DeepSeek's Liang Wenfeng Privately Admits China Is Behind

An episode of Dan's AI Intel

A leaked closed-door call caught DeepSeek's founder telling investors the opposite of China's victory story — two years behind on a twentieth of the compute, winning by restraint, and dangerous enough to scrub.

Published · By Dan Walter

Executive summary

For eighteen months the West has treated Liang Wenfeng as the face of a Chinese AI triumph — the quant-fund founder whose cheap, open models wiped nearly $600 billion off Nvidia in a single day and turned "DeepSeek" into shorthand for China catching up. Then, in late July 2026, a recording surfaced of him talking to his own investors behind closed doors, and the man in the room said almost the opposite of the legend. China is roughly two years behind the United States, he told them, and it is behind on compute, not talent — DeepSeek runs on perhaps a twentieth of the chips its American rivals command and still cannot train its largest models without Nvidia. His entire strategy, he explained, is restraint: deliberately refusing to build the products, chase the users, or book the profits that would raise more money, because doing less is how you raise the odds of reaching artificial general intelligence. Within hours the links carrying his words were scrubbed from the Chinese internet, and within days he had shelved a fundraising round that valued DeepSeek near $71 billion rather than let his candor keep circulating.

The most important thing the leak reveals is not a secret weapon. It is that the person doing the best work in Chinese AI privately holds the least triumphant view of it — and that the system around him found honest, calibrated modesty more dangerous than any boast. Read closely, the recording is a rare, unguarded map of how one of the most consequential companies on earth actually thinks: about money it doesn't need, chips it can't get, a culture it is trying to change, and a finish line it isn't sure it will reach first. This is a portrait of the man, drawn from that map — and of why his particular blend of discipline and doubt is a better lens on the real AI race than the swagger on either side of it.

Why a leaked Chinese investor call is worth an hour of your time

Most of what the world knows about frontier AI comes from performances: keynote launches, benchmark charts, funding announcements written to impress the next investor. Almost none of it is what these companies say when they think no one outside the room is listening. That is what makes this leak unusual. It is not a product demo or a manifesto; it is a founder walking new shareholders, line by line, through pricing math, chip counts, a bet on Huawei, and a dated roadmap to AGI — the internal reasoning that normally never leaves the building.

It matters for three reasons that sit at the center of this moment in the AI revolution. First, DeepSeek is the company that proved the frontier could be reached for far less money than anyone in Silicon Valley assumed, which reset the entire industry's cost expectations. Second, its founder is the clearest embodiment of a distinctly Chinese way of building AI — research over revenue, open over closed, collective over cult-of-founder — that now competes head to head with the American model. And third, the leak's own fate, vanishing from the internet within hours, is a data point about how the Chinese state relates to its AI champions. To understand where the technology is heading, you have to understand the people steering it; this recording is the closest we have come to hearing one of them think out loud. A necessary caution runs through everything that follows: DeepSeek has never confirmed the recording is authentic, so its specific quotes are treated here as alleged — but, as we will see, the parts that can be checked keep checking out.

What actually leaked, and how much to believe

The artifact at the center of the story is a recording said to run three hours and forty-four minutes — roughly a thirty-thousand-word transcript once reconciled against the Chinese audio — from a closed-door meeting on 20 May 2026. The occasion was mundane by design: Liang was briefing incoming investors from DeepSeek's first-ever outside financing, a round that closed in June 2026 at around $7.4 billion for a company valued near $52 billion. The recording surfaced online around 24–25 July 2026, ricocheted across Chinese social media and tech outlets, and then the WeChat posts carrying it were removed within hours. Whether a government body ordered the takedown for political, economic, or regulatory reasons, or whether it was a private request from the company, is genuinely unknown — and the ambiguity is part of the story, not a footnote to it.

How much should a listener trust it? DeepSeek has not authenticated the document, so the honest tier is "alleged, and corroborated in parts." Several independent details line up with the public record. The transcript's claim that DeepSeek prices inference to earn roughly six times its compute cost — a ten-month payback on the hardware — matches the 545% theoretical inference margin the company itself disclosed back in March 2025. Analysts who were initially skeptical grew more convinced precisely because DeepSeek then paused its second funding round; a company inventing a flattering leak does not usually respond by walking away from $71 billion. The safest reading is that the recording is very likely substantially real, that the load-bearing numbers cross-check, and that a few specifics remain single-source. That is enough to learn a great deal — as long as every claim is held at its true weight.

The shock that made him a symbol

To understand why the leak was so combustible, you have to understand the story it contradicted. On 20 January 2025, DeepSeek released R1, a reasoning model that matched the best American systems on hard benchmarks — and it did so, the company indicated, for a final training run costing on the order of $5.6 million, a figure that stunned an industry accustomed to nine- and ten-figure budgets. A week later, on 27 January, the shock reached Wall Street: Nvidia fell 17% and shed roughly $593 billion of market value in a single session, the largest one-day loss in the history of the stock market, as investors suddenly doubted that the path to AI had to run through ever-more-expensive chips. Marc Andreessen called it "AI's Sputnik moment," and the phrase stuck.

Two cautions keep that story honest, and both matter for what came later. The $5.6 million was the compute cost of one final training run, not the all-in cost of building DeepSeek — the years of research, the failed experiments, and above all the enormous GPU fleet inherited from a hedge fund are not in that number, so "China built a frontier model for the price of a house" was always a misreading. And the market's panic conflated a genuine efficiency breakthrough with a verdict that America's lead was over, which it was not. But the myth was more useful than the caveats, and inside China it hardened into a national narrative: export controls had failed, the upstart from Hangzhou had leapfrogged Silicon Valley on a shoestring, self-reliance had arrived. Liang, overnight, became the human face of that story — summoned to Beijing, celebrated as proof of Chinese ingenuity. The recording that surfaced eighteen months later is the same man, in private, gently taking that myth apart.

From a Guangdong village to a quant monk

Liang Wenfeng was born in 1985 in Zhanjiang, a coastal city in Guangdong province, the son of a primary-school teacher — a background that is unremarkable by the standards of Chinese tech founders, which is itself the point. He studied at Zhejiang University, one of China's strongest engineering schools, taking a bachelor's in electronic information engineering in 2007 and a master's in information and communication engineering in 2010. The formative shock was the 2008 financial crisis, which pulled his curiosity toward markets: with two classmates he began applying machine learning to trading, and in 2015 they founded High-Flyer, a quantitative hedge fund that would grow into one of China's largest.

The detail that explains everything downstream is what High-Flyer built to trade: enormous GPU clusters. Chasing an edge in systematic trading, Liang assembled some of the biggest privately held computing infrastructure in China years before he had any intention of building a chatbot. When he turned that hardware and that team toward fundamental AI research, spinning out DeepSeek in Hangzhou in 2023, he was not a startup founder pitching for the compute to begin — he already owned it. That inverts the usual sequence. Most AI founders raise money to buy chips; Liang had the chips, and the money, and went looking for a problem worth them. It is why he can say, and mean, that money has never been his constraint. High-Flyer remains a formidable machine in its own right, managing on the order of ¥70 billion (about $10 billion) and posting an average return above 56% in 2025 — a quant fund quietly bankrolling a research lab that refuses to behave like a business.

Restraint as a strategy, not a personality

The single most revealing idea in the leaked meeting is a sentence: restraint is a strategy used to increase the probability of achieving AGI. Liang reportedly framed almost his entire plan as a series of refusals. DeepSeek will not become the next super-app. It will not chase user numbers. It will not pursue what he called unreasonable profits. It will not go closed-source. It will not spread into video generation, 3-D, or world models — lucrative adjacent markets he is deliberately handing to the rest of the ecosystem. Even the founder-genius myth he waved away: not a genius, just a group of ordinary people trying to do one extraordinary thing.

