An analytical breakdown of Moonshots episode #281: Alvin Wang Graylin on why the US is misreading the game, how the chip embargo created China's GPU industry, and why the world is playing a stag hunt, not a prisoner's dilemma.
This conversation's value lies not in its forecasts but in deciphering how the Chinese side actually thinks. Alvin Wang Graylin, a US citizen born on a "re-education farm" during the Cultural Revolution, a dual MIT master's graduate, former HTC China president, digital fellow at Stanford HAI, and AI safety advisor to the upcoming US-China dialogue. His brother commands a US Navy nuclear submarine. A man who has physically sat on both sides of the table offers a counterargument to the entire American "race for ASI" paradigm.
The US is playing a prisoner's dilemma (zero-sum, "defect against the other"), when the actual game is a stag hunt (positive-sum, "hunt the stag together"). The world is not zero-sum, ASI has no finish line, and under superintelligence nation-states will lose their value. America has imposed on itself the wrong game with the wrong strategy.
Graylin is a rare figure at the intersection of two elites. Born on a "re-education farm" during the Cultural Revolution, he moved to the US in 1980 under his uncle's sponsorship. His grandmother was a New York Tribune reporter during the Second Sino-Japanese War and left his mother in Shanghai after the bombing of Pearl Harbor. Mixed heritage (Ashkenazi + Scottish + Chinese), dual MIT master's degrees in computer science and business, and an engineering education from the University of Washington.
Professionally: senior roles at HTC, Intel, IBM, Trend Micro, Watchguard, across all five layers of the AI stack (data centers, chips, PCs, phones, XR, applications). Founded 4 startups. Today, digital fellow at Stanford HAI, senior fellow at the Asia Society Policy Institute, professor of AI policy at the University of Washington, and advisor to the US government on the US-China AI safety dialogue.
Graylin's brother is a senior officer on a US Navy nuclear submarine (a "boomer" capable of destroying countries), and three of his four children are active-duty Navy officers. This is not a "China lobbyist", this is a man whose family literally sits on both sides of the strategic divide.
Graylin lays out three "heresies" against the Washington paradigm of the race for ASI:
The key takeaway: the US is playing a prisoner's dilemma (zero-sum), while the real game is a stag hunt (positive-sum). This is the core of the entire conversation, and Graylin returns to it repeatedly.
The central fact of the episode. Chinese open models grew from 2% to 61% of OpenRouter traffic in two years. Alibaba's Qwen : 1 billion downloads and 180,000 derivative models.
Graylin reveals the psychology behind DeepSeek: the company's CEO open-sourced the model against the government's advice. Xi Jinping initially asked "why give away such an excellent model for free?", and the labs got their hands slapped. But then soft power and the "local champion" PR effect flipped the view: DeepSeek became global news, its valuation soared, and now open source is blessed as national strategy.
Four to five years of embargoes on chips, EDA software, and lithography forced China to innovate under resource scarcity. By opening a model, a lab with 100–200 researchers gains 100,000+ external contributors. For inference this means hyperscalers and neoclouds worldwide host Chinese models, lifting from China the capex burden that American labs carry.
The harshest passage is first-hand testimony. Immediately after the ban on high-end chip sales, every one of Graylin's semiconductor-industry contacts got a call from the Chinese government: "Need funding? Resources? How can we accelerate? We'll give you customers."
The result: Chinese GPU companies that were 2–3 generations behind and that no one bought are suddenly unable to keep up with production. American policy literally created the Chinese GPU industry (Moore Threads, Cambricon, Biren, and others), "they would have gone bankrupt without American sanctions." Within 2–3 years they will begin exporting their chips.
The second blow: training has already moved offshore. Chinese labs train on Blackwell chips in international data centers and bring the weights back on a disk (~3 TB, an hour of transfer). The ASML embargo hits fabs (which can't be moved), but training is just "work" that relocates instantly to Hong Kong, Taiwan, or Europe. "The file moves around the world like a liquid," Diamandis summarizes.
You can embargo ASML machines, the fabs are physically on site and can't be moved overnight. But training is a job that relocates to a server in Hong Kong, Taiwan, or Europe instantly. A three-terabyte file transfers in an hour. The chip embargo has "completely failed."
Anthropic is the only company lobbying against "distillation attacks" (a term that, per Graylin, they coined themselves). His breakdown of the numbers kills the argument: the alleged "20,000 accounts from three labs, one to two million requests" amounts to $2–3 million in total traffic, and for DeepSeek, "thousands of dollars."
