Chip War author Chris Miller: China's $295B AI plan is $59B/year — less than one hyperscaler — and Beijing isn't AGI-pilled

China's $295B AI plan and the semiconductor supply chain

Chris Miller, economic historian and Chip War author, makes a point that reframes the headline figure immediately: China's $295 billion AI plan runs over five years, making it roughly $59 billion per year — less than what a single US hyperscaler spends annually. The number is large in absolute terms but underwhelming as a statement of strategic urgency.

Beijing isn't AGI-pilled

The more pointed observation is that China's revealed preferences suggest its government simply doesn't believe AI will be as transformative as Silicon Valley does. If Beijing were convinced, Miller argues, it would be buying H200 chips and aggressively building data centers. It isn't doing either. Chinese firms that want access to Western compute are routing through offshore locations like Malaysia instead. Meanwhile, the government is actively pushing domestic companies toward Huawei's Ascend chips, which Miller reads as a combination of supply-chain nationalism and genuine concern about backdoors in foreign hardware — not a sign of a country racing toward AGI.

The contrast with the US is structural as much as cultural. America's service-sector-heavy economy means AI delivers returns faster there than in manufacturing-heavy China. US cloud adoption dwarfs China's even relative to GDP — Chinese large enterprises have been slow to move to the cloud — and there's no shortage of venture capitalists and tech founders embedded in US government circles pushing AI urgency. China has none of that feedback loop.

“Those numbers are over five years. So annually, that's less money than Google or any of the hyperscalers... The puzzle is why isn't Xi Jinping more AGI pilled? If you thought AGI AI was important, you thought chips were an important ingredient, China's been underspending for the last four years on AI... I think the Chinese government just doesn't really believe that AI is gonna be nearly as important as we do.”

The semiconductor supply chain ceiling

Semiconductor spending as a share of global GDP was roughly flat for two decades, then doubled in four years, entirely driven by AI demand. Whether it can triple is the open question, and Miller says the constraint right now is manufacturing capacity, not demand. TSMC and ASML sit on the wrong side of the bullwhip: the deeper you are in the supply chain, the more skepticism accumulates about whether AI demand is real, and the more conservative your CapEx stance becomes. TSMC has historically been rewarded for that conservatism, investing at the bottom of cycles, but it creates genuine tension with how fast AI companies want capacity to scale.

Intel looks stronger than it did, with customer traction building, though volumes haven't materialized yet. Miller flags TeraFab (Elon Musk's fab venture) as the other variable to watch — transformative if it works, but a hard industry to break into. Samsung's Texas plant is also coming online, with Tesla reportedly among its customers.

On NVIDIA's moat, Miller is measured. ASICs will take some share, but NVIDIA acquiring Groq technology six months ago signals it can extend into inference-focused hardware. CUDA remains a durable advantage. Market share may slip below 90%, but NVIDIA remaining the dominant GPU supplier for the foreseeable future looks well on track.

Export controls and the Manus signal

Miller's practical advice on US export controls: ignore the daily noise and watch Congress, where bipartisan support for controls has been consistent. He also points to DeepSeek's Manus situation — Chinese authorities pulling passports from leading AI entrepreneurs — as a potentially chilling signal for China's own ecosystem. Liang Wenfeng, DeepSeek's founder, holds roughly 80% ownership of the company through his hedge fund High Flyer, making him one of the more exposed figures if Beijing decides to tighten its grip on AI entrepreneurs.

Russia barely registers. Hundreds of thousands of Russians, disproportionately educated and tech-sector workers, left after 2022. Nebius and other Yandex spinouts relocated to the West and are building there. GPU smuggling into Russia exists but is small-scale, with no real demand base for large data centers behind it.