The logic is that every one of those tempting directions is a tax on the one thing that matters. Products, in his framing, are by-products — evidence that the research is working, not the goal of the company. So DeepSeek concentrates its people and its scarce chips on what Liang sees as the actual road to general intelligence: language models, chains of thought, agents, and continual learning, the ability of a model to keep getting better after it ships. This is the opposite of the diversify-and-monetize instinct that governs almost every well-funded lab. And it reframes the "efficiency" that made DeepSeek famous: the low costs are not a growth hack, they are a symptom of a company that has stripped away everything not pointed at AGI. Restraint is cheaper because restraint is the strategy.

Exhibit — DeepSeek's edge is what it deliberately refuses to build. Source: Alleged leaked transcript of DeepSeek investor meeting, 20 May 2026 (unconfirmed by DeepSeek); Liang Wenfeng 2023–2024 interviews. Compiled by Dan's AI Intel.

The economics that scared everyone

To see why this is a strategy and not asceticism, follow the money through the compute. In March 2025, DeepSeek published a figure that sounded implausible: a theoretical cost-profit margin of 545% on its inference business. The arithmetic behind it is concrete. Assume a GPU rents for about $2 an hour; a day of serving DeepSeek's reasoning model then costs roughly $87,000. If every token generated in that day were billed at the model's list price, the theoretical revenue is about $562,000 — more than six times the cost. That is the same six-fold rule, the same ten-month hardware payback, that the leaked meeting attributes to Liang. A public 2025 disclosure and a private 2026 recording describe the identical economic engine, which is one of the strongest reasons to take the leak seriously.

The point is not that DeepSeek actually pockets 545% — real utilization is lower, free traffic is enormous, and much of the model is given away. The point is that even at prices that undercut everyone, the underlying inference is wildly profitable per unit. Cheap, for DeepSeek, is a weapon, not charity. It can start a price war and still be structurally in the money, which is exactly what it has done: its V4-Pro model serves output at about $0.87 per million tokens — roughly 29 times cheaper than one leading US frontier model and 35 times cheaper than another. When a competitor's entire business model assumes AI stays expensive, an opponent who has proven it can be nearly free is the most dangerous kind.

This is also why the leaked numbers were believed so quickly. DeepSeek has spent two years resetting the industry's sense of what intelligence should cost, dragging prices down across the whole market each time it ships — the reason a phrase like "the DeepSeek moment" came to mean a sudden collapse in the assumed price of capability. So when a transcript surfaced describing a six-fold margin and a ten-month payback, it did not read as invention; it read as the private version of a public strategy analysts had already watched play out. The efficiency is not a marketing claim bolted onto the models after the fact. It is the mechanism by which a lab with a twentieth of its rivals' compute stays in the fight at all: if you cannot win by spending more, you win by needing less, and then you price so aggressively that spending more looks like a mistake. The economics and the restraint are the same idea seen from two sides.

Exhibit — At list price, a day of DeepSeek inference bills at ~6x its compute cost. The same 6x rule appears in the leaked 2026 meeting — the public number and the private one describe one engine. Source: DeepSeek public disclosure, March 2025 (theoretical 545% cost-profit margin; GPU rental assumed $2/hr, all tokens billed at R1 list price). Compiled by Dan's AI Intel.
Exhibit — DeepSeek prices its frontier model ~30x under the US frontier. An opponent who has proven the frontier can be nearly free is lethal to any rival whose model assumes it stays expensive. Source: DeepSeek published V4 API pricing (Apr 2026): $0.87 per 1M output tokens, ~28.7x below Claude Opus 4.8 and ~34.5x below GPT-5.5. Basis: list output price per 1M tokens. Compiled by Dan's AI Intel.

A research monastery bankrolled by a money machine

Here is the contradiction that makes Liang legible. He preaches restraint, disavows profit, and refuses to commercialize — and he is, as of mid-2026, reportedly the wealthiest individual in AI, worth around $36 billion, ahead of the founders of OpenAI and Anthropic. The resolution is that his wealth does not come from DeepSeek behaving like a business; it comes from DeepSeek's soaring paper valuation and from High-Flyer's very real trading profits. The fund is the endowment; the lab is the monastery. Because a quant machine pays the bills, Liang can afford the one luxury almost no other frontier founder has: he does not need his investors.

That is why the fundraising numbers read the way they do. DeepSeek's first outside round in June 2026 valued it near $52 billion, and Liang personally put in about ¥20 billion — roughly $3 billion, some 40% of the raise — buying influence in his own company rather than surrendering it. Weeks later, talks for a second round pointed to a valuation near $71 billion, which would have made him, on paper, the richest founder in AI. And then, after the leak, he simply switched it off, telling prospective backers the agreements they expected to sign would be delayed indefinitely. A founder who needed the money could not have done that. Liang could, because the whole edifice was built so that money would never be the thing that told him what to do. His independence is not a temperament; it is a balance sheet — and it is the precondition for both his restraint and, as the leak showed, his candor.

There is a genuine tension worth naming here, because critics inside China have. DeepSeek preaches restraint and disavows profit while the fund underneath it, High-Flyer, quietly compounds billions — a dual narrative in which the monk's vows are underwritten by the trader's returns. It is fair to ask whether "we don't do this for money" is easier to say when a hedge fund already made the money. But the tension cuts the other way too: precisely because the profit motive lives in the fund and not the lab, DeepSeek is free to behave in ways a normal, investor-dependent startup never could — releasing its crown jewels for free, refusing the obvious commercial land-grabs, and, when its founder's honesty embarrassed the moment, choosing silence and delay over the round. The structure is unusual, and it is the whole explanation.

The candor that got scrubbed

Strip away the strategy and the economics, and the reason the leak detonated is simpler: Liang was honest in a way that Chinese AI discourse rarely permits in public. To his investors he reportedly said that China's gap with the United States is real, on the order of two years, and that it is fundamentally a gap in computing power rather than in talent — Chinese researchers, in his telling, are not behind their American peers; their chips are. DeepSeek, he acknowledged, still depends on Nvidia hardware and cannot yet train its very largest models at the scale US labs can. The country is running, by one figure attributed to him, on roughly a twentieth of the compute — and doing remarkably well on it, which is a boast and a confession in the same breath.

Read against the official register, this is heresy. The dominant narrative inside China, amplified since the R1 shock, is one of accelerating self-reliance: export controls failed, domestic AI has arrived, the gap is closing on its own. Liang's private assessment was more careful, more hedged, and closer to sober American analysis than to national triumphalism — and that mismatch is almost certainly why it was so combustible. It is not embarrassing to admit a two-year lag on one-twentieth the compute; by any fair reading it is astonishing. But it punctured a story the moment demands, which is why the links died within hours and why the normally unflappable founder was reported to be rattled. The most subversive thing an AI founder can do in 2026, it turns out, is decline to exaggerate.

Exhibit — Liang's private scorecard on the AI race is a confession dressed as a boast. Roughly two years behind on a twentieth of the compute is, read fairly, an extraordinary result — which is what made saying it out loud so dangerous. Source: Alleged leaked transcript, 20 May 2026 (unconfirmed); corroborated in tenor by Liang's on-record 2024 remark that chip bans, not money, are the problem. Compiled by Dan's AI Intel.

The Huawei bet, and the cracks in it

If chips are the gap, the plan to close it runs through Huawei — and this is where the leak's confidence meets an untidier reality. Liang reportedly walked investors through DeepSeek's effort to break free of Nvidia's CUDA, the two-decade-old software moat that makes American GPUs hard to abandon, by rewriting its core code for Huawei's rival CANN framework and Ascend accelerators. He argued that Nvidia's moat is eroding fast, and that China's domestic chips do not lack a software ecosystem so much as production capacity: the designs exist, the fabs can't yet make enough of them. He reportedly expects a genuine breakthrough on Chinese silicon within about a year.

The reality underneath is messier, and it is the clearest illustration of why calibration matters. DeepSeek has indeed poured months into porting its stack to Huawei, and Ascend chips increasingly handle its inference. But the harder task — training a frontier model on domestic hardware — has stumbled. Its long-awaited next-generation reasoning model, R2, has not shipped; reporting indicates a training run on Huawei's accelerators failed, forcing a pivot back to Nvidia GPUs for training while Ascend stays on inference. So the honest picture is a hedge: Huawei for serving, Nvidia still for building the biggest models, and a self-reliance milestone that is genuinely close but not yet reached. This is exactly the gap between the confident roadmap and the stubborn present that the leak exposed — and precisely the nuance a triumphant public narrative erases. The bet is real and may well pay off; it simply hasn't yet.

