Hi all, I acknowledge that I did not publish last week. The reason is 1/ I got super sick from food poisoning in Vietnam, 2/ I’ve felt very uninspired or brain-tangled with many thoughts and questions that are unresolved and without conclusion. There are currently too many loose ends in the China AI space, so I decided to lay them all out and see if a thread emerges. Hear my brain vomit out and help me out here.
Where is agentic AI actually taking us?
We wrote about agentic shopping led by Alibaba, agentic work obvs with Claude Cowork and all, agents like Manus, Kimi Agents and the action loop and all. People rave about its capabilities, but I keep coming back to the same question: how much value is it truly driving? Is it not just the 1% using agents right now? How much of it is just hype at this point? It's hard to understand the real economic impact. I found online this kind of random report by the St. Louis Fed saying that AI-related spending contributed roughly 0.9 percentage points to U.S. real GDP growth in the first three quarters of 2025. Sounds meaningful until you adjust for imports — then it’s about 0.4 to 0.5 percentage points, or roughly 20-25% of real GDP growth as of Jan this year.
And almost all of that is capex. Companies buying chips, building data centers, spinning up cloud. That’s spending on AI, not productivity from AI. The productivity flywheel everyone’s pricing in hasn’t started spinning. This is pretty symbolic for the whole industry across the U.S., but also in China.
So how do we justify these valuations? Price-to-earnings or price-to-dream?
If the economic value hasn’t materialized, what exactly are we paying for when we price these AI labs? The IPOs of Minimax and Z.ai were initially said to be undervalued compared to the valuations of American labs at the time, but what profit have they brought in? What sustainable business model have we seen? Let me put some numbers on it.
MiniMax generated $79 million in revenue in FY2025 and recorded $512 million in losses in just the first nine months, per its IPO prospectus. It currently trades at a market cap of roughly HKD 309 billion — about $40 billion according to their latest filings. There is no P/E ratio because there are no earnings. Zhipu did RMB 724 million (~$105 million) in 2025 revenue, up 132% year-on-year, but widened full-year losses to CNY 4.7 billion (~$683 million) according to its first post-IPO report. Its market cap sits around $44 billion. Shares are up nearly 600% since its January listing. Again, no earnings — so no P/E.
For context: JD.com has a market cap of roughly US$39-42 billion with a trailing P/E of about 16x according to Yahoo Finance. Alibaba is valued at roughly US$297-372 billion, with a trailing P/E of 18-23x. Tencent, still everyone’s favorite (‘Tencent, is the winner, it has to be right?’ - quote from U.S. investor), is at US$579 billion with a P/E of about 18-20x. Look, I’m not a finance person, but these numbers don’t make sense to me.
So MiniMax is valued in the same range as JD — a company doing over $150 billion in annual revenue with real earnings — while generating less than $80 million and losing (investing/spending) money at a rate of roughly RMB 2 billion a month. Zhipu is valued higher than JD on $105 million in revenue and nearly $700 million in annual losses.
If you remember Alan, the hedge fund PM we interviewed a while back here, he coined a term - the ‘price to dream ratio’ when we were chatting about this valuation disconnect. There’s no price-to-earnings here. This is a price-to-dream ratio. And the dream has a burn rate that gives these companies maybe 12-18 months of runway from IPO proceeds before they need fresh capital or a lifeline.
Am I not dreaming big enough, or is that saying in Chinese applicable to me right now, ‘我的(智商)贫穷限制了我的想象力’, my raw IQ is limiting my imagination.
What’s the clear strategy for any of these labs in China now?
If open source can’t sustain a business and the pressure to monetize continues, the playbook defaults to API sales. But that’s a race to the bottom. DeepSeek’s input pricing sits at roughly $0.028 per million tokens — about 1/180th of equivalent GPT pricing. DeepSeek V4 launched in March at $0.30 input / $0.50 output per million tokens. Everyone else — Moonshot, Qwen, Doubao — is or let’s say was competing on “cheaper and almost as good.”
The Chinese LLM space is the most competitive, with various labs carving a different angle: ByteDance on consumer scale, Alibaba on open weights, DeepSeek on pushing the frontier, Moonshot on agentic capabilities. Easily, the most competitive AI model market outside the U.S., with less investor money. But competitive doesn’t mean profitable. At these token prices, I don’t see how standalone API revenue covers compute costs, talent acquisition, or any of these expenditures, let alone how that can fund frontier research.
However, on the other end of the spectrum are people saying that people complaining about Anthropic making OpenClaw more expensive are just a bit delulu right now. The argument that once VC subsidized infrastructure disappears, the massive supply crunch is coming, and users will either need to host on their own device or leverage more open source or switch to much cheaper paid options, such as Chinese models. And that will be a win-win case for Chinese labs.
So this leads to the next question: Is this the end of China’s open-source AI era?
