Nvidia designs chips. TSMC builds them. That distinction sounds like a footnote. It is actually the most important fact in the AI race.
Jensen Huang made this plain in a Dwarkesh Patel podcast interview last week, though he did not frame it quite this way. TSMC manufactures what Nvidia architects. Nvidia controls the relationships with the fabrication partners, the memory suppliers, the packaging facilities, and the developer ecosystem that makes its hardware worth buying. TSMC is the world's most critical fabrication facility, located in a geopolitically contested territory, and Nvidia is its most important customer. The supply chain that AI runs on is not a network. It is a small number of very specific facilities, with very specific relationships, controlled by one company in California.
Three pressures are testing whether that control holds.
The first is custom silicon. Three weeks before the recording, Anthropic signed a multi-gigawatt deal with Broadcom and Google to run TPUs at a scale with no industry precedent. OpenAI is building its own accelerator, called Titan. Both companies are substantial Nvidia GPU customers. Both are systematically reducing that dependency.
Jensen's response in the interview: Anthropic is unique, not a trend. Without Anthropic, there would be no TPU growth at all. It is 100 percent Anthropic.
That answer worked before OpenAI disclosed Titan. If the labs accounting for a significant share of Nvidia's GPU purchases are independently designing their own chips, "unique" is doing considerable work in that sentence.
The second pressure is the China market and the export controls threatening to eliminate it. Chinese companies ordered more than 2 million H200 chips for 2026, according to TechPolicy.Press. Jensen has estimated the China market at $50 billion annually. Congress is moving to close it. The AI Overwatch Act would impose mandatory denial on chips more powerful than the H200, including Nvidia's Blackwell architecture. The MATCH Act would ban exports of advanced semiconductor manufacturing equipment to China entirely. Nvidia has export licenses and purchase orders in hand, and restarted H200 manufacturing at GTC, Tom's Hardware confirmed. But the market could be legislated away before the orders are fulfilled.
The third pressure is upstream fab capacity. SemiAnalysis reported in the podcast that AI will consume 60 percent of TSMC's N3 node this year and 86 percent of the N2 node next year. Nvidia is TSMC's biggest customer on N3 and one of the biggest on N2. Jensen called it a two to three year problem, solvable with enough capital expenditure and EUV machine production. His strongest counterargument is that energy infrastructure is the binding constraint, and that takes far longer to build than software iteration. He is correct. He also spent years persuading upstream CEOs to invest in capacity, which means the constraint he is currently managing is partly of his own making.
Jensen's strongest argument for why the three pressures do not erode his position is the CUDA ecosystem. CUDA is Nvidia's proprietary computing platform, and rewriting CUDA code for custom silicon requires years of engineering work. The total cost of ownership advantage of staying on Nvidia remains significant for most customers. The GPU computing market is still expanding, which means the economic pie is growing even if custom silicon takes a larger slice.
The counterarguments are real. On fixed pricing: Jensen said Nvidia never changes prices based on demand. H100 chips sold for four to five times their list price during the 2023 shortage. On the China ecosystem lock-in argument: that assumes China stays dependent on the US stack. Huawei and SMIC are not standing still. On custom silicon: the question is whether hyperscalers, the customer type that matters most to Nvidia's revenue, continue to need Blackwell at current order volumes if they have their own silicon programs.
If OpenAI's Titan succeeds, the labs accounting for a substantial share of GPU purchases will have made a credible choice to reduce Nvidia dependency. If the AI Overwatch Act and MATCH Act pass, the China market vanishes before purchase orders are fulfilled. If TSMC's N2 node reaches 86 percent AI utilization next year, energy infrastructure becomes the binding constraint on AI scaling regardless of how much capital Jensen commits upstream.
Three pressures. The supply chain that AI runs on is a small number of very specific facilities, controlled by one company. What happens to that control when all three arrive at once.
What to watch: OpenAI Titan's public disclosure, the Congressional markup schedule for the AI Overwatch Act, and TSMC's next earnings call, where the N2 node utilization figures will either confirm or undercut Jensen's two to three year timeline.