Graylin's counterargument: if a $1B model can be copied for $2M, the economics of frontier AI don't work. Meta spends $100–200M/month on Anthropic tokens and only recently shipped a competitive model, despite having the highest payroll and the most compute in the world. Why didn't a Kimi-level model come out of them six months ago?
Graylin recalls how Claude, asked "who are you" in Chinese, answered "I am Qwen", at Alibaba they laughed: "they're distilling from us, we just can't trace it." xAI, via Cursor (which post-trained on Claude traces), does exactly what the Chinese are accused of, just with two layers of deniability.
A set of numbers that together paint a picture of fragile debt:
| Metric | Dot-com peak (~2000) | AI era (2026) |
|---|---|---|
| Tech sector share of the stock market | ~30% | 45% |
| Buffett Indicator (market/GDP) | ~120% | 240% |
| Off-balance-sheet hyperscaler debt | — | $1.7T |
| Comparison: Enron debt | $200M | ×8,500 larger |
Nvidia announced a $500B deal with BlackRock, Carlyle, and Blackstone to securitize chips and compute capacity. "This sounds like subprime," Graylin says.
The historical parallel Graylin draws directly: the US today is the USSR in 1988. The USSR lost the Cold War not to missiles but to bankruptcy, 15–20% of GDP went to armaments with no real value. The US spends ten times more on data centers and gets "97% of the quality." Hundreds of billions of dollars are poured into infrastructure that, by his assessment, will not generate profit for the investors themselves, the value will flow to open-weight models on cheap global hardware.
More than half of Amazon's and OpenAI's AI revenue comes from just two companies. "That's not a very diversified revenue base." When commoditization arrives, companies with hundreds of billions in debt will be left with "stranded assets," while Chinese models run on global neoclouds at the price of electricity.
The key game-theoretic model. Prisoner's dilemma, zero-sum: both defect to suffer less. Stag hunt (Rousseau), positive-sum: two hunters can together catch a stag (a month of food) or, individually, hares (a few days).
The worst case is playing out right now: America is chasing the stag (AGI/ASI, "solve everything"), China is chasing hares (good-enough AI for the economy). The lone stag hunter ends up with nothing, if the US fails to reach AGI or can't control it, it gets zero. China keeps collecting small wins and survives.
Graylin's optimum: both slow down → both build good-enough → the technology safely diffuses into the economy → then they hunt the stag together. But the US has imposed a prisoner's dilemma on itself and is playing by someone else's rules.
While the US builds data centers, China builds energy. China generates 10 times more new renewable energy per year than the US, solar and wind farms in the west, ultra-high-voltage lines with <1% losses per thousand miles. Electricity cost: 2–3 cents/kWh in China versus 10–30 cents in the US.
| Energy metric | US | China |
|---|---|---|
| New renewable generation/year | ~1× | ~10× more |
| Electricity cost | $0.10–0.30 /kWh | $0.02–0.03 /kWh |
| Strategy | Data centers first, energy later | Energy first, data centers follow |
| EV penetration | ~10% | ~50% of new sales |
Chinese generation at 2–3 cents/kWh, delivered from western solar farms over ultra-high-voltage lines with <1% losses per thousand miles, means Chinese inference will always be cheaper. The US builds data centers first and thinks about energy second, a structural error.
Graylin's recent article (Cipher Brief, for national security) argues: model size does not correlate with real danger.
| Threat domain | Model size | Danger | Status |
|---|---|---|---|
| Chemical weapons | 10–50M parameters | High | Open-source, already available |
| Bio/gene engineering | 1–50B parameters | High | Open-source, already available |
| Cyberattacks | 1–100B parameters | High | Microsoft Mdash (100 small models) — 95 vs Mythos 83 |
| Frontier models | $1B+ training | Controllable | Cloud, telemetry, harness |
"The danger is already here", in bio/chemistry especially. A small model runs on a laptop in a basement, and these models are already openly available. The real question is not "how to control big models" but "how to control the precursors and synthesis machines."
The attacker needs a small model (one hole is enough); the defender needs a big one (all holes must be scanned). Big models are in fact more controllable: they sit in the cloud, with telemetry and harness constraints. The harness matters more than the model, it is what "neutralized" Mythos.
The gap in views is enormous: campuses say "5–20 years," Silicon Valley says "1–2 years, and shrinking every year." Graylin looks at the behavior of the Chinese government and concludes: they are not acting as if ASI is around the corner.
