Compute is now geopolitical

Artificial intelligence runs on compute — the chips, memory, power and buildings that turn electricity into model training — and Nvidia’s Jensen Huang said this week China already has enough of it to stay in the race. He said on an April 15 podcast that China has “enormous” compute capacity, including underused “ghost data centers,” even under U.S. export controls. (dwarkesh.com, tech.yahoo.com) Huang’s remarks landed as Taiwan Semiconductor Manufacturing Co., the main contract chipmaker for Nvidia, reported first-quarter revenue of $35.9 billion on April 16, above its own guidance range of $34.6 billion to $35.8 billion. TSMC posted a 66.2% gross margin and guided second-quarter revenue to $39.0 billion to $40.2 billion. (investor.tsmc.com, investor.tsmc.com) TSMC’s first-quarter net income reached NT$572.48 billion, up from NT$361.56 billion a year earlier, and earnings per share rose to NT$22.08 from NT$13.94. Its investor site said the company reported first-quarter results on April 16, 2026, after March revenue had already pointed to stronger demand. (investor.tsmc.com, pr.tsmc.com) Compute has become a supply-chain contest as much as a software contest. Huang told Dwarkesh Patel that Nvidia’s edge depends on access to logic dies from TSMC, high-bandwidth memory from SK Hynix, Micron and Samsung, and rack assembly in Taiwan — a chain that spans several countries before a model is trained. (dwarkesh.com) That makes datacenters part factory, part power plant. The Financial Times reported that Microsoft, Alphabet, Amazon and Meta planned to lift capital spending above $300 billion in 2025, while Gartner estimated $475 billion in datacenter spending for the year, driven by denser AI sites that use far more electricity than older cloud facilities. (ft.com) China’s side of that buildout is uneven, not imaginary. MIT Technology Review reported in March 2025 that China had built hundreds of AI datacenters during the boom and that many later sat underused, while the Financial Times described major projects in places including Xinjiang and Inner Mongolia as local governments tried to expand AI infrastructure. (technologyreview.com, ft.com) The U.S. has tried to slow that expansion with export controls on advanced computing chips and semiconductor tools. The Bureau of Industry and Security says it imposed major controls on the People’s Republic of China in October 2022 and October 2023, and Nvidia has repeatedly had to redesign or delay products for China as those rules changed. (bis.doc.gov, cnbc.com) Even with those limits, Nvidia said in March it had received purchase orders from China and was restarting manufacturing for H200 shipments after getting clearance from both governments. CNBC reported that China had once accounted for at least one-fifth of Nvidia’s datacenter revenue, showing why access to that market still shapes the company’s public argument. (cnbc.com, nvidianews.nvidia.com) TSMC’s numbers point to the same pressure from the other end: demand is still outrunning easy supply. Bloomberg reported on April 16 that TSMC raised its 2026 revenue outlook to growth of more than 30% and signaled capital spending toward the upper end of its existing range, as AI customers kept ordering despite wider geopolitical risk. (bloomberg.com) The result is that “compute” now means more than faster chips. It means who can secure wafers, memory, packaging, power, land, export licenses and enough datacenter shells — occupied or not — to turn those parts into national AI capacity. (dwarkesh.com, investor.tsmc.com, ft.com)