The world’s most important chipmaker has run out of polite ways to say no. TSMC, the Taiwanese foundry that manufactures the processors behind nearly every AI system in production, said publicly this week that it simply cannot make enough chips to satisfy demand, and that using its new Arizona factory to close the gap could take a very long time.
“We can only support so much,” a senior executive said during an investor event, in a rare admission from a company that usually lets its order book speak for itself. The remarks, reported by multiple outlets covering the event, were the closest TSMC has come to declaring that the AI boom has outrun the physical capacity of the global chip supply chain.
The bottleneck is straightforward. Every leading AI accelerator from Nvidia and AMD is built on TSMC’s most advanced processes, the same ones used for Apple’s newest iPhone processors and Qualcomm’s mobile chips. Those lines are effectively full. TSMC has been expanding for three consecutive years, breaking ground on new fabs in Taiwan, Japan, Arizona and Germany, but a modern chip factory takes two to three years from first shovel to first wafer, and no amount of money compresses that schedule.
The reference to Arizona carries its own weight. TSMC’s U.S. plant, a $65 billion project, is the company’s largest overseas bet, built at the urging of Washington to secure advanced chip supply on American soil. But the executive’s caution that meeting customer needs from Arizona could take a very long time suggests the fab will not relieve the global crunch soon. Ramping yields on a new process in a new location, with a new workforce, is a slow and unpredictable process, engineers who have worked on such transfers say.
For Nvidia and AMD, the message is unwelcome but unsurprising. Both companies sell every chip they can get, and both have told investors that supply, not demand, is the binding constraint on their revenue. Nvidia’s chief executive has publicly lamented the gap between what customers want to order and what TSMC can produce. The foundry’s statement effectively confirms that the shortage will persist through the next product cycles.
The response, already visible, is a wave of price increases. Wall Street analysts now expect wafer pricing to enter a new rising cycle, with the most advanced nodes seeing the steepest increases. Chipmakers will pass those costs to their customers, which means every AI company that rents or buys accelerators faces a higher hardware bill next year. For startups operating on venture funding, the increase is existential; for the largest cloud providers, it is a line item measured in billions.
The pricing power marks a structural shift. For most of the industry’s history, foundry prices fell year after year as processes matured, the same way consumer electronics prices decline. That logic has broken. With demand for AI chips growing faster than capacity, TSMC can charge more for scarcity, and customers have no alternative supplier for the most advanced nodes. Samsung and Intel both offer competing processes, but neither has matched TSMC’s yields on the cutting edge, and switching is not a matter of a single order.
The consequences ripple through AI economics. Model training and inference are compute-intensive, and hardware is the largest single cost for frontier AI companies. If chip prices rise while model sizes keep growing, the cost per query climbs, squeezing gross margins across the industry. Cloud providers will try to absorb the increase through efficiency gains, but analysts say a meaningful portion will land on end customers, either through higher API prices or higher subscription fees.
The shortage is not only about silicon. The other bottleneck is advanced packaging, the process of stitching together the chiplets that make up a modern AI accelerator, and TSMC controls most of the world’s capacity there as well. Customers have told investors that packaging, not just wafer production, determines how many chips they can ship, and TSMC has been expanding that capacity as fast as it can build clean rooms. The packaging constraint is less visible than the wafer shortage, but it is just as binding, and it is one reason the industry’s own forecasts keep slipping.
Customers are responding with long-term commitments. Cloud providers and AI companies have signed agreements that guarantee future capacity at fixed prices, in some cases paying for capacity years before it exists, according to people familiar with the contracts. The prepayments give TSMC the cash to fund its expansion, but they also lock customers into prices that may look expensive if the market softens. For now, the risk is one-sided: the shortage is real, and the customers with the deepest pockets are the ones securing the most capacity.
The capacity question is also geopolitical. Washington has pushed for more chip production on American soil, and TSMC’s Arizona plant is the centerpiece of that policy, but the physics of ramping a new fab do not move for politics. The executive’s caution about Arizona was a plain-spoken way of saying that a factory in the desert does not become a supply chain overnight, and that the gap between policy ambition and industrial reality is measured in years, not quarters.
Some relief is coming, but slowly. TSMC’s Arizona plant is expected to contribute volume in the coming years, and the company is also expanding advanced packaging, the other scarce resource in AI computing. Yet executives at the foundry and its customers alike caution that even with all announced capacity, supply will trail demand for at least the next two years. The world is discovering that AI runs on physics as much as on software, and physics, unlike a model release, cannot be accelerated.
