public and private sectors and push ahead with one agenda has been a major catalyst to its economic success that we’ve seen over the last four decades.

In 2014, Guizhou was established as the National Big Data (Guizhou) Comprehensive Pilot Zone to create a national-level cluster for the development of the big data industry. By 2022, the province was already providing services for the likes of Huawei, JD.com, Alibaba, Tencent, and Foxconn, as well as leading international companies such as Apple, Microsoft, Hewlett-Packard, and NIIT. Its speed is unparalleled, and in just under a decade, there have beenIts speed is unparalleled, and in just under a decade, there have been over 23 key data centers in operation or under construction in the province, including eight ultra-large ones, making it one of the world's most concentrated regions of ultra-large data centers.

This coordination, however, extends beyond the computational stack and into energy infrastructure. The last ten years of investment by the state in renewable energy are paying off, as it plays a crucial role in powering China's energy transition at large. It is now also powering a critical sector—the burgeoning data center landscape. While coal remains a significant part of China's energy mix—accounting for about 60% of electricity generation—national orders will move to greener sources in 2025.

China’s infrastructure advantage also leans into its competitive electricity prices,which has positioned it well to get a head start on this front. To note, western regions such as Shanxi and Guizhou offer some of the lowest industrial power rates globally, providing a significant advantage over countries like the United States, which is facing major power shortage concerns and rising energy prices.

By coordinating efforts across sectors—energy production and data center development—China is capitalizing on its renewable resources to support increasing demands for computing power and artificial intelligence. This approach is a strategic move to promote sustainability and address regional disparities in computing capacity, as it has also been proven to work in the initial stages of the renewable energy infrastructure build-up.

Still, it's important to take a step back and see the bigger picture. While Chinese cloud firms have permissive ownership structures that allow for closer state-business relations, data center deployment in China still lags behind that in the U.S. China’s top-down strategy has been proven to be efficient when ramping up an industry, but whether it is enough to compete head to head with the U.S. remains to be seen.

If you found this piece insightful/ interesting/ entertaining, please support our work by subscribing to AI Proem & Value Added. Feel free to get in touch via DM or leave a comment below. Thank you!

Below, I’ve highlighted a few key points from the EIU piece and elaborated more. The article represents my views, and mine only, on China’s AI infrastructure build-out.

TL;DR

• China’s data center and cloud market will likely continue growing in 2025- 2026. AI demand will largely drive this growth, which will boost the country’s fixed investment (power plants, data centers, cooling systems, infrastructure).

• Government policies such as the “Eastern Data, Western Computing” (“东数西算”) initiative have effectively optimized regional resources.

• The country’s top-down approach has faced some challenges but overall had a massive advantage in rallying public and private sectors together to get behind AI, much like it did with renewable energy a decade ago.

• There have been signs of the excess buildout of data centers, similar to what happened with cloud infrastructure built in its initial stages.

Subscribe now

Tech-related investment and tangentially related industries vastly scaled back during the 2022-2023 regulatory storm. However, over the last year, due to the growth demand for generative AI, China has been mirroring global trends in pushing ahead with computing infrastructure and increasing capital expenditure, according to the latest EIU report.

Data from the government-backed think tank CCID Consulting indicates that more than 250 internet data centers have been built or are under construction as of June 2024. However, there is a worry that the state’s overly enthusiastic involvement could lead to less efficient use of computing power and eventual oversupply. Concerns were already raised mid-last year that China is experiencing an oversupply of computing power due to the AI boom (SCMP reported).

Some experts warned of a potential glut, as many data centers currently sit idle. However, this may not be a significant concern; the demand for computing power is expected to surge as AI technologies advance. However, during this growth stage, it is probably preferable to have excess capacity rather than face a shortage.

Source: EIU’s report China’s cloud, data centre and AI investment wave

After all, China’s overall spending on IT infrastructure seems to still pale relative to the U.S. According to the EIU’s estimates, China spends about 2% of its GDP on software, hardware, and IT services, compared to 5% in the U.S.

Source: EIU’s report China’s cloud, data centre and AI investment wave

That oversupply narrative was temporarily silenced when a September 2024 Morgan Stanley report suggested that hyperscalers in China have absorbed previous oversupply more quickly than anticipated, primarily driven by companies like Alibaba, Tencent, and ByteDance. Data center demand has surged recently, indicating a robust appetite for computational resources in the face of evolving technological needs.

Share

Eastern Data, Western Computing: A Strategic National Initiative

China's ambitious “Eastern Data, Western Computing" strategy reshapes its digital infrastructure landscape. It addresses the growing demand for computing power in economically thriving eastern regions while leveraging the abundant renewable energy resources in western regions. The government is encouraging private and private sectors to follow the guidelines of this initiative as it not only aims to enhance computing capacity but also aligns with China's commitment to long-term sustainable development goals.

