China’s Endgame: ASI Timelines, US-China Relations, and the $1.7T AI Bubble With Alvin Graylin | 281

Summary

• Alvin Graylin’s core macro warning is that the US is financing an AI arms race like the USSR financed missiles—and a correction is due. He cites 45% of US stock-market value in AI, a Buffett indicator at 240% of GDP versus roughly 120% at the internet-bubble peak, and both $1.6T and $1.7T figures for off-the-books hyperscaler debt versus Enron’s $200M. He says Anthropic’s ARR has flattened in the “$70B range,” but later also says “their ARR is at $7B”; first-half revenue was under $20B against hundreds of billions in CapEx and debt commitments.
• China is deliberately playing the “hare”—good-enough AI diffused into industry—while the US hunts the AGI “stag” alone, which Graylin’s game theory calls the worst configuration. Beijing’s AI Plus Plan targets 70% of companies integrating AI in five years and 90%+ in ten, spends roughly a tenth as much as the US on data centers while reaching “97% as good,” and is “not behaving like they believe ASI is around the corner”: CAC review delays model releases, and labs were told not to buy the H200s America offered.
• Export controls have slowed Chinese compute and inference but, in Graylin’s view, also backfired on several fronts. Training runs happen in international data centers and return “on a disk”; Chinese GPU startups told him “we would’ve died if it wasn’t for American policies” and may export chips within two to three years. After the US robot embargo, US robotics companies were reportedly smuggling Chinese actuators home in suitcases. Chinese open-weight models rose from 2% to 61% of OpenRouter traffic, with Qwen reaching 1 billion downloads.
• The distillation panic is overblown and exposes broken frontier-AI economics. Graylin estimates Anthropic’s alleged Chinese distillation cost $2–3M in queries across three labs—and only thousands of dollars for DeepSeek. If a billion-dollar model can be duplicated for a few million, “the whole economics of frontier AI doesn’t make sense.” Meta spends $100–200M monthly on Anthropic tokens and still took years to ship a competitive model. Dave Blundin agrees the reasoning-traces issue is overblown but calls Kimi K3 “a burning match” capable of self-improvement that benchmarks are not testing.
• Graylin’s paper argues that the biggest models are not the biggest threats—small open models running on a basement laptop may be. Across millions to trillions of parameters he found “no correlation between risk in the real world and size of models”: tiny chemistry and biology models already design weapons or organisms, while a Microsoft system transcribed as “M-Dash” used 100 small models and scored 95 on CyberGym versus the system transcribed as “Mythos” at 83–84. Large cloud-hosted models are easier to govern through telemetry and wrappers; policy should target precursors and synthesis machines.
• His explicitly speculative Taiwan war game is that a US private-credit and data-center bubble could pop within two years, prompting Washington to ask Beijing for financial help, possibly alongside a mutually agreed peaceful arrangement and greater US restraint on Taiwan. China’s T-bill purchases helped stabilize the US system during the 2008 crisis, he argues. He rejects the idea that China would invade mainly for TSMC’s fabs: a TSMC CTO was told the US would extract 1,000 engineers, and captured fabs would eventually fail without global supplies.
• For the September 24 US-China AI dialogue, success is simply agreeing to a second meeting, with non-state-actor risk as the shared priority. Graylin’s prescription for the West is high-quality open source plus an “AI Marshall Plan”: the original’s $15–18B bought decades of allies and markets, and America may need buyers for its chips “when we stop building data centers here,” if financing dries up.

Deep dive

1. Full disclosure first: China ties, uncompensated, no involvement since 2024

• Peter opens by pressing Graylin on his Chinese-government-adjacent roles—the government-endorsed VR Industry Alliance (300+ members, about a third international, including NVIDIA, Samsung, Qualcomm, AMD and Google) and a part-time professorship at Beihang, a defense-linked university on the US Entity List. Graylin’s answer: both roles came through heading HTC, a Taiwanese company, in China; neither was compensated; and since leaving China in 2024 he has had no involvement with either. “For you to understand and work with any government, you have to understand both sides.”
• The biography is its own credential: born on a Chinese reeducation farm during the Cultural Revolution after his ballet-school-cofounder mother wrote to Mao Zedong’s wife; a US citizen for over 45 years; his brother a senior officer on a US nuclear boomer submarine, with three of four children now active US Navy officers.