Not Silicon Valley — and not the other Chinese tigers either

Liang's model is often described as "the Chinese approach," but that flattens two distinctions the leak sharpens. Against Western labs, the contrast is stark. OpenAI and Anthropic have moved toward closed models behind API walls, aggressive commercialization, and the machinery of eventual public listings, led by founders who are household names. DeepSeek keeps its best models open for anyone to download, refuses to optimize for revenue, and is fronted by a man who has spent years trying to stay in the background. Where the American frontier treats the model as a product to be metered, Liang treats it as research to be shared, and cost leadership as a way to build an ecosystem moat rather than a margin.

But he is just as distinct from his domestic rivals — the cohort sometimes grouped with Alibaba's Qwen, ByteDance's Doubao, Moonshot's Kimi, Zhipu's GLM, and Baidu's ERNIE. Those are, for the most part, commercialization plays: Qwen is an infrastructure layer wired into Alibaba's cloud; Doubao is a consumer engine inside ByteDance's apps; and as we covered in our Kimi K3 episode, number 33, released last week, Moonshot has begun pricing its newest open model like a premium US flagship rather than undercutting on cost. They are open where it helps them sell something. DeepSeek is open because Liang believes the ecosystem is the point, and he has explicitly rejected the ByteDance–Tencent path of turning a model into a super-app. Among the Chinese labs, in other words, DeepSeek is the purist — the one that most resembles a research institute that happens to have a hedge fund attached.

Exhibit — DeepSeek sits in a quadrant of its own: open AND research-first. The others are open where it helps them sell; DeepSeek is open because it believes the ecosystem — not the product — is the prize. Source: Liang Wenfeng 2023–2026 interviews and alleged leaked meeting; public positioning of Qwen, Doubao, Kimi. Compiled by Dan's AI Intel.

What a culture does with a Liang Wenfeng

To read Liang only as an individual is to miss half of him; he is also a product of, and a bet by, a particular system. DeepSeek is one of Hangzhou's "six little dragons," a clutch of hard-tech firms — alongside the robotics maker Unitree and the studio behind a hit domestic game — nurtured by years of Zhejiang provincial policy, including a "chain leader" system that wires the party-state into local supply chains. In February 2025, weeks after R1, Liang sat among the tech elite as Xi Jinping urged China's founders to "show their talent" in the contest with America. He is, in a real sense, a national champion — state-adjacent, politically useful, and expected to embody self-reliance. Which is exactly why an honest word about being two years behind could not be allowed to stand.

The deeper cultural signal, though, is in how he builds and how he talks. DeepSeek's org is famously flat — bottom-up, with researchers, many of them fresh graduates, handed compute and autonomy, no rigid hierarchy, hiring for "passion and curiosity" over experience. One of its signature efficiency breakthroughs reportedly began as a young researcher's personal side-interest. And in interviews going back to 2023, Liang has returned again and again to a single theme that reads as a cultural mission statement: China, he argues, has spent thirty years of the information-technology wave as a follower and a copier, and its real deficit is not capital but confidence — a belief that Chinese firms can do original, "hardcore" innovation rather than commercialize other people's. "China should gradually become a contributor rather than a free-rider," he has said. When DeepSeek's earlier V2 model startled Silicon Valley, his reaction was pointedly deflating: among the enormous volume of innovation happening in the US every day, he noted, this was an ordinary one — what surprised people was simply that a Chinese company had joined the game as an innovator rather than a copier. His diagnosis of China's real weakness is not money or brains but nerve: "Our issue isn't capital," he has argued, "but rather a lack of confidence and a lack of knowledge on how to organize top-tier talent for effective innovation." His restraint, his open-sourcing, his refusal to chase quick money — all of it is, in his own framing, an attempt to prove that a Chinese company can lead on ideas, not just undercut on price.

That mission is also where the culture contrast bites hardest. The Silicon Valley story is individual and urgent — a named founder racing to own a winner-take-all market, monetizing fast to fund the next lap. Liang's is collective and patient: hire for curiosity over credentials, hand young researchers real autonomy, give the work away, and measure success in whether China becomes a place that originates technology rather than commercializes it. It is telling that he insists he is not a genius and that DeepSeek is "a group of ordinary people" — a register almost unthinkable from an American founder selling a vision to investors. Where one culture treats the model as property and the founder as prophet, the other treats the model as a contribution and the founder as a custodian of a longer national project. That, more than any benchmark, is what the leak accidentally laid bare.

Where this goes — for DeepSeek, and for the race

Forecasting from here means naming the hinges. For DeepSeek, the near-term one is R2 and the domestic-silicon bet: if the next reasoning model ships strong and, eventually, trains on Chinese chips at scale, Liang's confident roadmap is vindicated and the compute gap starts to look temporary. If R2 keeps slipping and frontier training stays chained to Nvidia, then the leak's quiet admission — that the biggest models still can't be built without American hardware — becomes the defining constraint, and export controls will have done more than Beijing admits. Restraint buys focus, but it cannot manufacture chips.

For the broader race, the recording reframes the contest. If Liang is right that the gap is compute rather than talent, then the American lead is real but rented — a function of chip access that policy created and policy could erode, not a permanent edge in genius. The efficiency China has been forced into by scarcity could become structural: models that do more with less travel better into a world of cheaper, more distributed compute. And the wildest variable is what Liang himself keeps pointing at — continual learning and autonomous iteration, the step from models that ship frozen to models that keep improving, with embodied intelligence as the endpoint. If that ladder holds, restraint aimed squarely at AGI may age better than diversification aimed at revenue. The uncomfortable takeaway for the West is that its clearest competitor is not a boaster to be dismissed but a disciplined, self-critical operator who told his own investors the sober truth — and got punished for it. In an arms race, the honest player is easy to underestimate. That may be the most dangerous thing about him.

Exhibit — Liang's roadmap climbs one ladder — every rung aimed at AGI, not revenue. Source: Alleged leaked transcript, 20 May 2026 (unconfirmed by DeepSeek); Liang Wenfeng 2023–2024 interviews. Compiled by Dan's AI Intel.

Bottom line

The leak did not expose a hidden weapon or a fraud. It exposed a temperament — disciplined, self-critical, allergic to hype — attached to a balance sheet that lets it act on principle, inside a system that would rather he perform confidence than tell the truth. Liang Wenfeng runs the opposite of the swaggering AI founder the world imagines: he says China is behind, he wins by doing less, he gives his best work away, and he answers to a hedge fund instead of investors so that no one can tell him to stop. Whether restraint plus open weights plus forced efficiency actually reaches the frontier before America's compute advantage compounds is the open question of the next two years. But anyone trying to read the AI race should keep this recording close, because it is the rarest thing in the whole contest: one of its most important players, caught telling the truth.

Sources

Transcript

Alex: Picture the most important AI founder in China sitting down with his own investors, behind a closed door, and quietly telling them the opposite of everything the world believes about him.

Sam: Wait — the opposite? This is the guy who's supposed to be China's answer to Silicon Valley.

Alex: He tells them China is roughly two years behind the United States. And within hours of those words leaking, they're scrubbed off the Chinese internet.

Sam: Okay. That's the kind of sentence that makes you stop walking. Behind means behind — from the man everyone's holding up as proof they've caught up.

Alex: Welcome to Dan's AI Intel — the show where we take the one question in AI that actually matters that week and dig past the hype and the fear until we can see what's really going on underneath.

Sam: I'm Sam, and as always I'm here with Alex, who is going to keep us honest on the numbers, because today the numbers are the whole story.

Alex: They really are. So here's the setup. Late July, 2026. A recording surfaces of a man named Liang Wenfeng — the founder of DeepSeek — talking to a room of new shareholders. And the reason the whole tech world lunged at it is that this is a company that already shook the earth once.

