2026-08-18

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

Peter Diamandis and the Moonshots co-hosts sit down with Alvin Wang Graylin, a US-China AI insider who just returned from the World AI Conference in Shanghai, to unpack why China does not behave like it believes ASI is imminent while Silicon Valley labs act like it is days away. Graylin argues the US-China AI race is wrongly framed as a prisoner's dilemma when it should be a stag hunt, that US export controls on chips and Fable-class models backfired by accelerating Chinese open-source innovation (Qwen, GLM, Kimi K3) and forcing China to build its own chip industry, and that AI training has quietly moved offshore because weights move like liquid while fabs cannot. He warns that 45% of US stock market value sitting in AI, $1.6-1.7 trillion in hyperscaler off-balance-sheet debt, and Anthropic's flattening revenue point to a serious correction, and sketches a scenario where a popped AI credit bubble forces the US to ask China for financial help in exchange for backing off Taiwan. The conversation also covers humanoid robots' poor commercial fit, why smaller AI models may pose bigger bio/cyber risks than frontier ones, China's WACO initiative versus the US's PAX Sinica, and Graylin's pitch for an 'AI Marshall Plan' ahead of the September 24 US-China AI safety dialogue in DC.

▶ Watch on YouTube

Topics

China's real ASI timeline vs Silicon Valley's AI ▶ 0:23
Graylin argues Chinese labs and regulators are not behaving as if ASI is near (still restricting H200 purchases, keeping slow-moving AI regulations), unlike AGI-pilled Silicon Valley labs.
Chinese open-weight surge and export-control backfire AI ▶ 12:33
Chinese open models went from 2% to 61% of OpenRouter traffic in two years; Qwen has ~1 billion downloads and 180,000 derivatives; US chip and model export controls pushed China toward open source rather than slowing it.
Stag hunt vs prisoner's dilemma framing Geopolitics ▶ 1:44:35
Graylin's game-theory framework: the US is racing for a winner-take-all AGI 'stag' alone while China plays it safe with 'good enough' AI, creating the worst-case payoff structure instead of a cooperative equilibrium.
Training compute quietly moved offshore Compute ▶ 56:37
Because China lacks Blackwell-class chips domestically, frontier training reportedly happens in international data centers and the resulting weights are shipped back on disk, making the chip embargo more optics than substance.
Distillation dispute between Anthropic and Chinese labs AI ▶ 47:45
Graylin disputes Anthropic's 'distillation attack' framing, arguing the actual dollar cost was a few million versus a claimed billion-dollar frontier build, and that all labs (including US ones) distill from each other.
Humanoid robot overhype vs commercial reality Robotics ▶ 1:05:45
Despite 150-200+ Chinese humanoid robot companies, real commercial deployment is minimal; legged humanoids are impractical, and Unitree is shifting toward wheeled/upper-torso designs for commercial sales.
WACO vs PAX Sinica: competing global AI blocs Geopolitics ▶ 28:45
China's World AI Cooperation Organization (announced at WAIC, 29 countries) is framed as an open, shareable-AI alternative to the US-led PAX Sinica (~25 countries) alliance model.
Smaller models, bigger real-world danger AI ▶ 1:24:35
Graylin's new paper argues no correlation exists between model size and real-world risk; small (10M-50B parameter) models already enable bio/chemical weapon design and run on a laptop, an under-addressed threat.
The coming AI financial correction Economy ▶ 1:54:47
45% of US stock market value sits in AI, the Buffett indicator is at 240% of GDP (vs 120% at the dot-com peak), and $1.6-1.7T in hyperscaler off-balance-sheet debt signal a fragile, overleveraged AI buildout.
China's AI Plus plan vs the American AI Action Plan Economy ▶ 1:10:26
China's plan targets AI diffusion into industry (70% adoption in 5 years, 90%+ in 10), while the US plan focuses on supply-side dominance in models and chips with no adoption target.
Labor displacement: white-collar workers vs manufacturing Economy ▶ 1:19:33
70% of the US workforce is white-collar and most exposed to AGI displacement versus 40% in China; Graylin argues future labor absorption will be human-to-human services (nursing, elder care, teaching), not manufacturing.
Taiwan, TSMC, and the chip-embargo endgame Geopolitics ▶ 2:03:35
Graylin argues China's interest in Taiwan is civilizational/political, not about seizing TSMC's fabs (which are unsustainable without global supply chains), and floats a scenario where a US financial crisis leads to a Taiwan-related quid pro quo with China.