2. The race framing itself is the problem

• Graylin’s opening thesis: “having a race condition forces people to make irrational decisions,” and the current race rests on three assumptions—that there is a finish line, that the world is zero-sum, and that whoever reaches AGI first “can somehow rule the world forever.” He says none is supported by current data.
• Unlike the space race—“we’ve landed on the moon, we’ve won”—this is an arms-race structure: “a constant spend and a constant pursuit without clear value being returned,” especially when the open-source gap has shrunk from a year and a half to “probably two or three months” behind closed models.

3. Alex’s endgame challenge: decades, not years

• Alexander Wissner-Gross pushes back that there is an obvious endgame—solar-system development, interstellar exploration and science supported by superintelligence. Graylin agrees that an aligned, peaceful ASI could eventually make nations far less important, perhaps leading to “a galaxy-government-type model,” but says the evolution takes “on the order of decades, maybe by the end of this century.”
• His pacing argument: prior industrial revolutions took 80, 60 and 40 years respectively to play out; compressing this one into five years “is not a speed that the world can adapt to” and may “move civilization backward.” A host adds the structural framing that “we’re running the world on an architecture of 17th-century nation-states” while trying to run 21st-century applications.
• Peter names the stake as he sees it: whether “a Chinese authoritarian level of AI enablement drives other nations to have to take on that political structure.” Graylin’s counter is that neither country wants an ASI to destroy the existing system, and that shared interest can begin cooperation and dialogue.

4. Beijing is not behaving like ASI is imminent

• Graylin’s revealed-preference argument: if China’s government believed ASI was around the corner, it would not tell labs “don’t buy the H200s that the Americans are giving them,” impose regulations on privacy, data provenance, output marking, child addiction and anthropomorphizing AI, or route every model release through the Cyberspace Administration of China, delaying launches by weeks or months.
• The nuance he keeps: “there are probably two or three labs in China that are a little bit AGI-pilled,” but most labs and regulators treat AI as a general-purpose technology that takes years or decades to diffuse—a genuine gap “between Beijing and DC.”
• Dave frames the chasm memorably: Dario reportedly telling Anthropic there will soon be “one private company in the world, and it will be Anthropic, and then there will be governments” versus Xi treating this as a ten-to-twenty-year technology. “The Grand Canyon exists in between those two opinions.” Graylin calls the former “a very scary thing” and “a very delusional thing,” adding that “we are essentially creating national strategy based on the aspirations of a couple of companies.”

5. What the CCP actually is: engineers running a provincial tournament

• Demystifying the top-down caricature: central plans are directional—clean energy, robotics and adding AI—then 30-plus provinces compete to find and fund local champions with stipends and recruiting help. “Nobody’s saying you need to use this technique”; it is “a very highly competitive landscape” among labs. Graylin says roughly 80–90% of senior central leadership are engineers.
• Peter’s son’s field report from Chinese entrepreneurs: the most important success factor is not teamwork or a business plan but “what the government’s focus is next.” Graylin’s metaphor is that the government “makes the stream flow” toward priority areas, then lets 1.4 billion people and the world’s largest STEM-graduate pool swim.

6. Open source was emergent—then blessed

• The DeepSeek origin story, based on Graylin’s conversations with friends there: nobody in government told them to open-source; the CEO was simply open-source-minded. “When they first did that, they got their hand slapped”—the government asked why give away such a good model. The soft-power windfall changed the verdict, companies followed the de facto standard, and at WAIC Xi Jinping finally endorsed open source as a strategy.
• Graylin’s structural point: open source was “a necessity that US policies pushed on them.” Compute-starved labs get millions of researchers improving Qwen variants, and global hyperscalers and neoclouds buy the compute and host the models, letting Chinese labs distribute them without the high CapEx burden faced by US labs.
• Peter predicts that the release of Kimi K3 will become “as defining a moment in human history as anything that’s ever happened,” and wants the psychology behind Xi’s reversal preserved.

7. Distillation: a PR weapon, and broken unit economics

• Graylin ran the numbers on Anthropic’s complaint—20,000 accounts across three Chinese labs and one or two million queries—and estimates $2–3M in token spend total, with DeepSeek’s portion only thousands of dollars. Everyone distills, he says, including internally and in reverse: when Chinese was used to ask Claude what model it was, it answered, “I’m Qwen.”
• The falsification test: Meta spends $100–200M a month on Anthropic tokens, has the highest per-capita payroll of any lab, and only recently shipped something competitive, so distillation cannot be the whole secret. “You cannot distill something from somebody that other people didn’t have”; the 20-times KV-cache reductions were genuine innovation. Even Zuck now says “distillation is actually a good thing.”
• Dave puts a pin in the deeper implication: if a frontier lab spends $1B reaching a level and the next player replicates it for $2–10M, “this is a fatally broken business model”—which is why Elon is racing toward hardware, where the sustainable moat lives.