Sam: This is the DeepSeek that wiped, what, hundreds of billions off Nvidia in a single day?

Alex: Nearly six hundred billion dollars off Nvidia in one session. Cheap, open AI models that made Silicon Valley wonder if it had been wildly overspending to reach the frontier. That DeepSeek.

Sam: So the deeper question I can't let go of isn't really "what did he say." It's this: what happens when the person doing the best work in a national race privately holds the least triumphant view of it? And what does it tell you that his own system found his honesty more dangerous than any boast?

Alex: That's exactly the thread. Today we walk through what actually leaked and how much of it to trust; who this man really is, from a village in Guangdong to a quant-fund monk; the strategy he calls restraint; the economics that terrified the market; and then the culture question — what China and America each reveal about themselves in how they build this stuff.

Sam: And there's a turn in here — a reason this particular leak got buried instead of celebrated — that genuinely surprised me. We'll get there.

Alex: Before we get into it: if you're finding these useful, follow the show wherever you're listening. It's free, it's one tap, and it means the next one lands in your feed automatically.

Sam: So here's my first honest reaction, before we even get to who he is. Why does a leaked investor call matter this much? Companies leak all the time.

Alex: Because almost everything the world knows about frontier AI comes from performances — keynote launches, benchmark charts, funding announcements written to impress the next investor. This is the opposite of a performance. It's a founder, in a closed room, walking new shareholders line by line through his pricing, his chip counts, his bet on Huawei, and a dated roadmap to artificial general intelligence.

Sam: So we're basically eavesdropping on how one of the most consequential companies on earth actually reasons when it thinks no one's listening.

Alex: That's the rarity. We get keynotes by the dozen. We almost never get the internal monologue. And it matters for three reasons that sit right at the center of this moment. One: DeepSeek is the company that proved the frontier could be reached for far less money than Silicon Valley assumed — it reset the whole industry's cost expectations. Two: its founder is the clearest embodiment of a distinctly Chinese way of building AI — research over revenue, open over closed, collective over cult-of-founder. And three: the leak's own fate — vanishing within hours — is itself a data point about how the Chinese state relates to its AI champions.

Sam: So the disappearance isn't just a footnote. It's part of the evidence. Okay. I'm sold on why it matters. Now make me believe it's real. Okay, first honest question, because I want to know how much weight to put on this. Someone posts a recording online. Why should anyone believe a word of it?

Alex: Right skepticism, and it's the first thing we have to nail down. So the artifact itself: it's said to be a recording running three hours and forty-four minutes. Reconciled against the Chinese audio, that's roughly a thirty-thousand-word transcript.

Sam: Three hours and forty-four minutes. That is not a soundbite. That's someone thinking out loud for an entire afternoon.

Alex: And the occasion was mundane by design. May twentieth, 2026. Liang is briefing incoming investors from DeepSeek's first ever outside funding round — a round that closed in June at around seven point four billion dollars, for a company valued near fifty-two billion.

Sam: So this isn't a manifesto or a product launch. It's the boring internal version. The pricing math, the chip counts, the actual bets.

Alex: Exactly. The stuff that normally never leaves the building. It surfaces online around the twenty-fourth, twenty-fifth of July, tears across Chinese social media — and then the posts carrying it get removed within hours.

Sam: Removed by who?

Alex: And that's the honest answer: genuinely unknown. Could be a government body, for political or economic or regulatory reasons. Could be a private request from the company. Nobody outside the room knows, and the ambiguity is part of the story, not a footnote to it.

Sam: See, that's the part that gives me pause. If you can't verify who pulled it down, how do you trust what's in it?

Alex: So you separate two questions. One: is it authentic? DeepSeek has never confirmed it is, so honestly the highest tier we can claim is "alleged, and corroborated in parts." Two: do the load-bearing pieces cross-check against the public record? And there, it holds up unnervingly well.

Sam: Give me the strongest one.

Alex: The transcript claims DeepSeek prices its AI to earn roughly six times what the computing costs it — a ten-month payback on the hardware. Now here's the thing: back in March 2025, DeepSeek itself publicly disclosed a theoretical inference margin of five hundred and forty-five percent. That's the same six-fold rule. A public number from 2025 and a private number from 2026 describing the identical engine.

Sam: Huh. So the leak isn't introducing a new claim there — it's the private-language version of something they'd already said out loud.

Alex: And then the behavioral tell. When the leak went viral, DeepSeek paused a funding round that would have valued it near seventy-one billion dollars. A company that fabricates a flattering leak does not usually respond by walking away from seventy-one billion.

Sam: No. You don't fake good news and then set fire to the money. Okay — so where does that leave a careful listener? Because I don't want to overclaim.

Alex: And here's a detail I love, because it shows how the credibility built over time. The analysts who first looked at this were skeptical — of course they were, it's an unverified recording. What changed their minds wasn't a new document. It was that DeepSeek then paused the round. The skepticism actually inverted: the company's own behavior became the corroboration.

Sam: That's a weird kind of proof, isn't it. Not "here's evidence the leak is true," but "here's the company acting exactly like a company that got caught being honest."

Alex: Actions confirming words that were never meant to be heard. So we hold it in tiers. The pricing math and the six-fold margin — corroborated against the public record. The two-years-behind assessment — consistent in tenor with things he's said on the record before. And then a handful of very specific quotes stay single-source, and we flag those as we go.

Sam: So it's not "believe all of it" or "believe none of it." It's a layered thing. Load-bearing beams, checked. Decorative details, noted.

Alex: The safest reading: very likely substantially real, the big numbers cross-check, and a few specifics stay single-source. Which is enough to learn a huge amount — as long as we hold every claim at its true weight. When it's confirmed, I'll say so. When it's alleged, I'll say that too.

Sam: I like that rule. Keep the price tag on every fact. So we trust it, carefully. That means the real question becomes: why was telling the truth the dangerous move? And to feel that, we have to go back to the myth he was quietly taking apart.

Alex: So rewind to the thing that made Liang a symbol in the first place. January twentieth, 2025. DeepSeek releases a reasoning model called R1. And it matches the best American systems on the hard benchmarks.

Sam: And this is the one with the famous price tag, right? The number everyone repeated?

Alex: The company indicated the final training run cost on the order of five point six million dollars. In an industry used to nine- and ten-figure budgets, that landed like a bomb.

Sam: Five point six million. For something that goes toe to toe with the American frontier. I can see why people lost their minds.

Alex: And a week later, January twenty-seventh, the shock hits Wall Street. Nvidia falls seventeen percent and sheds roughly five hundred and ninety-three billion dollars of market value in a single session. The largest one-day loss in the history of the stock market.

Sam: In one day. Because suddenly investors are asking — wait, if you can do this cheap, why have we been buying mountains of expensive chips?

Alex: Marc Andreessen called it AI's Sputnik moment, and the phrase stuck. And think about what that phrase does — Sputnik wasn't just a satellite, it was the moment America realized it might be losing a race it assumed it owned. That's the emotional charge people attached to R1.

Sam: A wake-up call with a body count of half a trillion dollars.

Alex: But here's where I have to slow us down, because two cautions keep that story honest, and both matter for everything after.

Sam: Go.

Alex: One. That five point six million was the cost of one final training run. It is not the all-in cost of building DeepSeek. The years of research, the failed experiments, and above all an enormous fleet of GPUs they'd inherited from a hedge fund — none of that is in the number.

Sam: Oh. So "China built a frontier model for the price of a nice house" was always a misread. That's the highlight reel, not the budget.

Alex: Exactly. And two: the market panic conflated a real efficiency breakthrough with a verdict that America's lead was over. Which it wasn't. The breakthrough was genuine. The obituary for the American lead was premature.

Sam: But the myth was way more fun than the caveats.

Alex: The myth was more useful than the caveats. And this is the pattern to watch for, because it repeats. A real, narrow technical fact — we trained the final run cheaply — gets inflated into a sweeping verdict: the whole cost of AI has collapsed, and America's advantage is gone. The fact was true. The verdict was a leap.