This strategic shift aims to alleviate the pressure on eastern data centers while utilizing lower average temperatures for cooling purposes, lower electricity tariffs and costs, and more readily available land in the west. The grand initiative, introduced in 2020, is a national effort backed by five federal-level agencies, including the National Development and Reform Commission.

Source: Premia Partners

The policy's rollout was applauded domestically, as it “helped to foster the data center and cloud service markets for over a decade,” wrote the EIU. This broadly aligns with President Xi Jinping’s goals to promote technological breakthroughs, enhance the resilience of industrial supply chains, and enable enterprises to become more productive and move up the value chain. Data center and cloud services, especially those directly related to AI development, naturally fell into that favorable basket.

The plan involves constructing eight computing hubs and ten data center clusters to facilitate a balanced distribution of computing power across China. Its goal is to create an integrated national system of data centers that enhances the planning and intelligent scheduling of computing resources. This effort aligns with China’s National Economic and Social Development strategy and Vision 2035, aiming to optimize resource allocation while supporting the country's ambitious AI objectives.

What led to this is that the majority of China’s data centers are currently located in the east (near the big cities such as Beijing, Shanghai, Guangzhou, and Shenzhen ‘北上广深’), where they are becoming increasingly costly and difficult to maintain. Lack of access to renewable energy also means many of these data centers rely on fossil fuels, contributing further to pollution and climate issues.

Some concerns relocating data centers to the West may lead to latency issues have been raised, but that’s a complete different discussion.

Renewable Energy Abundance in Western China

This kind of policy is not new in China. Over the years, the Chinese government has nurtured (jumpstarted) many industries with such policies, including the renewable energy sector. As mentioned previously, the state’s “invisible hand” has always been involved in China’s strategic development.

The EIU wrote that the central government has long played a significant role in guiding the market and local/ regional authorities in establishing a “national grid” for computing power to allocate resources better.

China's competitive electricity prices make it an attractive destination for energy-intensive industries. Western regions such as Shanxi and Guizhou offer some of the lowest industrial power rates globally, providing a significant advantage for data center build-out.

Renewable energy, albeit a way for a green transition, has also been preparing for this moment. It plays a crucial role in powering China's burgeoning data center landscape. While coal remains a significant part of China's energy mix, accounting for about 60% of electricity generation, ambitious targets aim to shift towards greener sources by 2025.

New data centers are expected to source over 80% of their electricity from renewables, a goal supported by current trends in which wind and solar capacities surpass traditional coal-fired plants. Current industry forecasts suggest that wind and solar capacity will overtake coal by 2026.

By strategically coordinating efforts across sectors—energy production and data center development—China can capitalize on its renewable resources to support increasing demands for computing power and artificial intelligence.

Subscribe now

A few significant provinces that are already key to the initiative:

• Inner Mongolia: This province is a leader in wind and solar energy production, boasting some of the largest wind farms in the country. Inner Mongolia's energy infrastructure supports low-cost electricity generation, which is crucial for powering data centers.

• Guangdong Province: The southern province focuses on solar installations, capitalizing on its sunny climate to contribute to the overall renewable energy mix. It is also being encouraged to move data centers undersea to cut costs.

• Sichuan Province: Renowned for its extensive hydroelectric capabilities, Sichuan generates approximately 85% of its electricity from hydropower. The province's surplus energy is often exported to more developed coastal regions, meeting the high demand for electricity in urban centers. It is also home to many data centers.

Future Demand for Data Centers Will Continue to Rise in China

Demand for data centers is anticipated to grow significantly as AI technologies advance. As highlighted by Premia Partners, AI could demand approximately 134 TWh of electricity annually by 2027—an amount comparable to the annual consumption of entire countries like Sweden or the Netherlands.

William Huang, Founder, Chairman, and CEO of GDS Holdings Ltd, said during the company’s most recent earnings call that the new demand for data centers, approximately ~ 70%, “was driven by the AI-type requirement” in China, for both training and also inference needs. The other 30% is driven by Internet companies and the traditional cloud business.

While short-term challenges related to oversupply exist, China's strategic investments in AI infrastructure and renewable energy position it favorably for future growth. Balancing supply with demand while ensuring sustainable practices will be essential. Time will tell whether all these factors can progress simultaneously.

FYI, I’m publishing a deep dive into the differences between China and the U.S. in cloud infrastructure build-out next week. Stay tuned.

Side rant: I am so impressed by Chinese acronyms and how beautifully four characters crammed together can sound so harmonious and make so much sense and compact so much meaning in it, “东数西算”- genius modern “chengyu.”