TSMC has informed Nvidia and Broadcom that it cannot provide as much production capacity as they want, The Information reported this week. The disclosure arrives just days after TSMC posted record quarterly earnings and announced capital expenditure plans that blew past analyst expectations.
The timing is awkward. TSMC just committed to spending between $52 billion and $56 billion this year, a 37% increase over 2025, with most of it earmarked for leading-edge manufacturing. And yet its two biggest AI chip customers were apparently told: still not enough.
"Not enough, not enough, still not enough"
TSMC CEO C.C. Wei made this explicit back in November at the Semiconductor Industry Association awards, where he estimated that the company's existing advanced-node capacity is still roughly three times short of what its major customers plan to consume. He joked about wearing a T-shirt reading "No more wafers."
Coming from Wei, who rarely gives off-the-cuff remarks, the comment landed hard. TSMC produces roughly 90% of the world's most advanced chips. Every Nvidia GPU, every AMD processor, every Apple M-series chip, and most AI accelerators come from TSMC fabs. When Wei says capacity is insufficient, there's nowhere else to go.
The bottleneck isn't just raw wafer production. TSMC's CoWoS advanced packaging technology, which stacks high-bandwidth memory alongside processors on a silicon interposer, has become the real chokepoint. In mid-2024, TSMC revealed that its advanced packaging capacity for 2024 and 2025 was fully booked by just two clients: Nvidia and AMD.
The Ripple Effects
Google has already felt the squeeze. Alphabet reportedly reduced its 2026 production target for Tensor Processing Units from around 4 million to 3 million units due to limited CoWoS access. Nvidia has locked in over half of TSMC's CoWoS capacity through 2026 and 2027, leaving competitors scrambling for what remains.
Some customers have started looking elsewhere. Tesla signed a $16.5 billion deal with Samsung last summer to produce its next-generation AI6 chip at Samsung's troubled Taylor, Texas facility. Elon Musk called the deal's strategic importance "hard to overstate." Translation: TSMC couldn't give Tesla what it needed either.
That Samsung deal says something about the current moment. Samsung's foundry business has struggled to attract major clients and holds just 8% of the global market versus TSMC's 67%. Yet Tesla took the bet, which tells you how constrained TSMC's capacity really is.
The Money Won't Solve It Fast
TSMC's $56 billion spending plan sounds massive. It is massive. But new fabs take years to build and ramp. Of the projected $52-56 billion capital budget, approximately 70-80% will be allocated to leading-edge process technologies, with 10-20% going to advanced packaging.
The Arizona expansion that's supposed to help diversify capacity won't move the needle soon. TSMC is accelerating its timeline there, but high-volume 2nm production in Arizona won't happen until the second half of this year at earliest. Even then, packaging remains an issue. American customers still rely on production capacity in Taiwan for the packaging process. Chips manufactured at TSMC's US wafer fab need to be airlifted to Taiwan for packaging.
A CoWoS packaging facility in Arizona is planned, but it won't be operational until 2029.
Who wins when supply can't meet demand?
The obvious answer is TSMC. The company just posted its eighth consecutive quarter of profit growth, with net income hitting a record NT$505.7 billion. Gross margins are above 62%. When you're the only game in town for the chips everyone needs, pricing power follows.
The less obvious answer is Intel. Intel does not need to displace TSMC as the industry leader to benefit from the current environment. Instead, it can act as a pressure valve for an overheated supply chain, offering something that has become increasingly valuable: available capacity. Intel's foundry business is scaling its New Mexico facility's EMIB packaging capacity by 30%, and both Google and Apple have reportedly explored using Intel for future chips.
Still, don't expect the shortage to end soon. TSMC management acknowledged on last week's earnings call that they're "very nervous" about the scale of spending required. CEO Wei says he's satisfied that AI demand from customers is real, but admitted the industry faces uncertainty about whether semiconductor growth can continue at this pace for years.
The structural problem: AI chips generate vastly more revenue per wafer than consumer products. An Nvidia H100 sells for $25,000 or more. Consumer CPUs can't compete for allocation when the economics look like that.
TSMC's Q1 2026 guidance projects revenue of $34.6-35.8 billion, up 38% year-over-year. The AI boom shows no signs of slowing, but neither does the capacity crunch.