Sam: And once a market makes that leap, it moves half a trillion dollars in an afternoon on the leap, not the fact.

Alex: Half a trillion on the vibe. Which is worth holding onto, because the leak we're discussing is, in a sense, the founder himself walking that leap back — quietly telling his own investors, "no, don't over-read what we did." And inside China it had already hardened into a national story: export controls failed, the upstart from Hangzhou leapfrogged Silicon Valley on a shoestring, self-reliance had arrived. Liang, overnight, becomes the human face of that story. Summoned to Beijing. Celebrated as living proof of Chinese ingenuity.

Sam: And he becomes the face of it whether he wants to or not. Summoned to Beijing, held up as proof of Chinese ingenuity.

Alex: Turned into a symbol. And that's the trap of the whole thing — a symbol isn't allowed to be nuanced. A symbol has one job: embody the story. So the moment the symbol quietly says "actually, the story's more complicated," he's not just being honest, he's breaking character.

Sam: So let me make sure I've got the shape of this. Eighteen months ago he's crowned as the man who proved China caught up. And the leak we're talking about is the same man, in a private room, gently taking that crown apart.

Alex: That's the whole tension of the episode in two sentences. Which raises the obvious question — who is he? Because a person doesn't casually dismantle a national myth about himself unless he's wired in a very particular way.

Sam: Yeah, I want the human being. Where does someone like this even come from?

Alex: Born 1985, in Zhanjiang — a coastal city in Guangdong province. Son of a primary-school teacher. And that ordinariness is actually the point: by the standards of Chinese tech founders, this is an unremarkable background.

Sam: Not a princeling, not a returnee from Stanford. A teacher's kid from the coast.

Alex: He studies at Zhejiang University — one of China's strongest engineering schools. Bachelor's in electronic information engineering in 2007, master's in information and communication engineering in 2010. And the formative shock is the 2008 financial crisis, which pulls his curiosity toward markets.

Sam: Markets, not AI. Interesting. So he doesn't start as a chatbot guy.

Alex: Not even close. And the way he came at markets tells you everything — not "how do I get rich," more "here's a giant, messy system full of patterns, can machine learning find them." The market was just the first hard problem he pointed the technique at.

Sam: So the through-line from day one is basically: point machine learning at a hard problem and see how far it goes. The problem just happened to be markets first.

Alex: That's a really clean way to put it. With two classmates he starts applying machine learning to trading, and in 2015 they found a quantitative hedge fund called High-Flyer, which grows into one of the largest in China. And here's the detail that explains everything downstream.

Sam: Which is?

Alex: To trade, High-Flyer built enormous GPU clusters. Chasing an edge in systematic trading, Liang assembled some of the biggest privately held computing infrastructure in China — years before he had any intention of building an AI model.

Sam: Oh, that's the whole trick, isn't it. Everyone else in AI is desperate to raise money to buy chips. He already owned the chips.

Alex: You just skipped ahead of me, and you're exactly right. Most founders raise money to buy compute so they can start. Liang had the compute, and the money, and went looking for a problem worth them. He spins DeepSeek out in Hangzhou in 2023 — not as a founder pitching for the resources to begin, but as a man who already owned them.

Sam: So when he says money has never been his constraint — that's not a flex. It's just literally true.

Alex: It's a balance sheet, not a boast. And it inverts the usual founder story completely. The normal sequence is: have an idea, raise money, buy compute, hope it works. His sequence was: already have the compute, already have the money, go find the most important problem in the world to point them at. That reversal is the seed of everything strange about him.

Sam: Right, because if you never had to beg for the resources, you never learned to think like someone who has to please the people with the resources. You're just... free.

Alex: You're free in a way almost no one at the frontier is. And the fund is no side project. High-Flyer manages on the order of seventy billion yuan — about ten billion dollars — and posted an average return above fifty-six percent in 2025.

Sam: Fifty-six percent. So the hedge fund is quietly crushing it, and using the winnings to bankroll a research lab that flatly refuses to behave like a business.

Alex: A quant machine paying for a monastery. Hold onto that image, because it's the key that unlocks the strangest thing he said in that room — the idea he builds his entire company around. He calls it restraint.

Sam: Restraint. Okay, that's a weird word for a strategy. Normally a strategy is a list of things you're going to do. What does he actually mean?

Alex: The single most revealing sentence in the whole leaked meeting is basically this: restraint is a strategy used to increase the probability of achieving AGI. And he frames almost his entire plan as a series of refusals.

Sam: Refusals. Like what — walk me through them.

Alex: DeepSeek will not become the next super-app. It will not chase user numbers. It will not pursue what he called unreasonable profits. It will not go closed-source. And it will not spread into video generation, or 3-D, or world models — lucrative adjacent markets he's deliberately handing to the rest of the ecosystem.

Sam: That is genuinely bizarre for a founder. He's listing the doors to the money and saying, I'm not walking through any of them.

Alex: And even the myth about himself, he waves away. Not a genius, he says — just a group of ordinary people trying to do one extraordinary thing.

Sam: Which, again, is such a strange thing to say to investors. The whole genre of the founder pitch is "back me, the visionary." He's saying "there's no visionary here, just a team."

Alex: It's almost anti-charismatic. And that's not false modesty for effect — it's connected to the strategy. If the company is a cult of one genius, it lives and dies by that genius. If it's a group of ordinary people with a clear method, the method is the asset, and it's repeatable.

Sam: So why? What's the logic? Because that can't just be modesty. Nobody runs a company on modesty.

Alex: The logic is that every one of those tempting directions is a tax on the one thing that matters. In his framing, products are by-products. They're evidence the research is working — not the goal of the company.

Sam: Ohh. So the app isn't the point. The app is a receipt. Proof the research is real.

Alex: That's the exact reframe. So DeepSeek concentrates its people and its scarce chips on what Liang sees as the actual road to general intelligence: language models, chains of thought, agents, and continual learning — the ability of a model to keep getting better after it ships.

Sam: And everything he refuses — video, 3-D, world models, the super-app — those are the detours off that road.

Alex: Every one of them is a detour, and in his framing a detour is a tax. Not because those markets are bad — they're lucrative, that's the point. It's that they'd pull people and chips off the main climb. He'd rather hand those markets to the wider ecosystem and keep his scarce compute pointed at one thing.

Sam: This is the opposite instinct to almost every well-funded lab, right? The normal move is diversify and monetize — build the model, then the products, then the platform.

Alex: The normal move is a portfolio. His is a spear. And that's genuinely rare — a founder with the money to diversify, choosing not to, on principle.

Sam: Okay, give me the analogy, because I can feel this one is important and I want it to stick.

Alex: Think of a climber going for a summit no one's reached. Everyone else at base camp is opening gift shops, running the café, selling the parking. Profitable, sure. But every hour and every porter you spend on the gift shop is an hour and a porter not carrying you up the mountain. Liang's bet is: strip away everything that isn't the climb, and you reach the top sooner. The refusals aren't self-denial. They're focus.

Sam: And that completely reframes the "cheap" thing, doesn't it. Everyone talks about DeepSeek being cheap like it's a growth hack. But if you strip away everything not pointed at the summit —

Alex: — then low cost isn't the trick. Low cost is a symptom. It's what a company looks like when it's refused to build anything that isn't the climb. Restraint is cheaper because restraint is the strategy.

Sam: Right, I had it backwards. I thought cheap was the weapon and restraint was the vibe. It's the reverse.

Alex: It's the reverse. And notice what it costs him to hold that line. Video generation, 3-D, world models — those aren't hobby markets, they're some of the most lucrative territory in AI right now. He's looking at real money on the table and walking past it, on purpose, to keep his people and his chips on the one road.

Sam: That's the part that would break most founders. It's easy to say no to a bad idea. Saying no to a genuinely profitable one, over and over, because it's not the summit — that takes a very particular kind of discipline.

Alex: And that discipline is the whole company. That is the perfect place to actually follow the money — because when you do the arithmetic on that "cheap," it turns into something that genuinely scared an entire industry.

Sam: All right, I've been promised scary arithmetic. Scare me.

