China's AI Labs Went From 15 to Six. The Compute Gap Is 15-to-1.

Jordan Schneider, host of the China Talk newsletter and podcast, joins Jacob Effron on Redpoint's *Unsupervised Learning*. The two have been friends for about 15 years. Effron asks for the wide shot: what happened in the Chinese AI ecosystem before ChatGPT and after it. Schneider's first cut is a paradigm split. Pre-ChatGPT, Chinese companies were doing AI, and some of the vision work was world class, with early inklings of robotics models. They were not, however, putting much money into the "scaling is all you need" transformer stack that produced OpenAI, Anthropic, and the world of the last three years.

ChatGPT was a genuine wakeup call. There was hand-wringing that America had built a cool thing China did not have. There was also energy, because capital, talent, compute, and electricity were still there after a long dry spell. Schneider dates that drought from the 2018–2019 tech lash through the fall of 2022, COVID included. Over 2023 and 2024 a cluster of labs spun up. If you draw the circle of "credible" generously, he puts the count at about 12 or 15.

Effron wants to know how coordinated that burst was. Schneider's answer starts with a long-running preference: after the reform era, and under Xi in particular, the Chinese government has favored hardware over software. Software companies were not conceived in Beijing as contributors to national power. That is the root, in his telling, of the drama around Ant Financial and Didi. Decision-makers did not see software-led firms as doing the fundamental work that would make China flourish the way the party envisioned.

The implied social theory was harsher than neglect. Great engineers working on algorithms and models were wasting their time rotting children's brains when they could have been designing cars and weapon systems. That is why Schneider finds 2023 fascinating: the expansion was almost entirely private-money driven. Domestic VCs were, in his phrase, really just sad. Tearjerker articles asked what they were supposed to invest in after the consumer-internet crackdown. Large models gave them a new story to tell. And because Beijing banned Western closed models from deploying inside China very early, those labs inherited a protected market — a pattern Schneider says has shown up in many Chinese corporate stories over 30 years.

Hardware is the other ledger. China has spent hundreds of billions of dollars over 15 years trying to build a semiconductor stack that is not hostage to TSMC, Tokyo Electron, and ASML. Software got very little of that money. DeepSeek, Alibaba's Qwen, MiniMax, and Kimi, Schneider insists, are not Chinese-government innovation stories. They are Chinese engineering, Chinese private capital, and big software companies that got hated on for five years, then delivered models that compete globally.

The major players are a mix. DeepSeek is the weird exception: a hedge-fund founder who thought models were cool, bought a lot of Nvidia chips before they became impossible to buy, and moved a team from high-frequency trading models to large language models. The lore, Schneider says, is that Liang Wenfeng's English was not even good enough for China's top graduate schools, so he was not about to move to the United States. He made a point of hiring only domestic talent because he thought it was underpriced. Alibaba is the other pole: China's biggest cloud provider, owner of the largest e-commerce platform, plus payments and finance, with enough computer-science talent to keep spending. ByteDance had talent, capital, and compute. A handful of smaller labs raised venture money — eight figures, low nine figures, business model TBD — often on the back of Western-lab pedigrees.

Those Western credentials still do double duty. They are how researchers prove themselves to established Chinese tech firms and how they raise money. Schneider says that is changing, and the open question of the past five years — can Chinese engineers do this, are they creative enough — has been answered in the affirmative. It is a big country with a lot of smart people. Look at Western labs and about a third of the last names are Chinese; of that third, he estimates two-thirds to three-quarters grew up in mainland China. The exchange of people and ideas has continued even as geopolitics has gotten louder.

Whether the collaboration lasts is less clear. Schneider is headed to NeurIPS in San Diego. This is the first year the conference has had to spin up a second, off-site gathering in Mexico because Chinese researchers struggled to get visas to the most important AI meeting in the world. He treats that as emblematic. This is dual-use technology and the growth story of the decade. The two ecosystems are not in a perfect open market, and playing nice is getting harder.

After DeepSeek, Chinese labs moved collectively toward open-sourcing models. Schneider's reading is that it was a market-adoption play and a race to the bottom. If a really good model is open and free, it is hard to sell a closed model that is not much better. The United States, over the past three years, did have a considerable gap: if you wanted the best capabilities, you paid. Meta tried the open strategy and then backed out. What also changed is that you could get pretty close without tens of billions of dollars in training runs, which made it easier for Chinese players to sell something usable.

The fun research year is over. The 15 labs they started with are now down to five or six. DeepSeek is still an exception; most of the rest are aligned with Tencent or Alibaba. Use cases have gotten more straightforward. If you are Amazon, you want a model so ChatGPT does not take over the buying ecosystem; Alibaba's logic is the same. OpenAI has not really been born in China. You have a few small players fussing around with pretty good models, and a handful of giant companies that can pay for frontier work because other businesses throw off cash.

