Hi hi hi, I’ve been sitting on this piece for two weeks, but every few days there’s a new seismic shift in the industry, so I’ve just been digesting and readjusting my thinking on this. Anyhow, it’s 5 days before Christmas, and I decided it was time to put a mental time stamp on this and let it blast out to the world.
So to start, two emails landed in my inbox last week and put a smile on my face. One was from an editor I met nearly a decade ago in New York, and the other was from the director of a program I did over a decade ago. Both wrote for the same reason: they’d been following AI Proem!
Now, many have asked you what Proem means. Well, this is my opportunity to explain it to you. Proem means preface, and I named this newsletter AI Proem because the industry is evolving every day, we’re learning every day, and everything we know today is a preface to what will happen tomorrow. So, here’s my wrap of 2025 — and my proem to 2026.
What 2025 Actually Changed About AI
Vertical AI, agent interfaces, and an energy-constrained China–U.S. stack are reshaping how AI is built and commercialized.
This year, we moved from comparing models to interrogating the systems surrounding them: power grids, super-apps, cloud stacks, industrial bases, and the interfaces through which people actually experience and deploy AI. (FYI 2025 Second Half Top Stories List)
The more I looked at those systems, the harder it became to believe that the future of AI will be decided by a single breakthrough model. It will be decided by the convergence of infrastructure, software distribution, and the institutional contexts in which AI is adopted.
Across my analyses on AI Proem and conversations on Differentiated Understanding, I kept coming back to the same pattern: AI in practice is shaped less by abstract capability and more by structural forces and practical demands underneath it. In China, those forces are super-app distribution, industrial policy, manufacturing capacity, and an emerging open-weight ecosystem. In the U.S., there are cloud platforms, enterprise software adoption, and the bottlenecked physical infrastructure needed to expand data centers. Once I put those layers side by side, the “AI race” narrative stopped feeling like the most useful lens. What I see instead are two different stacks, operating under surprisingly similar constraints.
Before I go further, I want to do a small self-audit — not in a victory-lap way, but because this industry moves so fast it’s easy to forget what you believed a year ago. When I look back at my own writing, a few threads I flagged early ended up becoming more central than I expected: first, that China would lean hard into open-weight and open-source as a distribution and ecosystem strategy after DeepSeek; second, that the “interface” question would matter more than model rankings, and that we’d move beyond mobile + text into multimodality, long-form memory, and voice control as the default interaction mode. Those weren’t the loudest talking points at the time, but they were where incentives seemed to be pointing.
So, this is a piece where I analyze the key themes I believe we saw in 2025 and will see in 2026. Not dismissing any safety concerns risks we may see that come with AI.
Energy Became the Hard Ceiling
Let’s start with infrastructure, because it’s the least glamorous layer and the most determinative.
Capex by hyperscalers is the headline, but the more important question is what capex means in an AI world. It’s data centers, chips, talent, and energy — except those don’t scale symmetrically. GPUs can be bought; power and time-to-power cannot.
Back in late 2024, I wrote that the “AI arms race” framing missed something basic: chips are only half the game; the other half is land, power, permits, and time-to-grid-connection.
2025 was the year that stopped sounding like a contrarian line and started sounding like the core constraint. In the U.S., bottlenecks manifested as multi-year delays on grid interconnections and substation buildouts. In China, they manifested as accelerated activity: new compute zones, renewable integration, and the ongoing “East Data, West Compute” strategy that has been in motion for over a decade. Power became a strategic asset, not just an operational line item.
DeepSeek’s breakthroughs reinforced another dynamic I’ve been writing about all year: lower model costs don’t decrease energy demand; they multiply use cases. Every improvement at the model or kernel layer kicks off another adoption loop in adjacent domains. That was the heart of my AI Proem piece The Jevons Paradox in AI Infrastructure,” which highlights it’s efficiency accelerates consumption, not the other way around.
My conversations with David Fishman, Chim Lee, and Kyle Chan deepened this perspective: AI is now inseparable from energy planning, industrial organization, and long-term infrastructure regimes. It’s not just an AI story. It’s a power story.