Alex: March 2025, DeepSeek publishes a figure that sounds implausible: a theoretical cost-profit margin of five hundred and forty-five percent on its inference business. And the arithmetic behind it is surprisingly concrete.

Sam: Inference — that's the running-the-model part, not the training part, right? The cost every time someone actually uses it.

Alex: Exactly right — training is building the model once; inference is the cost each time it answers. So: assume a GPU rents for about two dollars an hour. A full day of serving DeepSeek's reasoning model then costs roughly eighty-seven thousand dollars.

Sam: Okay, eighty-seven thousand a day to keep the lights on. Got it.

Alex: Now, if every token the model generated in that day were billed at its list price, the theoretical revenue is about five hundred and sixty-two thousand dollars. More than six times the cost.

Sam: There's the six-fold rule again — the exact one that shows up in the leaked meeting.

Alex: The same six-fold rule, the same ten-month hardware payback. A public 2025 disclosure and a private 2026 recording describing the identical economic engine. Which, by the way, is one of the strongest reasons to take the leak seriously.

Sam: Ten-month payback — meaning the hardware pays for itself in under a year, and everything after that is gravy.

Alex: Under a year to recoup the chip, and then it keeps earning. That's an extraordinary return on a physical asset. Most infrastructure businesses would kill for a ten-month payback.

Sam: But hang on — I want to be careful. Does DeepSeek actually pocket five hundred and forty-five percent? Because that seems too good to be real.

Alex: No, and that's the crucial caveat. It's a theoretical margin — if every single token were sold at list price with the servers fully loaded. In reality, utilization is lower, an enormous amount of traffic is free, and a lot of the model is simply given away. They do not actually pocket five hundred and forty-five percent.

Sam: So it's a ceiling, not a bank statement. The best case if everything were billed.

Alex: A ceiling. But a wildly high ceiling is still the whole story, because it tells you the floor is generous too.

Sam: So then what's the point of the number?

Alex: The point is that even at prices that undercut everyone on earth, the underlying unit is wildly profitable. Cheap, for DeepSeek, is a weapon — not charity. They can start a price war and still be structurally in the money.

Sam: And they have. Give me the price war in one number.

Alex: Their V4-Pro model serves output at about eighty-seven cents per million tokens. That's roughly twenty-nine times cheaper than one leading US frontier model, and thirty-five times cheaper than another.

Sam: Thirty-five times. Not thirty-five percent. Thirty-five times cheaper.

Alex: Times. To put that in human terms — if the American frontier model is a thirty-five dollar meal, DeepSeek is serving basically the same dish for a dollar. And not losing money doing it.

Sam: And a dollar-versus-thirty-five isn't a competitor you negotiate with. It's a competitor who's redefining what the thing is worth.

Alex: And here's why that's lethal. If your entire business model quietly assumes that AI stays expensive — that people will keep paying premium prices for intelligence — then an opponent who has publicly proven it can be nearly free is the most dangerous competitor you can face.

Sam: Because they're not just undercutting you on price. They're attacking the assumption your whole company is built on.

Alex: That's the second-order point most people miss. It's not a discount. It's an argument — that the thing you're metering should be nearly free. And here's the mechanism that makes it lethal: it's the only way a lab with a twentieth of its rivals' compute stays in the fight at all. If you can't win by spending more, you win by needing less — and then you price so aggressively that spending more starts to look like a mistake.

Sam: So the poverty and the pricing are the same weapon. The scarcity forced the efficiency, and the efficiency became a price war.

Alex: Scarcity forged the blade. And they've spent two years swinging it by shipping. Every time DeepSeek releases, prices drop across the whole market. It's literally why "the DeepSeek moment" came to mean a sudden collapse in the assumed price of capability.

Sam: So "the DeepSeek moment" basically became shorthand for the day the assumed price of intelligence falls off a cliff.

Alex: That's exactly what the phrase means now. Not a product. A repricing event.

Sam: And quick aside, for anyone who wants to see this exact move from another angle — we did a whole episode last week on Moonshot's Kimi K3, "China Hit the AI Frontier and Gave It Away," it's number thirty-three. Same country, very different pricing choice. Worth a listen after this.

Alex: Good pointer. And it sharpens the contrast we'll get to. But hold onto the mechanism for one more beat, because it ties the whole strategy back together. Remember the restraint — the refusing to build anything that isn't the climb?

Sam: The economics and the restraint are the same idea seen from two sides.

Alex: Two sides of one coin. Which leads straight into the contradiction that, honestly, took me a second to get my head around. The man preaching "we don't do this for money" is the richest person in AI.

Sam: Wait, what? The restraint guy — the products-are-by-products, we-don't-chase-profit guy — is the richest person in AI?

Alex: As of mid-2026, reportedly worth around thirty-six billion dollars. Ahead of the founders of OpenAI and Anthropic.

Sam: That does not compute. How do you disavow profit and end up the wealthiest person in the entire field?

Alex: Because his wealth doesn't come from DeepSeek behaving like a business. It comes from two other places: DeepSeek's soaring paper valuation, and High-Flyer's very real trading profits. The fund is the endowment. The lab is the monastery.

Sam: Ohh. So it's like a university, almost. The endowment throws off returns, and that pays for the researchers who are explicitly not supposed to make money.

Alex: That's a genuinely good analogy — the fund is the endowment, the lab is the monastery. And because a quant machine pays the bills, Liang can afford the one luxury almost no other frontier founder has. He does not need his investors.

Sam: And "doesn't need his investors" isn't a small thing. Every other frontier lab is, on some level, performing for the next check. The roadmap bends toward whatever keeps the money flowing.

Alex: Whereas his roadmap can bend toward whatever he actually believes reaches AGI, even when that's giving the models away and refusing the profitable detours. The freedom to be stubborn is the rarest asset in the whole field, and he bought it with a hedge fund.

Sam: And I bet the funding numbers show exactly that.

Alex: They do, and they're wild. DeepSeek's first outside round, June 2026, valued it near fifty-two billion. And Liang personally put in about twenty billion yuan — roughly three billion dollars, some forty percent of the raise.

Sam: Hold on. He put money into his own funding round? Founders raise money to take it out, not to buy in.

Alex: He was buying influence in his own company rather than surrendering it. And then weeks later, talks for a second round point to a valuation near seventy-one billion — which would make him, on paper, the richest founder in AI. And after the leak, he just... switches it off. Tells prospective backers the agreements they expected to sign are delayed indefinitely.

Sam: A founder who needed the money physically could not do that. You can't walk away from seventy-one billion if payroll depends on it.

Alex: And think about what shelving that round actually costs him personally. On paper, that second round would have crowned him the richest founder in AI. He gave up the crown rather than let his own honesty keep circulating. That's the price of the candor, in dollars.

Sam: He'd rather not be the richest founder in AI than let the room's honesty stand. That reorders my whole sense of the guy.

Alex: He could do it, because the whole edifice was built so that money would never be the thing that told him what to do. His independence isn't a temperament. It's a balance sheet. And it's the precondition for both his restraint and, as the leak showed, his candor.

Sam: Okay, but I have to push back a little, because this is a bit convenient, right? "We're not in it for the money" is a lot easier to say when a hedge fund already made you the money.

Alex: That's the fair criticism, and it's one that Chinese critics have made themselves. There's a real tension: the monk's vows are underwritten by the trader's returns. You're allowed to raise an eyebrow at that.

Sam: I am raising it. It's a little like a trust-fund kid saying money doesn't matter. Technically true, easy to say from where they're standing.

Alex: That's a fair shot, and the critics land it. But notice it cuts the other way too. Precisely because the profit motive lives in the fund and not the lab, DeepSeek is free to do things a normal, investor-dependent startup never could. Release its crown jewels for free. Refuse the obvious land-grabs. And when the founder's honesty embarrassed the moment — choose silence and delay over the seventy-one billion.

Sam: So the structure isn't a footnote. The structure is the whole explanation. The hedge fund is why he can afford to be honest.