The structural difference is venture capital. There is more VC money sloshing around in the West to scale an OpenAI or an Anthropic until it can compete with Google. Schneider does not see that relative independence in any of the new Chinese entrants. Beijing, for now, is willing to let firms fight it out. BYD was not anointed. Huawei was not anointed; it battled for decades. He does not think the technology today is an *AI 2027* "one run to rule them all." It is a repeated game, and business models are still being figured out, let alone whether AGI arrives once you collect enough chips in a basement. Chinese firms can still rent the best Nvidia chips abroad — ByteDance on Oracle in America is the headline version — so the compute constraint is as much capital as coordination.

China's published AI action plan, Schneider says, is fine and not super dramatic: more AI in education and healthcare, open models are cool, plus platitudes such as 70 percent adoption at all companies by 2035. The feared crackdown on model outputs has not really arrived. People once argued you could not make a good model under Chinese censorship. Schneider's retort is that if you can make a model not racist, you can make it not say politically controversial things, and those queries are a tiny slice. Users are asking about basketball shoes, not 1989. Semiconductor industrial policy, by contrast, was rocky: a roughly $100 billion fund whose top seven people have been disappeared for corruption, reportedly after Xi grew frustrated with the pace of indigenization. Software AI has been surprisingly low-touch. There is more downside than upside, he thinks, if the "eye of Sauron" turns this way.

The Politburo still does Mao-era study sessions: bring in experts, sit through a lecture, ask questions. There was an AI one in May. The guest list, which China Talk wrote up, was anodyne. AI is important; we need to care about it. Not much doom. There is no Chinese Eliezer. The right lens, Schneider says, is national power and political stability. Job displacement is understood as both opportunity and risk, and officials compare it favorably to the early-1990s state-owned-enterprise reforms. They are relatively confident the system can absorb the shock. Safety talk from the government is a line here or there. The more interesting signal is private: a top AI person at Alibaba spent 30 minutes on AGI at the company's annual conference, in language Schneider found very analogous to San Francisco, which was remarkable coming from such an established player.

The information flow is one-way in a way that frustrates him. Entire Chinese media organizations translate Silicon Valley interviews. In the other direction he counts about ten people whose job is to explain the Chinese AI ecosystem to the United States. Day to day, though, founders are not living in a race narrative. About 90 percent of the worry is competitive pressure to stay alive. American AI labs have started working 996; China never stopped. Liang Wenfeng's interviews have a humanistic register — a quest to open new scientific vistas. The other pole is Zhang Yiming. Old social posts from the late 2000s and early 2010s made him look liberal and frustrated with censorship, running a tech-enabled BuzzFeed equivalent. As Douyin grew, he and his leadership team disappeared for a few days, then at 5:12 on a Monday morning he posted an apology: content moderation had been poor, they needed to serve the party better, they had not upheld socialist values. Schneider's conclusion is grim. Many leading lights have a cosmopolitan view of what they are building. They live in a system where that does not matter. If the government keeps being run the way Xi has run it for 15 years, the operating space for positive-sum interaction will shrink.

Effron asks whether the financing gap forces Beijing to step in. Schneider wants a net assessment. To deploy great models at scale you need talent, energy, and compute. China has as much talent as anyone. It probably has enough energy for the next few years. Compute is the real constraint. The Trump administration has continued Biden-era policy of not giving the Chinese mainland as much access as it would want. The contest, then, is whether Western finance plus TSMC plus the entire Western semiconductor-equipment supply chain can outscale Huawei from a lead of probably 10x, maybe 15x.

Electric vehicles and solar are Schneider's caution. Those were moderately difficult emerging technologies that China scaled in a way that dwarfs the rest of the world. He still thinks semiconductors are different. Beijing has had an open tab on the hardware ecosystem, spend-as-much-as-it-takes, and they are where they are today. "The Chinese government can spend a lot of money, but global capitalism can spend a whole lot more." Alibaba is worth about $200 billion; Nvidia is worth $4 trillion or $5 trillion. The swing sources of capital are coming America's way. He does not see a cavalry in the form of Japan switching sides.

Asked whether the gap between American and Chinese models will be larger or smaller by the end of next year, he separates the model from the ability to serve it. If you think Claude and OpenAI rate limits are bad, try a Chinese model. Serving takes compute and money. As long as the wall keeps China from buying the best chips from a producer that makes ten times as many as the domestic competitor, the West has a big edge. That edge may not show up as a huge closed-model quality gap. It shows up in how many tokens each person gets, which is the thing that is going to be constrained in China relative to the West for a long time.