Predictions made

open Alvin Wang Graylin: A private-credit bubble financing the US AI data center buildout will pop, prompting the US (possibly Trump) to ask China for financial help, in exchange for the US backing off its involvement in Taiwan.
EP #? · · due: within 2 years (~2028) · ▶ watch
“maybe at that point uh the Americans uh or maybe Trump will will give a call to she and say hey can you help us out again”
Your call:
open Alvin Wang Graylin: Chinese GPU makers (referenced as Moore Threads, Cambricon, Biren) will catch up to US chip technology and begin exporting their chips internationally.
EP #? · · due: within 2-3 years (~2028-2029) · ▶ watch
“within the next you know two or three years they will start to catch up to to what what America is doing and they will start to export their chips”
Your call:
open Alvin Wang Graylin: The 150-200+ Chinese humanoid robot companies demoing products will consolidate down to a single-digit number of survivors.
EP #? · · due: within 1-2 years (~2027-2028) · ▶ watch
“in the next year or two uh you'll see that 150 go down to you know probably you know a single-digit number of surviving um uh human and robot companies”
Your call:
open Alvin Wang Graylin: China's share of global manufacturing capability will rise from roughly 35% today to 40-45%, while the US share (~15%) stagnates.
EP #? · · due: 5-10 years (~2031-2036) · ▶ watch
“I think the forecast is it's going to go to 40 or 45% over the next 5 or 10 years”
Your call:
open Alvin Wang Graylin: The current AI-driven US stock market concentration (45% of market value, Buffett indicator at 240% of GDP) will trigger an economic correction.
EP #? · · due: unspecified · ▶ watch
“that to me is a sign that we are in a very very fragile place and a economic correction is due”
Your call:
open Dave Blundin: Kimi K3, when run as thousands of parallel instances, is capable of recursively improving itself, and Dave states he will run 5,000 instances to demonstrate this.
EP #? · · due: unspecified (near-term) · ▶ watch
“I have 5,000 Kimies running tomorrow. 5,000. I guarantee you it can improve itself.”
Your call:
open Alvin Wang Graylin (citing China's AI Plus plan): 70% of Chinese companies will have integrated AI into their operations within 5 years, rising to over 90% within 10 years.
EP #? · · due: 5 years / 10 years from plan announcement (~2029 / ~2034) · ▶ watch
“within the next 5 years we want to have 70% of companies integrate AI into their their business... within the next 10 years we want to have 90 plus%”
Your call:

Numbers that matter

Worth digging into

🕳️ The AI-credit-bubble-to-Taiwan-quid-pro-quo war game
A specific, checkable geopolitical-financial forecast tying Nvidia's $500B securitization deal and hyperscaler debt directly to US Taiwan policy leverage -- if it plays out, it reshapes the entire cross-strait calculus.
🕳️ Dario Amodei's alleged 'there will be Anthropic and then governments' remark
Relayed secondhand via a Sacks/Baker podcast, this suggests a monopolistic single-company ASI endgame belief inside a leading lab -- worth confirming against primary sources before treating as established.
🕳️ Graylin's 'Biggest AI Models Are Not the Biggest Threats' paper and the UK AISI cyber-capability report
Reframes AI safety priorities from frontier-lab governance toward controlling small (10M-50B parameter) bio/chem-capable models and synthesis-machine precursors, a policy angle largely absent from mainstream AI safety debate.
🕳️ Offshore training as an export-control loophole
The claim that Chinese labs train on foreign Blackwell-equipped data centers and ship weights back on disk would mean the entire chip-embargo strategy is largely symbolic -- a major policy-relevance finding if verified.
🕳️ The Anthropic vs. Chinese-labs distillation dispute
Competing claims (Anthropic's 'distillation attack' framing vs. Graylin's $2-3M cost estimate and the anecdote of Claude answering 'I'm Qwen') shape US policy on restricting Chinese lab access to US model APIs.