Alex: The hedge fund is why he can afford to be honest. Which is the perfect bridge, because the honesty is where this all detonates. Strip away the strategy and the economics, and the reason the leak went nuclear is much simpler than either. He just... told the truth.

Sam: So what did he actually say? The line that got scrubbed.

Alex: To his investors, he reportedly said the gap between China and the United States is real — on the order of two years. And, crucially, that it's fundamentally a gap in computing power, not in talent.

Sam: Not talent. So he's saying the Chinese researchers are just as good — they're just short on chips.

Alex: In his telling, Chinese researchers are not behind their American peers. Their chips are. DeepSeek, he acknowledged, still depends on Nvidia hardware and cannot yet train its very largest models at the scale US labs can. The country is running, by one figure attributed to him, on roughly a twentieth of the compute.

Sam: And this is the private version — the man himself saying the quiet part. Not an outside analyst estimating it. The founder, telling his own backers.

Alex: The founder, to the people writing him checks, with no reason to sandbag. Which is exactly what makes it credible, and exactly what makes it dangerous. He's not managing a public image in that room. He's leveling with the people who most need the real picture.

Sam: A twentieth. So one unit of chips for every twenty the Americans have — and still keeping pace on the results.

Alex: And that's the line that's a boast and a confession in the same breath. Doing remarkably well on a twentieth of the compute is astonishing. But admitting you're on a twentieth of the compute is a confession that the gap is real.

Sam: It's the two-sided sentence. Say it one way, it's the proudest thing you could claim. Say it the other way, you've just conceded the race isn't won.

Alex: And he chose to say the honest, two-sided version to the people who most needed the accurate map — not the one-sided victory version the outside world wanted. That choice is the whole character study in miniature.

Sam: Okay, help me understand why that's heresy, though. Because to my ear that's... kind of an incredible thing to be able to say? Two years behind on one-twentieth the resources — that's a flex.

Alex: You'd think so. But read it against the official register inside China, and it's the wrong tune entirely. The dominant narrative since the R1 shock is accelerating self-reliance: export controls failed, domestic AI has arrived, the gap is closing on its own. Liang's private assessment was more careful, more hedged, and honestly closer to sober American analysis than to national triumphalism.

Sam: So it's not that he said something shameful. It's that he declined to sing the anthem.

Alex: That's it exactly. It's not embarrassing to admit a two-year lag on a twentieth of the compute — by any fair reading it's remarkable. But it punctured a story the moment demands. Which is why the links died within hours, and why this normally unflappable, boring founder was reported to be genuinely rattled.

Sam: So the danger wasn't the content. It was the calibration. He was more accurate than the moment allowed.

Alex: Exactly — and there's corroboration for the tenor of it, by the way. On the record, back in 2024, he'd already said the thing holding China back is chip bans, not money. So the private two-years-behind line isn't a bolt from nowhere. It's the fuller, blunter version of a view he'd hinted at in public.

Sam: There's something almost dark-comic about that. The most subversive thing an AI founder can do in 2026 is decline to exaggerate.

Alex: Decline to exaggerate. In a field where everyone inflates, calibrated honesty is the contraband. And it points us right at the hardest, messiest part of his plan — because if chips are the gap, then the plan to close the gap is where the confident story runs into a much more stubborn reality.

Sam: Right — if the whole problem is chips, and he can't get the American ones, what's the escape hatch? How do you close a compute gap you're banned from buying your way out of?

Alex: The escape hatch runs through Huawei. And this is where the leak's confidence meets an untidier reality. Liang walked investors through DeepSeek's effort to break free of Nvidia's CUDA.

Sam: CUDA. Okay, define that for me, because I hear it constantly and I nod like I know.

Alex: CUDA is Nvidia's software layer — a two-decade-old ecosystem of tools that basically every AI researcher has learned to build on. It's the moat. Think of it like a country whose language everyone in the industry already speaks. The chips are the hardware, but CUDA is the reason it's so painful to leave — you'd have to rewrite everything, in a new language, from scratch.

Sam: So switching off Nvidia isn't just buying a different chip. It's making your whole team learn a new language. And rewrite every book they ever wrote in it.

Alex: That's the real cost, and it's why the moat held for twenty years — not because the chips were magic, but because leaving meant redoing all your work. Any one company defecting is expensive and lonely.

Sam: Unless enough of them defect together that the new language stops being lonely.

Alex: And that's precisely Liang's bet — that the moat is eroding because the whole Chinese ecosystem is being pushed to the new language at once. DeepSeek's move is to rewrite its core code for Huawei's rival framework — it's called CANN — running on Huawei's Ascend accelerators. Liang argues Nvidia's moat is eroding fast, and that China's real problem isn't the software ecosystem so much as production capacity. The chip designs exist; the factories just can't make enough of them yet.

Sam: That's an important distinction, actually. He's not saying "we can't design a good chip." He's saying "we can design it, we just can't manufacture it at volume." That's a factory problem, not a genius problem.

Alex: And it maps right back onto his whole thesis — the gap is physical, not intellectual. Same argument as talent-versus-compute, one level down. The brains can design the silicon; the fabs can't yet print enough of it.

Sam: And he's confident? Timeline?

Alex: He reportedly expects a genuine breakthrough on Chinese silicon within about a year.

Sam: I'm sensing a "but" coming.

Alex: Big but. The reality underneath is messier, and it's the clearest illustration of why calibration matters. Yes — DeepSeek has poured months into porting its stack to Huawei, and Ascend chips increasingly handle its inference, the running part.

Sam: But the harder part — the training, the actually-building-the-model part —

Alex: — has stumbled. Their long-awaited next-generation reasoning model, R2, hasn't shipped. Reporting indicates a training run on Huawei's accelerators failed, and they had to pivot back to Nvidia GPUs for training, while Ascend stays on inference.

Sam: So the honest picture is a hedge. Huawei for serving the model, Nvidia still for building the biggest ones.

Alex: A self-reliance milestone that is genuinely close but not yet reached. And that's the gap the leak exposed in miniature: the confident public roadmap versus the stubborn present. The bet is real, and it may well pay off. It simply hasn't yet.

Sam: And that R2 delay isn't a small thing, right? R2 was supposed to be the follow-up to the model that shook the world. If it keeps slipping because the domestic chips can't train it —

Alex: — then the whole "compute gap is temporary" story is on hold until they fix it. R2 is the test case for the entire self-reliance thesis. If it ships strong and eventually trains on Chinese silicon, he's vindicated. If it keeps stumbling back onto Nvidia, the leak's quiet admission is the real story.

Sam: And a triumphant national narrative just... erases that nuance. "We're self-reliant now" — no asterisk.

Alex: The asterisk is the whole truth. Which is a nice segue, actually, because we keep saying "the Chinese approach" as if it's one thing. And the leak sharpens two distinctions that flatten out from a distance — DeepSeek versus the West, and DeepSeek versus its own neighbors.

Sam: Okay, start with the West, because that contrast feels obvious — but tell me if it's more than obvious.

Alex: It's stark. OpenAI and Anthropic have moved toward closed models behind API walls, aggressive commercialization, and the machinery of eventual public listings — led by founders who are household names. DeepSeek keeps its best models open for anyone to download, refuses to optimize for revenue, and is fronted by a man who's spent years trying to stay in the background.

Sam: So where the American frontier treats the model as a product to be metered —

Alex: — Liang treats it as research to be shared. And he treats cost leadership as a way to build an ecosystem moat rather than a margin. Different verb for the same object. They sell the model. He seeds it.

Sam: Right. But here's where I'd push, because everyone lumps all the Chinese labs together. Is he actually different from the other big Chinese players? Or is "open and cheap" just... the China playbook?

Alex: That's the sharper question, and the answer is he's just as distinct from his domestic rivals as from the Americans. Think of the cohort usually grouped with him — Alibaba's Qwen, ByteDance's Doubao, Moonshot's Kimi, Zhipu's GLM, Baidu's ERNIE.

Sam: The tigers.