If they believed it, they wouldn't ban labs from buying H200s, wouldn't impose regulations (CAC, censorship, week- to month-long delays for every model), wouldn't slow themselves down. China's leadership sees AI as "just another technology", a general purpose technology that diffuses into society over years, if not decades. Only 2–3 labs in China are "AGI-obsessed," and even then idealistically.
Beijing sees AI as a tool for diffusing into the economy; Washington sees it as a race for ASI. This divergence colors all policy: one regulates and deploys, the other accelerates and spends. Graylin: "if the CCP believed ASI was around the corner, it would not be restraining its own labs."
A recent tour of Unitree's headquarters and factories gives Graylin a sober view of the humanoid hysteria. 150+ companies showcased robots at the World AI Conference, but the global annual production volume is in the tens of thousands, 80–90% from 2–3 Chinese companies. "There is no market for this many companies right now," he says bluntly. Within 1–2 years, 150 will compress to a handful.
The critique of the form factor is radical: the humanoid is a terrible form factor. Legs are a liability (balancing, falling, maintenance). Unitree is moving toward "torso-only" models, and they sell better. Figure demoed 10-hour package sorting, but it's theater: a single-arm machine would do the same for pennies.
American robotics companies now smuggle Chinese components in suitcases, just as the Chinese once smuggled GPUs. 90% of robot components come from China. "The world is interconnected; we can't stuff the genie back into the bottle."
The most underestimated factor. In China 80% of the population is pro-AI; in the US, 80% is against. The reasons: 40 years of visible improvement in living standards (800 million lifted out of poverty), state media amplifying successes, and, critically, legal precedents: Chinese courts have ruled that you cannot fire a person by replacing them with AI without providing another job. This removes the fear that paralyzes American adoption.
| China | US | |
|---|---|---|
| Sentiment toward AI | 80% pro-AI | 80% anti-AI |
| Youth unemployment | ~20% (42% graduate underemployment) | ~9% |
| White-collar workers | ~40% | ~70% (more vulnerable to AI) |
| Successful AI deployments | Higher (top-down, less inertia) | 6% (McKinsey) |
McKinsey: only 6% of corporate AI deployments in the US actually work. The technology is already good enough, the bottleneck is organizational resistance. China's top-down model and worker protections accelerate adaptation.
In the US, 70% of the workforce is white-collar (the first victims of AGI); in China, 40%; in Africa, 10–20%. The US races toward an AGI that will "replace precisely the sectors where America is strong", finance, consulting, creative.
Graylin debunks the myth of "a monolithic CCP issuing orders from the top." The reality is a distributed, competitive system: the center sets direction ("the next 5 years, clean energy, robots, AI"), while 30+ provinces compete among themselves for "provincial champions." No one orders specific technologies, the labs compete fiercely.
Diamandis's key observation: in the US, politicians are lawyers; in China, they are engineers (80–90% of top leadership is technical). This explains the difference in industrial policy: China understands technology, the US understands law.
China's "AI Plus" plan: the goal is 70% of companies integrating AI within 5 years, 90%+ within 10. The focus is on deployment and diffusion into manufacturing, education, and medicine. The American AI Action Plan is supply-side: better models, better chips, "dominate", and nothing about what comes after.
The difference in philosophy: China asks "how to deploy safely and grow the economy," the US asks "how to be first." Meanwhile China is already executing the "Eastern Data, Western Computing" megaproject (compute in the west, where energy is cheap; data in the east, where the megacities are).
Graylin introduces a distinction: the West confuses China's anxiety (the century of humiliation, the desire never to repeat it) with ambition (the desire to dominate). China is not trying to "conquer the world", it is trying to become strong and independent so that it is never humiliated again. This changes the entire frame.
After WWII the US had 50% of world manufacturing; today 15%; China has 35% and is growing toward 40–45%. America is strong in finance, creative, and consulting, precisely the sectors AGI will devalue first.
The prescription for the West: reindustrialization. Hard manufacturing will remain necessary even under AGI. But he adds a nuance: the long-term redistribution of labor is not back into factories but into human-centered services (teachers, nurses, elder care). This is the sector that will absorb 60–70% of displaced workers.
Consulting, accounting, and law firms are already saying "we no longer need junior staff, a few seniors plus AI will do more, faster." If you're a mid-level McKinsey employee, it's time to worry.
Chinese state media is an "always good news network": inventions, new buildings, successes, instead of disasters and murders. Graylin doesn't defend it, but he explains the effect: the population sees more good than bad, and is therefore optimistic about the future. Even in video games the blood is green (violence censorship).
