Miller: China Underspending on AI, Not Buying Nvidia H200s

Chris Miller, author of “Chip War,” said Tuesday on TBPN that China has been “underspending for the last four years on AI.” The comments challenge the prevailing narrative of a Beijing-led AI arms race and carry direct implications for Nvidia (NVDA) and Taiwan Semiconductor Manufacturing (TSM).

Reports have circulated that China is preparing a roughly $295 billion nationwide AI investment plan. Miller pushed back, noting the figure is spread over five years, making annual outlays lower than the capital expenditures of leading U.S. cloud and AI companies. “If you thought AI was important and chips were an important ingredient, China’s been underspending,” he said. He added: “The puzzle is why isn’t Xi Jinping more AGI-pilled?”

Beijing is steering firms toward Huawei’s domestic chip ecosystem, citing potential data security concerns and fears of backdoors in foreign hardware. That shift could reduce long-term demand for American GPUs. Miller also said Chinese firms have still not purchased Nvidia’s H200s, even though the Trump administration permitted their sale for roughly six months.

China’s chip exports hit a record $31 billion in April, roughly doubling year over year. Miller’s analysis suggests much of that volume is in lower-complexity chips, not the advanced GPUs required for AI training. The semiconductor industry remains constrained by manufacturing capacity at TSM and by bottlenecks in advanced packaging and related supply-chain segments. Miller described the dynamic as a “bullwhip effect,” where demand fluctuations amplify further down the supply chain.

Semiconductor spending as a share of GDP has nearly doubled over the past four years, driven entirely by AI demand, Miller said. The constraint is not demand but manufacturing. For Nvidia, the key open question is whether Chinese buyers eventually order H200s. The stock carries an Alpha Score of 69 out of 100; TSM scores 68. Both face tail risks if China’s AI buildout remains slower than the U.S. hyperscaler cycle.