Alex: The tigers. And for the most part, those are commercialization plays. Qwen is an infrastructure layer wired into Alibaba's cloud — a model that exists to sell cloud contracts. Doubao is a consumer engine inside ByteDance's apps — a model that exists to keep you scrolling. Baidu's ERNIE, Zhipu's GLM — variations on the theme. And Moonshot has started pricing its newest open model like a premium US flagship rather than undercutting on cost.

Sam: So even when they open-source, there's a business reason underneath. The model is a lever for something else — the cloud, the app, the brand.

Alex: The model is always in service of a business. It's a means. For Liang, the model — the research, the ecosystem — is the end.

Sam: Which is exactly the Kimi K3 story from our episode last week — number thirty-three, if you want the deep dive. Open, but priced like a luxury good.

Alex: Right. So they are open where it helps them sell something. That's the tell. DeepSeek is open because Liang believes the ecosystem itself is the point. He's explicitly rejected the ByteDance-Tencent path of turning a model into a super-app.

Sam: So if you drew it as a grid — open versus closed on one axis, research versus commercialization on the other — the American labs are closed-and-commercial, the tigers are open-but-still-commercial, and DeepSeek is off in its own corner. Open and research-first.

Alex: You just drew the whole map. And that empty corner is the interesting part — because almost nobody else chooses it. Open weights and no commercialization pressure means you're giving away your best work and refusing to monetize it. On paper that's a business that shouldn't be able to exist.

Sam: Except it can exist, because the hedge fund is paying for the corner nobody else can afford to stand in.

Alex: That's the click. The quadrant that looks financially impossible is exactly the one the endowment structure makes possible. Everything routes back to that. DeepSeek is the purist — the one lab in the group that most resembles a research institute that happens to have a hedge fund attached.

Sam: Which brings us right back to the monastery. Okay — but a person like this doesn't appear from nowhere. He's not just an individual. He's a product of a system, and a bet by that system.

Alex: That's exactly the right lens for the last big idea, and it's where the culture question really bites. So place him in his setting. DeepSeek is one of Hangzhou's "six little dragons" — a clutch of hard-tech firms, alongside the robotics maker Unitree and the studio behind a hit domestic game — nurtured by years of Zhejiang provincial policy.

Sam: Six little dragons. That's a great phrase. So the city, the province, is deliberately incubating these companies.

Alex: Deliberately — including a "chain leader" system that wires the party-state into local supply chains. It's industrial policy at the city level: pick promising hard-tech firms, plug them into the supply chain, and grow your own champions. DeepSeek didn't emerge from a vacuum. It emerged from a machine designed to produce exactly this.

Sam: So the flat, curious, bottom-up lab is itself sitting inside a very top-down national plan. Those two things coexisting is kind of the whole China-AI story, isn't it.

Alex: That's the tension in one image. And in February 2025, weeks after R1, Liang sat among the tech elite as Xi Jinping urged China's founders to, quote, show their talent in the contest with America.

Sam: So he's a national champion. Which — oh. That's why the two-years-behind line couldn't stand. He's not a private citizen musing. He's a state-adjacent symbol who's expected to embody self-reliance.

Alex: You just closed the loop the whole episode's been circling. He's politically useful, and an honest word about being two years behind directly undercuts the job he's been assigned. That's the deepest reason the honesty was dangerous.

Sam: But you said there was a deeper cultural signal — in how he builds and talks, not just where he sits.

Alex: There is, and it's the part I find most human. DeepSeek's organization is famously flat. Bottom-up. Researchers — many of them fresh graduates — handed compute and real autonomy, no rigid hierarchy, hiring for passion and curiosity over experience. One of their signature efficiency breakthroughs reportedly started as a young researcher's personal side-interest.

Sam: A side project became a core breakthrough. You don't get that in a place where everyone's terrified of their manager.

Alex: You don't. And it's a deliberate design choice — hire for passion and curiosity over experience, hand a fresh graduate real compute and real autonomy, flatten the hierarchy so a good idea can come from anywhere. It's the org chart as a bet on where innovation actually comes from.

Sam: Which is a very particular bet. That the next breakthrough is more likely to come from a curious twenty-four-year-old with room to run than from a committee of veterans.

Alex: And in interviews going back to 2023, he keeps circling one theme that reads like a mission statement. China, he argues, spent thirty years of the information-technology wave as a follower and a copier. And its real deficit isn't capital — it's confidence. A belief that Chinese firms can do original, hardcore innovation, not just commercialize other people's.

Sam: That's a striking self-diagnosis. Not "we lack money," not "we lack brains" — "we lack nerve."

Alex: And that word — copier — is doing a lot of work for him. His point is that China got extraordinarily good at one half of innovation: taking someone else's breakthrough and commercializing it at scale. What it never built the muscle for was originating the breakthrough in the first place. And that's a habit of mind, not a shortage of resources.

Sam: So he's trying to change a habit, almost. Break the reflex of "wait for the Americans to invent it, then build it cheaper."

Alex: His words: China should gradually become a contributor rather than a free-rider. And when DeepSeek's earlier V2 model startled Silicon Valley, his reaction was pointedly deflating. Among the enormous volume of innovation happening in the US every day, he said, this was an ordinary one. What surprised people was simply that a Chinese company had joined the game as an innovator rather than a copier.

Sam: So the restraint, the open-sourcing, the refusing quick money — it's all in service of proving a point about confidence. It's a psychological project as much as a technical one.

Alex: That's the synthesis. He literally said the issue isn't capital, but a lack of confidence and a lack of knowledge on how to organize top-tier talent for effective innovation. And that's where the culture contrast bites hardest. The Silicon Valley story is individual and urgent — a named founder racing to own a winner-take-all market, monetizing fast to fund the next lap. The founder is the brand, the model is the property, and speed is the whole game.

Sam: And you can feel why that culture produces what it produces — that urgency is a real engine, not a flaw. It's just a very different machine from the one he's describing.

Alex: A completely different machine. And his is collective and patient. Hire for curiosity, hand young people autonomy, give the work away, and measure success in whether China becomes a place that originates technology.

Sam: So the yardstick itself is different. Silicon Valley measures "did we win the market." He's measuring "did we change what China is capable of." Those aren't even the same game.

Alex: They're not, and that's the deepest cut. Where one culture treats the model as property and the founder as a prophet, the other treats the model as a contribution and the founder as a custodian of a longer national project. It's telling that he insists he's not a genius — that DeepSeek is "a group of ordinary people." That register is almost unthinkable from an American founder selling a vision to investors.

Sam: And I want to be careful not to romanticize it, because "custodian of a national project" is also exactly why the state can lean on him. The collective, patient, give-it-away model isn't purely noble — it's also very useful to a government that wants a champion.

Alex: That's the right complication to hold. The same openness that reads as generosity from one angle reads as national strategy from another. He can be sincere about the confidence mission and be a useful instrument of the state at the same time. Both things are true, and the leak sits right on the seam between them.

Sam: And that, more than any benchmark, is what the leak accidentally laid bare. Not a secret weapon. A worldview.

Alex: A worldview. Which is exactly what we should carry into the last question — where does all of this actually go?

Sam: Okay, forecasting. And I want you to be honest about the uncertainty, not just pick a winner.

Alex: Deal. Forecasting from here means naming the hinges — the specific things that, if they swing one way or the other, change the answer. For DeepSeek, the near-term hinge is R2 and the domestic-silicon bet.

Sam: The model that hasn't shipped, on the chips that failed the training run.

Alex: Right. If R2 ships strong, and eventually trains on Chinese chips at scale, then Liang's confident roadmap is vindicated and the compute gap starts to look temporary. If R2 keeps slipping and frontier training stays chained to Nvidia — then the leak's quiet admission becomes the defining constraint. And export controls will have done more than Beijing admits.

Sam: Because restraint buys you focus. But restraint cannot manufacture a chip.

Alex: Restraint buys focus; it cannot manufacture chips. That's the sentence. Now zoom out to the whole race, because the recording reframes the contest itself.

Sam: How so?

Alex: If Liang is right that the gap is compute rather than talent, then the American lead is real — but rented. It's a function of chip access that policy created, and policy could erode. Not a permanent edge in genius.

Sam: Rented, not owned. That's a very d…