China and the U.S. Diverged Further — and Also Quietly Converged
I’ve never loved the “race” framing, not because geopolitics don’t matter, but because it obscures the fact that China and the U.S. are building structurally different AI economies — and it tends to paint everything with one big geopolitical brush. There’s a joke in China about the AI race that goes, “it’s Chinese people in China vs Chinese people in Silicon Valley.” It’s funny because it points to something real: technology doesn’t stay inside borders the way narratives want it to.
At a high level, the two economies can seem very different, and each has obvious strengths, as Kyle Chan puts it succinctly here. And at the beginning of 2025, the sense of difference came through clearly in what I wrote in Why China’s AI Strategy Differs From That of the U.S., “Divergent Approaches to AI Commercialization, and in interviews with Jing Yang, Wency Chen, George Chen, and others.
The U.S. system channels AI through cloud platforms, enterprise software, and standardized workflows — the pursuit of frontier models, the obsession with AGI, and the integration of AI into systems of record. U.S. companies such as OpenAI, Anthropic, and Gemini remain the clear top choices for enterprise GenAI adoption, and we’re seeing GenAI embedded across a wider set of use cases.
The Chinese system channels AI through consumer super-apps and the real economy — WeChat search and mini-programs, Taobao and DingTalk, Douyin and Doubao, Meitu and Meituan — and diffusion in the real economy feels more pragmatic in many ways. Open-weight and open-source models have also increasingly become the default. In the coming year, diffusion into deeper vertical use cases will be prevalent.
And then, the convergence: both sides are hitting the same constraints — power, land, chip restrictions, and escalating compute demand. Both sides are moving toward agentic interfaces embedded inside existing ecosystems — WeChat on one side, Copilot and GPTs on the other. And both sides are discovering that vertical AI, not general chatbots, is where real commercial traction lies for companies beyond hyperscalers - and we’ll go into more details below.
2025 also felt like a sentiment inflection point for global investors investing in China. The mood shifted from “China is uninvestable” to “curious but selective.” But it wasn’t just capital interest coming back to China. It was a shift in the mainstream “China watching” space as well. Where Dan Wang’s book Breakneck, Kyle Chan’s NYT op-ed, and Kaiser Y Kuo’s essay The Great Reckoning were read not just by the few in DC but across investors, policymakers, business people, and academics. There appears to be broader recognition and interest in China’s industrial policy and innovation, and an awakening to the relationship between planning and execution in China’s economy, which has, as a result, made it a real contender for AI supremacy.
And after a few years of what was called “China shedding,” the new age of LLM companies embraced their Chinese heritage from Moonshot to DeepSeek. While Benchmark’s first China AI deal in Manus, structured via a Singapore parent company, was a microcosm showcasing companies and capital choosing sides - but only in an optical illusion way, ultimately, capital still wants exposure to Chinese AI talent and products, but via new legal and geopolitical feasible pathways.
Meanwhile, more players joined the story across Asia, from Southeast Asia to South Korea. Southeast Asian funds have quietly been recruiting Mandarin-speaking investors and courting Chinese founders to build AI companies in Indonesia and across the region - while remaining realistic about local knowledge and governance constraints. Singapore now plays a bigger role than people had expected in terms of capital, talent, and more.
Inside China Big Tech: Operating Styles as Strategy
One thing I’ve come to appreciate more this year is that strategy isn’t only about products. It’s also about operating style — and for technology companies, that often determines how well they capture and retain targeted users inside their ecosystem.
I thought, as people constantly question whether this is a bubble or a fad, Howard Marks answers it clearly.
“Struggling with whether to apply the “bubble” label can bog you down and interfere with proper judgment; we can accomplish a great deal by merely assessing what’s going on around us and drawing inferences with regard to proper behavior.”
To understand the capex vs. revenue angle, see Kevin Zhang’s article here. But to best understand what’s going on in terms of long-term strategy and where the Chinese AI players are placing their bets. We look at the BAT.