Even Alibaba’s Qwen seems to be faced with this dilemma now. With an end product- cloud to sell, the team is also faced with pressure to justify the AI spend and lack of profit. If selling cloud isn’t enough to cover the investment into AI infrastructure, and open-source model development doesn’t convert to revenue, then the economics simply don’t work. I think we’re watching the open-source window close in real time. And the startup labs that survive will either find a hyperscaler patron or figure out a new business model that does not exist right now. Anyway, that’s my consolidation thesis. And here Nathan Lambert writes about the need for an open source consortium.
This is where the concerns around privacy and security sit.
Open-source models can be jailbroken for malicious use — generating child exploitation material, assisting with bioweapons, you name it. There’s no self-regulatory mechanism in China’s open-source ecosystem, and that is likely now being flagged as regulations around the technology catch up with the technicalities.
At the recent Hong Kong AI governance conference, academics from both nations suggested the realistic path for U.S.-China cooperation might be a joint response to a major security breach. But what would that actually look like?
Anthropic just gave us a preview. They built Claude Mythos, a model so capable at discovering zero-day software vulnerabilities that they partially withheld it from public release and launched Project Glasswing to deploy it defensively on critical infrastructure. Azeem Azhar’s latest article goes into this too. But if that ‘worst case scenario’ were to happen, then that is literally the kind of threat those leading scholars and academics from both China and the U.S. were discussing.
For more context, Nathan Lambert writes about how this kind of general stroke of fear can be misguided. And a little self-promo from the event:
How much proper management is going through these labs? And how will they compete for compute?
Now, as we know, many of the top Chinese LLM labs have spun out of university research, but without the right business guidance and mindset, and a much less vibrant VC fundraising runway, they’re rushed to go public. How ready are these companies actually prepared for public scrutiny, pressure to turn a profit, and does that pressure change the initial purity of tech advancement into something else?
Beyond that is what we’ve written about, the obvious compute constraint and the fight for more. How does a lab fight against the hyperscalers? Compute is the binding constraint, and nobody’s seriously addressing how Chinese labs solve it.
U.S. export controls, now as we move into the first generation of models trained natively on the Grace Blackwells, what happens now? That capability gap widens? Will Huawei Ascends be actually viable at scale? This leads to the next question…
If AI diffusion continues, where does that lead for jobs and employment?
This is what keeps governments up at night. If the open-source era closes and AI concentrates in a few hyperscalers, the diffusion benefits to the broader economy narrow. But tbh, the worry isn’t about the top 1% — it’s sustaining the livelihood of the mass population. Should hyperscalers bear societal responsibility to create jobs, or at least not destroy them faster than the economy can absorb? If mass job losses happen, will there be a drop in fiscal income taxes? What happens to the governments?
According to Microsoft’s AI Diffusion report, it wrote that in 2H 2025, “The United States shows that leadership in innovation and infrastructure, while critical, does not by itself lead to broad AI adoption. The U.S. leads in both AI infrastructure and frontier model development, but it fell from 23rd to 24th place in AI usage among the working-age population, with a 28.3% usage rate. It lags far behind smaller, more highly digitized and AI-focused economies.” South Korea is punching way above its weight (which we’ll have a leading VC walk us through this soon on Differentiated Understanding season 2). And of course, post DeepSeek, China’s diffusion accelerated at rocketship speed.
But as diffusion happens, what happens to the job market, and what happens to the economy? Is AI really contributing to any GDP in a meaningful way, yet? And governments know that when the economy doesn’t work, other policies’ legitimacy sometimes finds itself with less support or legitimacy too. Especially in China, where political legitimacy is more directly tied to economic delivery, the stakes are sharper. This should be a discount factor on every AI investment thesis. Is ensuring AI does not disrupt the top knowledge workers more important, or is ensuring social stability through taking care of the masses more important? Matt Sheehan writes about that rising anxiety in China.
So, how much of the AI diffusion has really worked?
I'm a little skeptical. The Claw frenzy felt fleeting in China. Where are the stories now? Tencent is still everyone's favorite because of its ecosystem story, but why are there five Claw products? That fragmentation tells me nobody has found product-market fit yet. I'm questioning some of my own analyses and obsession with distribution now. We'll have someone from QClaw joining us on Differentiated Understanding season 2, so I'll push on this.
The (capability) human gap
The gap was first closed, then it was supposedly stretched again. According to CNBC's coverage, Anthropic's Mythos is soo advanced that it fears to release it to the public, worried that it could be maliciously used. How much of that is fear-mongering, how much of that is legit? Probably a bit of both. See Nathan’s write-up here.
The company hit $30 billion in annualized revenue as of April 2026, up from $9 billion at the end of 2025 — a 3x jump in four months. They now sit above OpenAI 's~$25 billion ARR. Claude Code alone reached $2.5 billion in annualized revenue. The company closed a $30 billion funding round at a $380 billion valuation in February, and is reportedly targeting a $60 billion+ IPO in October at a $400-500 billion valuation based on Bloomberg's numbers.
At ~$30 billion ARR and $380 billion valuation, Anthropic trades at roughly 13x revenue — expensive, but anchored to real, rapidly growing income. MiniMax would need to multiply its revenue by roughly 39x just to justify its current market cap at the same ratio. That's the distance between a price-to-earnings world and a price-to-dream world. And what about the capabilities gap? How will that be closed? I hear more and more - it's just the people, and some are ‘‘just the sun of the orbit" kind of thing, but of the AI-verse.
