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

Alvin Graylin, a U.S. citizen with over 35 years operating inside both Chinese and American technology ecosystems, argues the U.S.-China AI competition is structurally misframed. He describes it not as a prisoner's dilemma but as a stag hunt, a coordination game with two stable Nash equilibria, where treating the competition as a race forces irrational decisions and produces suboptimal outcomes for both sides. He contends the assumption that whoever reaches AGI first rules the world forever is unsupported by data and that, unlike the space race, there is no clear finish line.

China is not behaving as though ASI is imminent. Every AI model released in China must pass review by the Cyberspace Administration of China, delaying releases by weeks or months. Chinese regulators told domestic labs not to purchase H-200 chips that Americans were offering. Only two or three Chinese labs are meaningfully AGI-focused, and not to the level of Silicon Valley counterparts. China's stated strategy is embedding good-enough AI into industry and exporting those industries globally through the Belt and Road initiative, which currently covers approximately 150 partner countries. Graylin characterizes the CCP's primary AI motivation as expanding global economic influence rather than political domination.

DeepSeek was not government-funded and was not directed to open source its models. The Chinese government initially reprimanded DeepSeek for releasing a capable model publicly, but the global PR value prompted Xi Jinping to publicly endorse open source AI strategy at the WAIC, after which most new Chinese companies oriented toward open source. Graylin puts total Qwen downloads at 1 billion across over 100,000 model variants. Open sourcing allows global hyperscalers to host Chinese models worldwide without Chinese labs bearing the capital expenditure burden that U.S. labs carry. The gap between open source and closed source model performance has narrowed from roughly 18 months to two or three months.

U.S. chip export controls functioned as evolutionary pressure rather than a substantive barrier. Chinese labs conduct training runs in international data centers in Hong Kong, Taiwan, and Europe where Blackwell-generation chips are accessible, then transport results back to China on disk, with a trained model file of approximately three terabytes transmissible in roughly one hour. The embargo saved Chinese GPU companies including Moore Threads, Cambricon, and Biren from bankruptcy by forcing domestic demand, and Graylin predicts Chinese GPU companies will reach parity with American capabilities and begin exporting within two to three years. Chinese open weight models climbed from 2 percent to 61 percent of open router traffic in the last two years.

Graylin characterizes the distillation attack framing as largely a lobbying construct invented by Anthropic. Anthropic's own evidence flagged roughly 20,000 accounts across three Chinese labs generating one to two million queries at a total estimated cost of only two to three million dollars. By contrast, Meta spends between 100 and 200 million dollars per month on Anthropic tokens and has only recently produced a somewhat competitive model, which Graylin says undermines the claim that distillation produces rapid capability transfer. If a billion-dollar frontier model can be replicated for two to ten million dollars through distillation, the business model of frontier AI development is structurally broken.

On valuations, Graylin puts major hyperscalers' off-balance-sheet debt at approximately 1.7 trillion dollars, contrasting that with the roughly 200 million dollars of off-balance-sheet debt that characterized the Enron era. He states 45 percent of U.S. stock market value is currently in the AI sector versus approximately 30 percent for internet companies at the dot-com peak, and the Buffett indicator sits at 240 percent of GDP versus 120 percent at the internet bubble peak. Anthropic's revenue growth has essentially flattened at an annualized run rate in the seventy billion dollar range despite commitments to hundreds of billions in capital expenditure. More than half of Amazon and OpenAI AI revenues come from just two companies each, making the revenue base highly undiversified.

Graylin warns that the U.S. racing toward AGI risks commoditizing financial services, creative services, and consulting, the sectors where America currently holds global preeminence, while the U.S. workforce at roughly 70 percent white collar is far more exposed to AGI-driven displacement than China at 40 percent. He proposes an AI Marshall Plan deploying data centers and technology globally to build markets and alliances, and calls for a red line hotline between the U.S. and China to prevent false flag misattribution incidents. He identifies the shared interest in preventing non-state actors from weaponizing AI as the most viable foundation for cooperation, a view he says was echoed by Treasury Secretary Bessent following the recent dialogue.