We just looked at ByteDance’s upgraded playbook. For a company of its size, reports have pegged ByteDance’s 2025 capex at RMB 150 billion+ to ~RMB 160 billion (~USD 22 billion). Meanwhile, reported figures put Alibaba’s AI + cloud infrastructure capex at ~RMB 120B over the past four quarters, with Ant Financials investing another ~RMB 30B, together roughly ~ RMB 150 billion (~USD 21 billion). The exact number matters less than what it signals: China’s biggest consumer internet companies are now spending like infrastructure companies, because inference and training at scale have become the new production line. The Great Wall Street talks about it here, too.
ByteDance is currently the only real contender today that can compete with Alibaba at the ecosystem level, and it sits somewhere in the middle culturally. We’ve been writing about Zhang Yiming driving AI in the background. Doubao remains in the top two in MAU, and its token consumption has been extremely high. Volcano Engine leads in the public cloud model-as-a-service (MaaS) market. And then ByteDance’s “App Factory” — Doubao, Maoxiang (“Cat Box”), Jimeng AI (video model that competes with Kling), Trae, CapCut — continues to be the silent winner in entertainment, creative, and prosumer workflows. ByteDance is spending heavily on inference, reportedly using roughly 30 trillion tokens per day internally for its Doubao model. The latest QuestMobile figures show 56.7 million MAU, making it China’s most actively used AI application.
But the biggest game-changer was ByteDance’s launch of its AI OS. On the surface, people ask why voice control is such a big deal — Amazon has had Alexa for a while, and Apple has had Siri. But if they succeed, this could be more disruptive than it looks. As smartphone technology advances with each generation, it creates opportunities for newcomers to challenge the existing operating system, and it’s something incumbents usually don’t get a chance to defend against. This is a moment when ByteDance’s newly launched “Doubao Mobile Assistant” could become the first AI-native OS for phones. I’ll come back to the voice-interface point later.
What’s even more interesting is that what we’re seeing now is a reflection of these companies’ personalities at their core, in this newly reignited “拼” — competition. With Alibaba’s clear top-down style, a command from the top has led to a fury of “AI in everything and AI everywhere” across business units and functions. Tencent’s “山头主义” - factionalism and ByteDance may be somewhere in the middle.
For much of 2025, Tencent didn’t feel like it had a clear high-level strategy. At least not in the way Alibaba’s mandate was visible. Yet Tencent’s advantage is obvious to anyone who has ever used WeChat: distribution. But they seem to be making a pivot. When I met them, executives said the company will double down on its proprietary models and see agentic AI using its own model inside WeChat as the next step. If that happens, it’s not just “Tencent adding AI.” It’s Tencent turning WeChat into an operating environment.
Pony Ma is famously hands-off, and he enables the “山头主义” vibes — turning a blind eye to internal competition as long as business group leaders continue to innovate. To an extreme, people have said that Tencent Games will rent a new building next door to Tencent’s new Shanghai flagship building, which is only 45% occupied. WeChat continues to be based in Guangdong, where the legendary product leader Allen Zhang prefers, while the HQ of the rest of the company's core business sits in Shenzhen.
But this time, there is a change in management efforts. A top-down push, and the courtship of Yao Shunyu, a 27-year-old scientist snatched from OpenAI, to lead the company’s AI lab as its Chief AI Scientist, is one signal. Another was a senior executive I met, saying that the company will be doubling down on its proprietary LLM. Seems like the original bet was that LLMs would become commoditized, but a late but steady push could put Tencent in a position where it wastes less and spends more strategically.
During a company tour, I also noted that WeChat enterprise is doing what Google is doing — but honestly, probably even more seamlessly. In the future, you can receive PDFs and docs via WeChat, ask AI to summarize them, generate logic-tree charts, and then directly save them to your drive. Similarly, the WeChat corporate call function can provide simultaneous translation for tens of languages. Tencent has seamlessly integrated AI into existing touchpoints. Now the missing piece is a strong model focused on agentic tasks that is truly native to its tools.