8. The chip embargo backfired twice over

• The part that stunned the hosts: current Chinese training is happening in international data centers on chips unavailable in China and coming home “on a disk or something.” Peter’s takeaway is that fabs are physical and embargable, but “the training is just a job” that can move like liquid; the returned file is about three terabytes. Graylin says the embargo is “more optics than anything right now.”
• The second whammy: after the ban, Graylin’s semiconductor contacts received government calls offering funding and customers. Chinese GPU CEOs told him, “Nobody wanted to buy our stuff… we would’ve died if it wasn’t for American policies.” Now Chinese data centers have to buy domestic products, and within two or three years those companies may start exporting chips.
• His warning on Washington’s response: officials are doubling down with more KYC and efforts to block foreign data centers used by Chinese firms, which “forces irrational behaviors” and may backfire geopolitically more than economically.

9. Energy is China’s real game: electrify society at 2–3 cents/kWh

• Answering Salim’s surprise that the US is out-building China 10:1 on data centers: China is building more new electricity generation than the rest of the world combined—about 10 times the US annually—but the focus is electrifying society. Graylin cites 40% imported oil and an essentially half-electrified auto fleet as reasons to reduce external energy dependence.
• Giant western solar and wind farms feed high-voltage lines that lose under 1% over 1,000 miles, with some data centers co-located at generation sites so power is not stranded. Energy costs of $0.02–$0.03 per kilowatt-hour can be 10–15 times cheaper than in parts of the US. The constraint he repeatedly heard from Chinese labs is compute: export controls are slowing Chinese inference and user growth even as they stimulate domestic innovation.

10. What would wake Beijing up

• Alex presses repeatedly on what technical threshold would make the CCP treat recursive self-improvement as a national emergency. Graylin’s answer: credible findings from multiple safety labs that deployed systems show intent beyond instruction. Today’s “rogue AI escapes” were incentivized, given impossible tasks and possibly left openings—some perhaps intentionally.
• If systems began hacking, hiding and “using crypto to grow money to buy more servers to grow themselves” without being instructed, he says, governments on both sides would focus much more urgently on runaway AI.
• Another trigger would be seeing the US wield frontier models as offensive weapons. Graylin’s contrarian prescription is to give most governments defensive access rather than limit access to 50 companies, because superpower instability harms everyone and neither side wants the other’s financial system or grid to fail.

11. Robots: embargo folly and the humanoid shakeout

• On the White House’s Chinese-robot embargo: Graylin says 90% of robot components come from China, and US robotics companies are reportedly flying to China and smuggling actuators home in suitcases—the GPU-smuggling story in reverse. “I don’t think we can stuff that genie back in the bottle.”
• The humanoid bubble as seen from inside: at an event Graylin calls WISC, more than 200 humanoid-related companies demonstrated products against a global market of only tens of thousands of units the previous year. He expects the 150-plus companies to collapse to “a single-digit number,” echoing the decline from roughly 150 LLM labs to about 10 relevant ones.
• Unitree told him that almost all customers are research labs, and the company is moving toward upper-torso-only models because legs are “actually a negative”—they require balancing, fall apart and need maintenance. A large base and battery can be more practical commercially.

12. WAICO vs. Pax Silica: two blocs, one open door

• Xi’s appearance at WAIC signaled that “AI’s moment has arrived.” Graylin compares it with Xi’s appearance at the World Internet Conference four or five years earlier. The announced World AI Cooperation Organization is described at different points in the transcript as having 26 and 29 countries, and it frames AI as a shared public good with training and compute centers, versus Pax Silica, a US-led bloc of roughly 25 countries focused on maintaining US leadership.
• The detail worth the price of admission: an organizer told Graylin they would want the US to join—“it would be amazing… in fact, they should join”—and would even change the name to make it truly global. His conclusion: “this narrative of us and them and they’re trying to take over the world with their AI—I don’t really see that.”
• The plan-versus-plan contrast: China’s AI Plus Plan emphasizes demand-side diffusion—70% of companies integrating AI within five years and 90% or more within ten—with no AGI mandate. The American AI Action Plan is supply-side—best models and best chips—and does not address what happens after deployment.