Schneider's policy history starts in 2015–2016, late Obama, when the American consensus shifted from "engagement liberalizes China" to discomfort with the direction Xi was taking. Trump's first term focused late on telecommunications and Huawei. When the Biden team arrived, a handful of people in the White House already saw how important compute would be to the new paradigm. He names Jason Matheny, Ben Buchanan, Chris McGuire, and Saif Khan, several of them around CSET, who had been reading the papers for years. In October 2022, a month before ChatGPT, the United States moved to control exports of AI chips and semiconductor manufacturing equipment. A month later the product landed, and people who had found the controls weird suddenly understood what scaling laws implied.

The Biden years were a dance. Nvidia made chips just under the limit; the government said no. Manufacturing-equipment parameters twisted over time. Washington wanted the Dutch and Japanese to move in sync, which delayed things. TSMC was caught selling a ton of wafers to a Huawei cutout and got smacked. Schneider's grade is not great, not terrible, but on the margin it slowed the Chinese semiconductor and broader AI ecosystem. Trump's second term began with a team that seemed aligned on controls. Then David Sacks entered, China hawks exited after Laura Loomer's interventions, and Jensen Huang entered Trump's life. The will-he-or-won't-he drama over selling the best American chips to China has been head-spinning.

The H20 wrinkle is the 2025 anomaly. Those chips are good, not great — a generation behind, limited in some ways, still powerful. The administration decided to sell them into China, and the Chinese government said it did not want them. Schneider says the community is still debating why. Maybe Xi has bad information and Huawei lobbyists claimed they had better and more chips than they do. Maybe it is a negotiating ploy for Blackwells instead of a "gimpy knockoff." Before the Seoul summit, Nvidia's stock jumped 10 percent on the expectation of a sale; Lutnick, Rubio, Bessent, and Greer huddled on Air Force One to talk Trump out of it. He got off the plane and did not sell. Now there is another story that they are considering a version of it. The administration, Schneider thinks, is of two minds: keep the edge, or keep China on Nvidia for a while.

The strongest case for selling is Jensen Huang's: accelerators are software as well as silicon, CUDA improves because a global developer community contributes, and thousands of excellent Chinese engineers would rather stay on Nvidia. Block the chips and that energy goes into Huawei's stack. Commerce Secretary Howard Lutnick put it more bluntly: get them addicted to American chips. Schneider does not buy the Goldilocks landing that Ben Thompson and SemiAnalysis's Dylan Patel have written about, where you sell enough that China does not bother to indigenize. If the product is a nationally important strategic resource, domestic effort is coming whether you like it or not, supercharged by the choke points. Huawei went from a second source nobody wanted to deal with to front and center. The next administration is only three years away, and these companies are not stupid. He has tried, across many China Talk episodes, to find an expert making the sell case who is not financially tied to Nvidia's upside. He has not found that person. Consensus among people who take it seriously leans against selling the chips, and definitely against selling the manufacturing equipment.

The status quo is leakier than a cartoon embargo. Nothing stops Chinese firms from renting the best chips abroad. They are still training on Nvidia — in Malaysia, on chips they already have, or on chips they have smuggled — because Nvidia is still a lot better than Huawei. Legal sales into China might mean $10 billion, $20 billion, maybe $50 billion of extra Nvidia revenue over five years, and it is unclear how much of that is incremental versus money they would have spent renting overseas anyway. Biden's diffusion rule has been put in a weird limbo: BIS said in an FAQ that it does not exist, but it has not been crossed out on paper, which Schneider reads as an internal fight. A lot of this is TBD. He would not project a grand strategy onto an administration that, in his view, optimizes for vibes and headlines.

On Huawei's shot, he is specific. In 2020 and 2021 the chips Huawei was designing and SMIC was making were basically as good as Nvidia's from a design perspective. The bind is scaling inside Huawei and SMIC fabs without EUV. That leaves them stuck around 7 nanometers, walking down technology trees that Western fabs decided were less efficient than EUV lithography. One irony of Biden-era equipment controls is that semiconductor-equipment stocks roughly doubled from 2022 to 2023: Chinese buyers saw the writing on the wall and hoarded both chips and tools. Loopholes remain — 45-percent-owned subsidiaries, a fab controlled on one side of the street and not the other, lags between U.S., Dutch, and Japanese rules. They have a lot of equipment in the building. Relative compute is still the question. Schneider's vibe-coded infographic of the Huawei roadmap versus Nvidia and TSMC is "a 15-to-1 ratio at least for the next three years." Different assumptions about high-bandwidth memory or Huawei design skill can move that to 13 or 10. Ten or twenty years out is TBD. Near to moderate horizon, the Western edge is real.