On the topic of China’s 2C app winners, an unnoticed mover is Ant. It’s been quietly 操作ing - working at it. After the regulatory storm in 2020, what was supposed to be the largest IPO ever never happened. Over the last few years, Ant restructured and revisited some of the functions that drew scrutiny, such as P2P loans. And then, quietly, it captured two major vertical AI tracks in China: a healthcare product called Ant Afu (supposedly 15 million MAU) and a consumer finance AI product called Ant Xiaocai, plus its own LLM Lingguang. For outsiders, Ant is a separate legal entity with a separate management team from Alibaba. But anyone inside the China big tech ecosystem knows they still operate in cohesion and collaboration, and that is an understatement of their intimate relationship.
Which means when you zoom out and look at Alibaba’s posture, which is heavy investment in cloud infrastructure, the Qwen series and Wan series models, and a broad product rollout from Qwen, the 2C app to Quarke to Qoder, with AI functions embedded across existing apps so it is clearer now that Alibaba’s ecosystem is much broader than many would have expected.
When I spoke with people at BAT, the consensus was that, although Alibaba has recently been shifting management from top leaders to mid-level staff, the company fundamentally operates on a top-down, mandate-driven style, almost like the military. When the top leader says “we’re all in on AI,” it really means every single person — product, engineering, e-commerce, logistics, pretty much every moving team goes all in. And as I’ve written before, Alibaba is using the Qwen app as an entry point to commerce activities like maps, delivery, ticketing, and shopping. This is what it looks like when a company treats AI not as a feature, but as a new operating system.
Diffusion Into the Real Economy Became Visible
One of the most striking features of 2025 was how AI left the lab and embedded itself into physical systems, especially in China, where diffusion into the real economy has been rapid and pragmatic. Pieces like Electric Dreams, Robot Reality, Rise of China’s Robotics Industry, and Robots Everywhere showed how EVs, robotics, drones, and industrial software became frontline AI adopters.
EV companies like BYD and Zeekr now behave more like robotics + software firms than traditional automakers. The joke is that EVs are really just larger drones, if you think about it. And robots are small EVs with arms? This may be oversimplifying it, but you get my point - there are a lot of technological similarities and components that borrow from each other. (and again, Kyle Chan explains it perfectly in China's overlapping tech-industrial ecosystems.)
And on the point of these EVs, there was no way that Chinese makers could compete with Japanese and German automakers on the traditional ICEs because they had decades of practice and honed craftsmanship. A traditional Mercedes or Audi still drives smoother than any other country’s sedan. However, their opportunity arose when a completely new technology emerged, and everyone started from zero.
And, citing a piece published nine years ago, it remains as relevant as ever.
“The modern tech industry we’ve created can be thought of as a paradigm shift machine. It’s an industry whose core process is replacing the scarce resources of other industries (and of itself) with new ones, and in doing so, rearranging their business models and ecosystems.” — Alex Danco, while he was still at Social Capital.
In 2025, we also saw humanoids and service robots enter factories and logistics nodes in low-friction ways. Agricultural drones reached mass deployment. It’s also really not about the flashy robot demos doing backflips, which, right now, have proven to have little real-life use. But it’s the vertical use cases that are more promising. Think about street cleaning robots, obviously industrial robots, manufacturing robots, agricultural-bots? tutoring robots? Healthcare robots, surgical use robots, and so on. And it is in these unexpected ‘non-glitzy’ sectors that are adopting vertical AI tools far ahead of expectations.
This is where China’s AI Plus and Internet Plus policies mattered not just as slogans, but as mechanisms that push first digital infrastructure and now AI deeper across the real economy. Tom Nunlist further explains the speed and scale of diffusion here, where he explains how the policy and industrial clusters create adoption pathways. And in the coming year, we’ll see diffusion across geographies, demographics, and sectors.
But constraints remain. In “The Bottlenecks for Embodied AI Development,” I wrote that physical data is thin, teleoperation and simulation are expensive, and we’re probably at least another 5–10 years away from humanoids being useful at scale in households. Enter 2026, a shift in focus on spatial intelligence, and advocated by the godmother of AI Fei-Fei Li.
Open-Source Embrace
"Show me the incentive, and I'll show you the outcome” — Charlie Munger
This quote from Munger encapsulates it. Anyway, I’m going to keep this section short and get to the point.
