Nvidia flags China compute, TSMC upbeat
Nvidia’s Jensen Huang warned that China has ‘enormous’ compute capacity and so‑called 'ghost' data centres that could rival U.S. AI infrastructure. (crypto.news) Taiwan Semiconductor gave Q2 guidance above expectations and said it sees a multi‑year AI growth cycle, underscoring the foundry's central role in the hardware stack. (techradar.com) (benzinga.com)
Nvidia and Taiwan Semiconductor are delivering the same message from opposite ends of the AI supply chain: China’s compute base is large, and chip demand is still climbing. (dwarkesh.com) (tsmc.com) On April 15, Jensen Huang told Dwarkesh Patel that the debate over selling AI chips to China sits alongside a bigger question about how much computing power China already has. The interview was published on April 15 on Patel’s site and circulated widely on April 17 through transcript and news writeups. (dwarkesh.com) (techflowpost.com) A day later, on April 16, Taiwan Semiconductor Manufacturing Co. reported first-quarter revenue of $35.9 billion and guided second-quarter revenue to $39.0 billion to $40.2 billion. The company also said 2026 revenue in U.S. dollar terms should grow by more than 30%, up from its earlier forecast of close to 30%. (tsmc.com) (reuters.com) AI compute is the industrial side of artificial intelligence: the chips, packaging, power and factory output needed to train and run models. Nvidia designs many of the processors, but TSMC manufactures the most advanced versions and assembles the production roadmap around nodes like 3 nanometer. (dwarkesh.com) (reuters.com) That is why Huang’s warning and TSMC’s guidance fit together. If China already has large pools of usable compute, export controls may shape which chips get sold, but they do not erase demand for data centers, networking gear, memory and foundry capacity. (dwarkesh.com) (cnbc.com) TSMC’s numbers show where that demand is landing now. First-quarter net income rose 58% to NT$572.48 billion, and Chief Executive C.C. Wei said AI-related demand remained “extremely robust” while the company planned capital spending at the high end of its $52 billion to $56 billion range. (cnbc.com) (reuters.com) The bottleneck is not only chip design. TSMC said production capacity remains tight, expanded 3-nanometer capacity across Taiwan, the United States and Japan, and tied part of that buildout to its broader $165 billion Arizona investment plan. (reuters.com) (tsmc.com) There is also a policy fight running through the story. Nvidia has already said new U.S. licensing rules for H20 exports to China could force a $5.5 billion charge, and Huang has argued that blocking sales can push Chinese buyers toward domestic alternatives rather than stop Chinese AI development. (cnbc.com) (dwarkesh.com) TSMC, by contrast, is speaking from the factory floor rather than the policy podium. Its April 16 results showed advanced 3-nanometer chips accounted for 25% of sales, up from 6% in the third quarter of 2023, a sign that the newest manufacturing processes are moving deeper into the AI buildout. (reuters.com) The next test is whether those orders keep turning into shipped systems through the rest of 2026. For now, Huang is warning that compute is spreading faster than Washington can fence it off, and TSMC is saying the fabs are still full. (dwarkesh.com) (reuters.com)