Diamandis's counterpoint: every night American news "trains the neural network" of 100 billion neurons on fear. Hence the gap in public sentiment, which matters more than any benchmark.
But Graylin doesn't hide China's dark side either: youth unemployment ~20%, actual graduate underemployment at 42%, the "lying flat" (躺平) phenomenon, the gig economy. He is honest: China has no solution; what's needed is international coordination of the labor transition.
A bombshell anecdote from a breakfast with TSMC's CTO and a former senior CIA official. The CTO's question: "Is it true you'd blow up our data centers if China attacked?" The CIA's answer: "I can neither confirm nor deny. But we have a thousand of your engineers, whom we'll evacuate before anything happens."
The meaning: the US has already planned the extraction of TSMC's human capital, the fabs themselves are not the prize. America is focused on duplicating manufacturing capability at home (the CHIPS Act).
Graylin: the reason for China's interest in Taiwan is not the fabs but political history (the unfinished civil war of 1949). The fabs are useless: they require materials, chemicals, and maintenance from around the world, after a seizure they'd quickly run out. He sees no interest or urgency in a near-term attack for chips.
The thesis from his substack "Great Reckoning before the Reconnecting": within the next 2 years the private-credit bubble bursts (the $1.7T in off-balance-sheet debt financing the data-center buildout). When frontier labs' financials go public (IPO), the market will realize: the value of AI innovation does not accrue to the companies building the infrastructure. Companies with hundreds of billions in debt and revenue concentrated in 2–3 clients will face a reckoning.
Historical precedent: in 2008 China saved the US by buying trillions in T-bills and holding rates down. Graylin's war game: the bubble bursts → a debt-burdened Washington calls Xi → asks for financial help → quid pro quo: the US de-escalates on Taiwan, accepting peaceful reunification. Not "here's Taiwan," but "we'll stop selling weapons and interfering."
Within the next 2 years the $1.7T bubble bursts → a debt-burdened US asks China for financial help (as in 2008, when T-bill purchases averted the crisis) → in return the US quietly de-escalates on Taiwan, enabling negotiations for peaceful reunification.
Graylin is an advisor to the American side at the US-China AI safety dialogue (September 24, Washington). His key premise: success is not a framework agreement but simply a "second discussion." China prepares for diplomacy for months; the American style of "negotiating a day before" strikes the Chinese as chaotic and unprofessional. Kissinger arrived months before Nixon, that is the model.
The dialogue's priority: not state-versus-state (that has been in balance for 80 years, and AI doesn't change it, no one wants a first strike because of escalation) but non-state actors, extortionists, cyberattacks (up 200–300% in a year), a thousand agents attacking the network. A common enemy could render the "national race" an obsolete concept.
If you translate the conversation into your coordinates (Shenzhen, WFOE, ARLI, robotics):
| ♾ What to Do | ☠ What to Avoid |
|---|---|
| Sell components, don't compete in hardware. 90% of robot components come from China. American companies are already smuggling Chinese actuators. Demand is desperate. | Don't build a humanoid company. 150 players are chasing a commercial market that doesn't exist. Within 1–2 years only a handful will remain. |
| Go where the US can't. The FCC has closed off the US — 190 countries remain open. The Middle East, SEA, LatAm, Africa. | Don't depend on US market access. It's closing, not opening. The chip embargo proves the trend. |
| Energy asymmetry is your moat. Chinese inference at 2–3 cents/kWh is always cheaper. ARLI on Chinese models is structurally cheaper than Western competitors. | Don't ignore energy cost. The long-term moat is in the price of inference, not model quality. |
| The bubble creates your entry point. When the $1.7T deflates, AI becomes a game of cost efficiency — Chinese components win on price. Build the supply chain now. | Don't time the market top. Demand will come after the correction, not at the peak. |
If Graylin's Great Reckoning is right, the next 2 years will bring a financial shock. This doesn't invalidate the strategy, but it means cash flow and low capex (trading, not manufacturing) are the right defense.
Deep Analysis · ARLI · August 2026 · Source: Moonshots #281 with Alvin Wang Graylin (YouTube: Slle5_AxBzs) · Analytical synthesis, in our own words · Gold accent for geopolitical/trade analysis
ARLI membership: the notes from the workshop
One site of yours, measured, with the price for the work in the same document
ARLI, Inc., a Delaware corporation. Questions: hq@arliwork.com.