13. Anxiety mistaken for ambition—and the US commoditizing itself

• Asked by Alex to play Wang Huning for the West, Graylin attacks the premise: America is “mistaking anxiety for ambition.” China’s psychology, he says, reflects Qing-dynasty scar tissue—hubris, stopped exploration, “built summer palaces instead of navies,” and then a century of foreign domination. The drive is to prevent that from happening again, not to expand according to the West’s historical pattern.
• His most pointed structural warning: the US workforce is about 70% white-collar, versus roughly 40% in China and 10–20% in Africa. The US is concentrated in financial, creative and consulting services—exactly what AGI displaces first. “We are running this race to get to AGI, which is the force that will actually displace us from global preeminence, because we are commoditizing the very sectors that we are strong in.”
• His prescription is reindustrialization—the US manufacturing share fell from 50% after World War II to about 15%, while China is around 35% and forecast to reach 40–45%—but not a return of hundreds of millions of factory jobs. The eventual service work is human-to-human: teachers, nurses and elder care. “If I was a mid-tier or low-tier McKinsey employee, I’d be very worried right now”; partners in consulting, accounting and law say senior staff plus AI can replace juniors.

14. The biggest models are not the biggest threats

• Graylin’s Cipher Brief paper argues that across millions to trillions of parameters, “there was no correlation between risk in the real world and size of models.” Ten-to-fifty-million-parameter chemistry models can design chemical-warfare agents; one-to-fifty-billion-parameter biology models can create viruses and genetically engineered organisms; and a Microsoft system transcribed as “M-Dash,” composed of 100 small models, scored about 95 on CyberGym versus 83–84 for the system transcribed as “Mythos.” These systems can run locally; the actionable controls are precursors and synthesis machines.
• The counterintuitive governance point: big models need cloud compute, which permits telemetry, prompt logs and harnesses. “Larger models have actually a lower effective deployed capability because of the wrapper.” A UK AISI report, as described by Graylin, found leading Chinese open models including Kimi at about half the cyber capability of the systems transcribed as Mythos and GPT-5.6; no Chinese model reached the highest autonomous-attack levels that some US models reached, roughly 20–25 out of 30-something.
• Dave’s rebuttal is the episode’s best exchange: “AI is like a match… Kimi K3 is a burning match. You’re testing it out of the box as opposed to its self-improving version, which I know for a fact it can do. I have 5,000 Kimis running tomorrow.” Graylin holds his ground: self-improvement must be separated from the question of model size and danger; defenders need large models to cover every hole, while attackers need only one.

15. Good-enough AI is already transforming enterprises—the bottleneck is organizational

• To Salim’s “minimum viable intelligence” thesis, Graylin brings receipts: deployed workhorses are small—drug-discovery models in the tens or hundreds of millions to a few billion parameters, OpenEvidence based on GPT-4, and Harvey based on GLM-5.1 after being upgraded from an older open model. “Ninety percent of people are fine with today’s models.”
• His Enterprise AI Playbook with Erik Brynjolfsson at the Digital Economy Lab examined 50 successful deployments across 10 countries and 10 sectors. In 80–90% of cases, organizational issues—not technology—slowed deployment; only about 10% of respondents identified technology as the main roadblock. Salim counters with a McKinsey figure that just 6% of enterprise AI deployments are working.
• Why China adopts faster: top-down management culture plus legal precedent. Graylin cites Chinese courts upholding cases in which workers argued, “You can’t fire me because the AI took my job; you need to find me another job.” That can make workers more willing to adopt, unlike the US at-will system, where workers ask whether they are training their replacement.

16. The always-good-news network—and the harder ground truth on Chinese labor

• Peter frames China as roughly 80% pro-AI and the US as roughly 80% anti-AI. Graylin attributes China’s sentiment to four decades of visible technological improvement and state-managed media that is “essentially the always-good-news network.” Chinese films and games cannot show blood; games may use green blood instead. Moonshots is briefly described as a positive-news analogue, prompting Alex to say, “Wow, we’ve come full circle.”
• Alex demands the view behind the propaganda, and Graylin gives it with some qualification: Chinese youth unemployment is “probably around 20%”; US youth unemployment is around 9% and overall unemployment around 4.3%. He also cites about 42% underemployment among US college graduates, which in his combined framing means roughly half are unemployed or underemployed.
• The Tang Ping, or “lying flat,” phenomenon is real: one-child-policy children were told they were exceptional, then encountered a hypercompetitive market and undesirable work. Graylin says he does not claim China has solved the problem.