Why do some models do so much better than others? And where I'm placing my bet.
Is it architecture? Data? Or is it the end of the day, just people? Someone at Microsoft working on their model, who has access to OpenAI's open weights, told me it comes down to talent. I kept asking, really though? With all those resources, the differentiating point is just people?
Which brings me to Moonshot. Out of all the Chinese labs right now, I feel most strongly about Moonshot. There's something about their commitment to technology and not frivolity. They launched Kimi K2.5 in January and Kimi Claw in February. DeepSeek and Moonshot seem like the only Chinese labs that will lead in genuine research and are still committed to this. By default, big tech will be the winners in distribution and reach but... they're not leading the breakthroughs.
So the endgame might look like the OpenAI x Microsoft deal (pre-blowup) replicated in China: closed research labs selling to or merging with big tech distributors. If my consolidation thesis is right, the question is who buys these labs and at what price.
Then, a side tangent on Moonshot and Openclaw. During a roundtable discussion in Beijing two weeks ago, Moonshot CEO Yang Zhilin said that the industry is structurally shifting from a training-dominated era to an inference-dominated one. Now, the bottleneck is serving tokens at scale, cheaply and fast enough for agents that run for hours or days. (see original CN source here)
And then on April 10, Yang Zhilin was one of the select business leaders invited to brief Premier Li Qiang at a State Council economic policy roundtable (an honor DeepSeek’s Liang Wenfeng received as well in January 2025). Beijing is effectively pre-endorsing the company as a national champion in the agent era, reducing perceived regulatory risk and signaling that the IPO pipeline for top-tier Chinese AI firms remains open and supported. A pretty special nod that Minimax and Z.ai didn’t get.
There is something mystical about Moonshot. I want to learn more. Any intros would be appreciated.
So then what is winning?
I keep circling this question. Is winning having the most raw compute and energy capacity? Then the U.S. hyperscalers win the first half, and China as a whole wins the second half of that equation. Is it building the best model? Then Anthropic and OpenAI are pulling away — Mythos proved that. Is it the fastest diffusion — getting AI into the most hands, the most apps, the most workflows? Then maybe China's open-source ecosystem/ Alibaba, and Tencent still have a claim, at least for now.
But here's what I keep coming back to: diffusion doesn't matter if there's no economic progression underneath it. You can put AI in every app in China, and if it doesn't generate real productivity, real revenue, real jobs — then it's just adoption without value. And if there's no economic value, then even the most legitimate policies lose their footing. Governments can't defend AI investment to their populations if the returns aren't showing up in paychecks and GDP. And that's not a China-specific problem.
So the thread through all ten of these points is really one question: can any of this translate into economic reality before the capital runs out and the political patience expires? Right now, the value of agentic AI is half proven/ half still hype. Based on the conversation between Luo Fuli from Xiaomi MiMo, Zhang Peng at Z.ai, and Yang Zhilin at Moonshot, the consensus is that the companies best positioned are those solving the token supply chain through efficient architectures, inference infrastructure, and agent orchestration that makes every token count. The next phase of AI value creation will be about making models work autonomously, efficiently, and for long enough to produce real output.
As of now, China lab valuations are still running on dreams, not earnings. The only real proof of step-change capability is coming from U.S. labs with massive compute access. China labs can't replicate that under current (unfixable) constraints, thus consolidation feels inevitable. And the bet — my bet — is on which lab survives long enough to be the one worth acquiring.
Someone help me out here…




If winning consists of raising worker productivity, aiding in fixing society's most pressing problems (like curing cancer), improving the wellbeing of individuals and more - then no one is 'winning' yet as it's very early innings.
I work in biopharma. Our company was mostly experimental labs and office staff for sales/clinical trials. I was the only computer guy for a few months. AI (Claude) helped me set up biocomputational pipelines and some information products for our lab, also helped me learn existing tools and deploy them. We hired a second person in their 60s who has good experience using non-AI tools in the space and their fluency and comfort with coding helped them use AI chats to quickly adapt and deploy new tools too. A lot of people in the office just use it as a search replacement b/c it wasn't magic at generating ppts or accurate at summarizing important documents like they had expected from marketing.
I think what gets lost in the weeds is that traditional computational models utilizing transformers and diffusion have made huge contributions to this type of work in life sciences (I also had Claude put together matrices comparing tools that utilize diffusion and comparing their features to older gold standard tools for binding affinity, structure prediction, etc). A lot of those are open source tools and permissively licensed. But as a life sciences company, a lot of the productivity is from people who can code or can come up with ideas and want to figure this stuff out. There are very talented people in our labs who are terrible with computers and software. AI is a better search for them but unless AI can orchestrate and multiplex their physical workflow (overlapping timelines for lab work), it will barely affect their output. Productivity is also limited in that we still need to do work in meatspace and are limited by physics and biology for cell lines to grow and procedures to progress.