Open-source isn’t new in science and technology. What changed in 2025 is that open-weight/open-source became a default posture for a critical mass of Chinese model builders and platform companies, and that posture is best explained by incentives rather than ideology.
First, open-weight enables Chinese companies to improve on each other’s work quickly — a practical, iterative “build on top of the last release” dynamic rather than treating models as isolated crown jewels.
Second, there’s a collaborative instinct here that I’ve described before as an “open-weight remix” culture — not just to ship faster, but to create a shared technical baseline that the ecosystem can compound on.
Third, there is the “zeal” aspect I wrote about via Kevin Xu’s framing of 开源情怀 — the desire to be cited, referenced, used globally, not just monetized domestically.
Fourth, open-source is also a reputational and narrative strategy: a way to prove originality and technical credibility to the world and to be seen as an originator, not a follower.
Finally, in 2025, there has also been explicit encouragement from the government for open approaches and ecosystem building. Whether one interprets that as industrial policy, security posture, or a diffusion mechanism, it reinforces the direction of travel.
This is the lens I’ve been using when I wrote “Has China Gone All In on Open-Source, Open-Weight, and Why?” — and it’s also why I don’t view open-weight as merely a “model release choice.” It’s an ecosystem strategy.
Proem to 2026: Three Working Theses
My head has been moving away from pure LLM competition for a while. Not because models don’t matter, but because the locus of competition is shifting. In 2026, I keep coming back to three things: vertical AI moats, interface shifts (especially voice), and the emerging “AI as an operating system” race.
Startups: vertical AI + data moats will matter more than “general copilot.”
First, vertical AI and data moats will matter more than “general copilot” now that we move away from LLM labs. This is a theme I’ve been writing about for months, and it’s the same thesis James Wang from Creative Ventures has articulated in his own way: durable moats are data, workflow integration, and distribution inside a specific vertical.
And this concept was further echoed by the team at Greylock: “With reduced barriers to building AI applications on LLMs, data is arguably the most important currency in building a differentiated position…An even stronger data position would be one where usage of the product itself generates a dataset (e.g., customers labeling their own data or developing a dataset on the interactions with the product itself).”
Following that, the next step, which Alex Immerman described in the a16z Big Ideas 2026 memo, is that vertical AI leaves “single-player mode.” Real work in housing, healthcare, finance, logistics, fashion, and industrial operations is multi-stakeholder. In 2026, AI systems start coordinating across these nodes — agents talking to agents, context persisting across handoffs, and compliance or reconciliation occurring across chains rather than within a single user’s chat window. This is why the strongest vertical AI companies will increasingly resemble operating environments for their sectors rather than standalone apps: work is collaborative, and value compounds when coordination becomes native.
Interface moves to voice
Second, interfaces shift toward agent-first design. And we’ve been writing about how the mobile phone and chatbot text format will not be the most natural AI interaction entry point. This is the thread I’ve been writing about for a year: will the interface still be the phone and the smartphone, or something else? Stephenie Zhang’s a16z Big Ideas 2026 piece sharpened it in a way I found clarifying: we start designing for agents, not humans. The internet was optimized for on-device. We needed our eyes to see and our hands to type/navigate. But if anything, the conversation at the Meta event in Singapore made me realize that I was limiting my thinking to my limited worldview, which is the knowledge workers’ world. For many who use technology, they would be better off with their eyes and hands freed up. Think food delivery workers who can then keep their eyes on the road while hearing instructions for navigation or seeing directions through their glasses.
Wearables then graduate from “notifications on your face” to “sensors for your agent.” I still don’t think smart glasses will be mass-market in 2026. Still, the direction of where this is going is clear: cameras, mics, displays, and your agent doing real-time multimodal understanding. And the technology goes from seeing to understanding - who you’re talking to, what document you’re looking at, and what’s on the whiteboard.
This is also where I want to be precise about what I mean when I say “the interface changes.” I don’t simply mean we swap typing for talking. The second layer of this is that AI becomes reactive rather than passive. Instead of a tool you open when you remember it exists, it becomes a layer that can absorb context as you move through the day, retrieve what matters, and surface the right action at the right moment. Multimodality and long-form memory stop being “features” and start becoming baseline.