Jake Sullivan, on Schneider's show, engaged with a hypothetical: should the United States have done it the hard way, a FATCA-style plea to allies that they simply could not sell, rather than the nice way. Chinese firms could not buy EUV, so they spent tens of billions on DUV, which is now powering the Huawei and legacy-chip buildout. The United States still holds cards because of American technology inside the machines, including ASML. Biden valued the alliances more, which produced two steps forward and one back. People hoped Trump would put the pedal down — if the Netherlands is worried about America leaving NATO, how much does one company's political economy matter — but rare-earth drama has put tighter equipment controls on ice. If Schneider could wave a wand, he would look past export controls to the CHIPS and Science Act's $13 billion for basic and translational semiconductor research. Biden was slow to spend it and did not make friends in industry. Trump canceled the nonprofit that was going to run it. Where that money goes — equity stakes, another vehicle, a corporate R&D subsidy — is open. Chunks of money at that scale for basic research do not come around often, and on a 10-to-20-year horizon that is what matters.

U.S. policy toward defending Taiwan has been strategic ambiguity: no promise to Taipei that would let it try to reinvade the mainland, no clarity to Beijing either. Biden ad-libbed several times that the United States would defend Taiwan. Inside the Trump-era right there is a live debate. Elbridge Colby wrote a book arguing America should care about almost nothing else because Taiwan makes the chips and is the way to contain China. Another strain, Schneider says, could not give two shits: 20 million people halfway across the world, versus Venezuela and drug runners. Before the Seoul summit there was a parallel fear that Trump would fold on Taiwan, including White House leaks of that concern. The new Japanese prime minister said the quiet part more out loud; China banned seafood imports. Schneider's view is that the package China would have to offer to make "selling Taiwan up a creek" acceptable to the American polity is hard to imagine.

The variable he wants people to watch is Taiwanese domestic politics. Two major parties: the DPP and the KMT. The KMT is Chiang Kai-shek's party, now, in a horseshoe twist, the one friendlier to Beijing and more open to rapprochement. The DPP is more left-wing, not formally pro-independence, but less comfortable with integration. The risk Schneider sketches is Xi getting impatient, seeing his own age, and reading trend lines that make political integration less likely. A KMT resurgence, conversely, lowers stress that Beijing will overreact into a short, sharp game. Everyone, he says, should care more about domestic Taiwanese politics; it is probably the most important variable in these dynamics.

His read on Xi is that this is not someone who goes willy-nilly invading places. More than a decade in power is a revealed preference. Scarborough Shoal, the Himalayan clash with India, fighter jets dancing in and out of Taiwanese airspace — none of that is what Putin looked like before Ukraine, after Georgia, Syria, and a first invasion of Ukraine. A world where Beijing dislikes an election result and flirts with something that looks like a blockade is nonzero. Over a five-year span he would put it in the single digits. They are not stupid about downside risk to the economy and even to party stability. The risk-reward matrix does not make a lot of sense. It did not for Putin in February 2022 either. This is not the thing keeping Schneider up at night. Ben Thompson's argument that America getting too far ahead would make China want to bomb the fabs is, he says, maybe, but hope springs eternal. The more boring fact is that the United States and China are each going to be about a quarter of world GDP for decades. A sci-fi split where America grows at 10 percent and China at 1 percent, Japan-in-1941 style, seems far-fetched.

Asked what is underdiscussed, Schneider points to robotics. The West has scaled chip manufacturing and a data-center buildout even faster than China. That is different from giant manufacturing. China has run that playbook with drones and electric vehicles. If manufacturing robots at scale and cheaply unlocks a lot of the 21st century, the broader Chinese ecosystem may end up with a real lead. Effron notes that a lot of robots are already being deployed in China, which is a data-collection advantage while the United States is still arguing about how to get machines out the door. Schneider remembers the pre-ChatGPT books that said China would win LLMs because of CCTV cameras and closed ecosystems. The relative weight of compute, data, and talent has ebbed and flowed. The United States is still stronger on models. Models and compute let you reap LLM benefits; robots at scale are what make the manufacturing business financially remunerative. A West dependent on Chinese robots is a comfortable steady state only until someone in government starts indexing on *I, Robot*. Jacob Helberg, now at the State Department, was tweeting in all caps years ago that Chinese humanoids had to be banned.

In the quickfire, Schneider says he has been spending time on military applications of AI. What you can be sure of is that militaries will not be perfectly fit for the technological moment when the next big war comes. Ukraine from 2022 to today already shows radically different configurations of weapons, humans, and drones, not because of incredible breakthroughs in three years but because of doctrine, manufacturing, and command-and-control chewing through a huge technological overhang. Enterprises get to that frontier with creative destruction and less death. Militaries play in the dark until the lights come on. He has been reading case histories, including John Ellis's 200-page *The Social History of the Machine Gun*, which layers invention, acquisition, tactics, and grand strategy. People who were wrong about the machine gun were not total idiots; the world could not see a revolution that would take hundreds of thousands of lives. Glimpses are possible — American generals who had watched the Crimean War were closer to realizing they could not line up as in 1807 — but not all lessons apply. The principle he wants is less certainty, more resilience into different futures. Technology shock is not unique. We have hit it many times.