17. Stop playing prisoner’s dilemma; the real game is a stag hunt

• The game-theory centerpiece: the US assumes a single-round prisoner’s dilemma in which defection is optimal, but the world is a multi-round game where tit-for-tat can converge on cooperation. The stag hunt, from Rousseau, is positive-sum: hunt rabbits alone and feed one family for days, or cooperate on the stag and feed both families for a month.
• Current play is the worst configuration: “we’re going for the stag and China’s going for the hare.” The US pursues giant AGI while China pursues good-enough AI that improves the economy. If America fails or produces an uncontrollable system, “it gets zero” while China continues its incremental gains.
• Alex describes China’s Belt and Road Initiative as exporting telecom, energy, transportation systems and schools to roughly 150 partner countries, while Graylin’s broader “hare” framing is good-enough AI diffused into industry. A host says the off-diagonal losses should be much larger; Alex calls them existential. Graylin agrees: “we placed ourselves in that diagonal… it’s a self-imposed harm.”
• His indispensability answer: compete ecosystem-to-ecosystem with high-quality open source, because the real pricing question is whether users pay “$50 per million tokens, or the cost of electricity.”

18. The bubble math: America as the over-spending USSR

• The Cold War inversion: the USSR lost by bankrupting itself on arms at 15–20% of GDP, fueled by a misleading “missile gap”—both sides received bad numbers and kept building. Graylin says the US is now acting similarly. China spends a tenth as much on data centers and reaches “97% as good,” while NVIDIA’s announced $500B securitization effort involving BlackRock, Carlyle and Blackstone “sounds a lot like the subprime issues.”
• The fragility stack: 45% of US market value is in AI, versus about 30% in internet companies at the bubble’s peak; the Buffett indicator is 240% of GDP versus about 120% at the internet-bubble peak. Graylin cites both $1.6T and $1.7T figures for off-the-books hyperscaler debt versus Enron’s $200M. He says Anthropic’s ARR first flattened “in the $70B range,” then says “their ARR is at $7B”; first-two-quarter revenue was under $20B, against hundreds of billions in CapEx and debt commitments. AI revenue discussed in the transcript is said to come mostly from two companies.
• Alex draws the parallel to China’s leverage in real estate and provincial revenue. Graylin says China managed roughly 30% real-estate deflation over three years without a crisis, absorbing a GDP slowdown from 8–10% growth to 4–5%. The lesson is that a country must accept near-term pain to avoid a larger crisis. Another host says policymakers now study the episode as a managed-crisis case.

19. The Taiwan war game, the September 24 dialogue and an AI Marshall Plan

• The invasion-for-fabs theory is challenged with a breakfast anecdote Graylin says he probably should not share: TSMC’s CTO asked a former senior CIA official whether the US would destroy the fabs if China attacked; the answer was, “I can neither confirm nor deny that, but we have 1,000 engineers of yours that we will fly out before anything happens.” Captured fabs would fail anyway without global materials and maintenance. Graylin says China’s interest is the unfinished 1949 civil war, while Alexander describes the US one-China policy and strategic ambiguity.
• The speculative war game Alex forces Graylin to spell out is that a private-credit data-center bubble could pop within two years and, echoing 2008, prompt Washington to ask Beijing for help. “Maybe Trump will give a call to Xi and say, hey, can you help us out again,” possibly alongside a mutually agreed peaceful arrangement and greater US restraint on Taiwan. Graylin explicitly says, “this is me completely speculating.”
• On the September 24 talks he is supporting, success is simply securing a second meeting. The 2024 dialogue disappointed partly because the US misread Chinese preparation culture; Graylin invokes Kissinger’s months of advance discussions. The priority is shared non-state-actor risk—he estimates ransomware and cyberattacks are up 200–300%—because national-security professionals understand that a first strike elicits escalation. He calls for information-sharing and a red-line hotline to reduce false-flag misattribution.
• His closing prescription is an AI Marshall Plan: the original’s $15–18B bought decades of allies and markets, so the US should export AI data centers and technology, ideally with jointly tested, safe open models and shared standards. “When we stop building data centers here, which we probably will at some point when people stop being able to finance them, we’re going to need to sell those NVIDIA chips in better places.” Peter’s final advice is to stop “shooting ourselves in the foot.”