Voice sits at the center of that shift. Gavin Baker also touched on this in industry podcasts, that once models are good enough, voice becomes the most natural and defensible interface. I agree — and the nuance is that voice only becomes transformative when paired with memory. It’s not people yelling commands at devices in public. It’s short voice snippets capturing meetings, ideas, and micro-tasks; wearables providing continuous context; and a long-form memory that makes the agent feel like it actually knows what you meant yesterday, not just what you typed five seconds ago.
Paradigm Shift: AI Native OS
Who will be the subsequent default AI-native OS? We wrote about how OpenAI is pursuing a WeChat Mini-program-like operating system. Now, ByteDance is rewriting the whole playbook with its partnership with ZTE. Each technological disruption and paradigm shift creates an opportunity for a new generation of companies to challenge the incumbents. From Nokia’s Symbian to Android and Apple’s iOS, we’ll see a new OS emerge from this situation.
Mini-app ecosystems become the “new browser tabs” in that world. This is where I find myself aligned with what Anish Acharya at a16z has highlighted about the mini-app trend: the winning distribution surface isn’t necessarily one super app; it’s the ability for an ecosystem to quickly spawn many small, task-specific experiences and let users — or agents — stitch them together. In the West, that looks like the GPT Store and browser extensions, and maybe whatever Apple does with on-device models and Siri.
In China, it seems like WeChat mini programs, Alibaba’s Qwen to agents, and quite literally the ByteDance “AI OS” are moving to this space.
The real competition now becomes the AI native operating systems - it goes beyond the entry points/ interface we know today. This is a thread I’ve been building toward across pieces on Tencent, Alibaba, and ByteDance: competition is moving from model-versus-model to operating system-versus-operating system. Technological paradigm shifts are fundamental changes in technology that reshape industries and society.
And an AI OS is no longer looking like a metaphor. What we’ll now see is a new operating system that orchestrates multiple models, supports multimodality capabilities, embeds long-term memory, integrates deeply with distribution (WeChat, TikTok, Taobao, Office, iOS), and provides extension layers such as mini apps, plugins, or agent frameworks.
As we enter 2026, standalone model APIs without distribution, data, or vertical depth will feel increasingly exposed. The real moat lies in ecosystems that can unify models, tools, interfaces, and users into coherent operating environments and seamless experiences, and the cherry on top will be proprietary industry data. That’s where the next phase of AI strategy and differentiation will unfold (I think).
I have a few interviews (Kyle Chan, Fellow at Brookings; Alan Zhang, Portfolio Manager at an HK-based hedge fund; Zixuan at Z.ai (Zhipu); and Kevin Xu at Interconnected) that I will release over the next two weeks. I hope you enjoy them.
And again, thank you for your support in 2025. I hope you have a restful and joyful holiday and a very, very, very happy New Year!
Links to everything referenced
AI Arms Race Far From Over: Chips is Only Half the Game, Infrastructure is the Other
AI Arms Race Far From Over: Chips is Only Half the Game, Infrastructure is the Other
For the more polished op-ed that was published in Fortune, see here. Below is the unedited (ramble) edition.
The Jevons Paradox in AI Infrastructure: DeepSeek Efficiency Breakthroughs to Drive Energy Demand
The Jevons Paradox in AI Infrastructure: DeepSeek Efficiency Breakthroughs to Drive Energy Demand
When steam engines became more efficient in the 1800s, coal consumption skyrocketed. As AI gets dramatically cheaper, history is about to repeat itself.
Behind the AI Arms Race: U.S. vs. China Cloud Computing
Behind the AI Arms Race: U.S. vs. China Cloud Computing
I had the pleasure of meeting JS Tan over the Christmas break in Hong Kong. JS was a software engineer with years of experience working in Microsoft’s Cloud and AI division. He’s currently a PhD candidate at MIT, researching the political economy of innovation with a focus on the cloud.
Why China’s AI Strategy Differs From That of the U.S.
Why China’s AI Strategy Differs from that of the U.S.
Hi everyone, as promised, this is Part Two of the collaboration series with JS Tan 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 c…
Divergent Approaches to AI Commercialization (Part 1): Comparing the U.S. and China, co-written with JS Tan
Divergent Approaches to AI Commercialization (Part 1): Comparing the U.S. and China
Hi all, this piece is again in collaboration with the wonderful JS Tan. His expertise in cloud is invaluable, and working with him is always a learning experience for me. For more of his thoughtful work on innovation, labor, and cloud, check out his newsletter,
Electric Dreams, Robot Reality: China’s EV-to-AI Evolution
Electric Dreams, Robot Reality: China's EV-to-AI Evolution
*Updated information about China’s “AI Plus” initiative, which was introduced during the most recent two sessions.
Rise of China’s Robotics Industry: from Manufacturing Arms to Embodied AI
Rise of China's Robotics Industry: from Manufacturing Arms to Embodied AI
Hi all, as I mentioned, I finally got my Manus AI invite, and I was stoked to try it out. Manus’s research capabilities supported this deep-dive piece, and I must say it's comprehensive and efficient. Even though the pro version of OpenAI limits deep research to twice a day (I guess that’s good; it limits people’s abuse of it), Manus was quite impressiv…
The Bottlenecks for Embodied AI Development & Talent Outflow
Has China Gone All In on Open-Source, Open-Weight, and Why?
Has China Gone All In on Open-Source, Open-Weight, and Why?
Hi all, I was so honored to be invited to join hosts Kevin and Aleks on BBC4’s show Artificial Human today (new series to be aired in June). We talked about our usual topics- China’s AI strategy, China’s embrace of AI, technical optimism, and the rise of DeepSeek, but one thing that caught me thinking when Kevin asked me the simple question of “what is …
Benchmark’s First China AI Investment & I Finally Tried out Manus AI
The Ghibli Hype: Next Frontier is in Video & Introducing China’s Leading AI Video App, Kling
The Ghibli Hype: Next Frontier is in Video & Introducing China’s Leading AI Video App, Kling
Last week, design software maker Figma submitted paperwork to the U.S. Securities and Exchange Commission for an initial public offering. This comes just over a year after it uncoupled from Adobe, following a public courtship of $20 billion, largely due to concerns about monopoly and regulatory controls.
Why Tencent’s Integration of DeepSeek Into its AI App is a BIG DEAL
Why Tencent's Integration of DeepSeek Into its AI App is a BIG DEAL
Hello to you all. Happy Valentine’s <3. Hope you have a good weekend.
Integration at the Speed of Light: Tencent’s WeChat Embeds DeepSeek
Integration at the Speed of Light: Tencent's WeChat Embeds DeepSeek
It seems so hard to keep up with the happenings of the AI world these days, especially with what’s happening with China’s AI industry. When DeepSeek R1 was released, I wrote a piece predicting that China's AI applications would truly take off by 2025
James Wang on Why American Investors Should Know About China on Differentiated Understanding (AI Proem)
David Fishman on China’s Grid Capacity on Differentiated Understanding (AI Proem)
Tom Nunlist on AI Plus and Understanding China’s Economic and Policy Goals On Differentiated Understanding (AI Proem)
Jing Yang on China’s Big Tech from Internet to AI, Capital Decoupling on Differentiated Understanding (AI Proem)
China vs. Silicon Valley: AI & Tech Culture with Jasmine Sun
Howard Marks (Oaktree) memo “Is It a Bubble?”
https://www.oaktreecapital.com/insights/memo/is-it-a-bubble
a16z Big Ideas 2026: Part 1 (includes Stephenie Zhang: “Creating for agents, not humans”)
https://a16z.com/newsletter/big-ideas-2026-part-1/
Gavin Baker on Patrick O’Shaughnessy’s Invest Like the Best
AI Eats the World: Benedict Evans on the Next Platform Shift
Paradigm Shift Machine, Part 1: Technology increases access to what is scarce
Fantastic piece, thank you Grace! That diffusion into the real economy is only going to become more visible in 2026, along with the energy (and water) supply as a (very) hard ceiling. 🙏
Truly excellent job, great review! Learned a lot. Keep up the great work 👍 Best wishes in 2026, Grace!