TSMC Has No Choice But To Trust The Sunny AI Forecasts Of Its Customers

If the GenAI expansion runs out of gas, Taiwan Semiconductor Manufacturing Co, the world’s most important foundry for advanced chippery, will be the first to know. And, should the GenAI market deflate, it will be because all of the big players in the market – the hyperscalers, the cloud builders, the model builders, and other large service providers – believed their own market projections with enough fervor that TSMC will shell out an entire year’s worth of net profits to build out its chip etching and packaging plants.

In going over the numbers for the final quarter of 2025 with Wall Street analysts, one of them asked CC Wei, the company’s chief executive officer, who not only talks to his chip designer customers, but their customers as well, to get a sense of whether or not this AI demand is real or not.

“I’m also very nervous about it,” Wei conceded. “You bet, because we have to invest about $52 billion to $56 billion for the capex. If we didn’t do it carefully, that would be big disaster to TSMC for sure. So of course, I spend a lot of time in the last three to four months talking to my customers and end customers’ customer. I want to make sure that my customers’ demand is real. So I talked to those cloud service providers – all of them. The answer is that I’m quite satisfied with the answer. Actually, they showed me the evidence that AI really helps their businesses. So they grow their business successfully and healthy in their financial return. I also double checked their financial status – they are very rich. That sounds much better than TSMC.”

Wei is being too modest. In 2025, TSMC set a new record for revenues, raking in a very impressive $122.42 billion, up 35.9 percent, with $55.18 billion in net income, up 51.3 percent and comprising 45.1 percent of revenues. The company spent $40.9 billion in capex over that time, which is a down payment on meeting future demand for chip etching and packaging and which is a hopeful economic act as it always is with foundries. They have to predict the relatively far future, whether they like it or not.

TSMC is focused on predicting demand, but it is also keenly aware that it is getting more expensive for every new process node to come to market. Jen-Chau Huang, chief financial officer at TSMC, said on the call that the cost of fabrication tools and the complexity of manufacturing processes have been on the rise for a long time, and it is getting more intense. The cost per 1,000 wafers for the N2 2 nanometer process is “substantially higher” than for the N3 3 nanometer process, and it will be an even bigger gap between N2 and the 1.4 nanometer A14 process. Moreover, expansion of the fabs in Arizona and elsewhere outside of Taiwan is diluting gross margins by 2 percent to 3 percent right now and as more advanced processes come online – Arizona is making 4 nanometer wafers right now – this will increase to a 3 percent to 4 percent hit to gross margins.

The good news is that TSMC has been very good at extracting more money out of each wafer it etches on behalf of customers because they have no choice but to adopt more expensive transistors to embody their designs because they need higher performing devices and lower power – particularly for AI supercomputer components – and they will pay more for that.

This is how you know AI has become a kind of HPC, because the mantra for HPC for the past six decades has been performance at any cost. (Which is different from going with the best price/performance per watt with acceptable performance, which is what clouds do, or 30 percent more performance than I need to deal with spikes, and at the lowest possible cost amortized over six or seven years, which is what large enterprises do.

In the fourth quarter, growth was a bit slower than in recent quarters, only rising 1.9 percent sequentially to a record-breaking $33.73 billion, still up a very respectable 25.5 percent. This can be partially attributed to product transitions for CPUs, GPUs, and AI XPUs lining up for the second half of 2026. So current products have already spiked and now are just growing and awaiting the next spike later this year. Net income for Q4 2025 was $16.31 billion, up 40.7 percent year on year and up 8 percent sequentially. The N3 process will cross over to the corporate gross margin average sometime this year, but that the N2 process will also start diluting profits as it ramps in the second half of 2026. We think a lot of these increased costs will be borne by chip designers, and they will pass these costs on to their customers, and so on and so on. . . .

In the sense that transistors are no longer getting cheaper – which is what really mattered over the past six decades to drive chips into more things and more powerful chips doing more stuff with each generation – Moore’s Law is dead. Moore’s Law joined the Choir Eternal several years ago, in fact, as Dennard’s Law did in the early 2000s. But none of this means advancements in transistors, chips, and packaging are unimportant, or that customers will not pay a bit more to get what amounts to much tougher engineering. Everything matters now.

Which is why despite all of the ebullience after TSMC announced its Q4 2025 results, if you listen to what Wei & Co said carefully, TSMC has to make bigger bets to drive future revenues, and charge more than history would suggest for each successive innovation, and that means chip designers have to bear those costs and pass them on, ultimate, to me and you in some fashion. And we all need to be OK with that.

It has taken $167 billion in capex and $30 billion in research and development over five years to get TSMC from the 5 nanometer ramp in late 2020 to the edge of the beginning of the 2 nanometer era as 2025 came to a close. The 7 nanometer process was still going strong during the first year of the coronavirus pandemic, and 3 nanometer processes were under early development.

Capex and R&D are on the rise to get TSMC to the end of the decade – and by how much Wei and Huang are not saying but it is reasonable to assume that it will be considerably more.

“In the last three years, our capex dollars amount totaled $101 billion, but is expected to be significantly higher in the next three years,” Huang explained. “Having said that, we continue to work closely with our customers to plan our capacity while sticking to our disciplines to ensure a healthy overall capacity utilization rate through the cycle. Our pricing will remain strategic, not opportunistic to earn our value. We will work diligently with our suppliers to drive greater cost improvements. We will also leverage our manufacturing excellence to generate more wafer output and drive greater a cross node capacity optimization in our fab operations to support our profitability.”

We would not be surprised at all if TSMC had to fork over $250 billion in capex from 2026 through 2030, inclusive. Capex was relatively flat between 2021 and 2024 as the GenAI Boom was starting, after all, averaging $31.65 billion over those four years and really only spiking a little higher in 2022 as 3 nanometer was coming online.

To get the same level of profitability as that TSMC currently enjoys – mainly because it has no real competition for high end processes and packaging – will require higher revenue levels (the ratio of revenue to capex plus R&D) than was the case in 2025. If competition ever materializes from Samsung or Intel for advanced processes, then TSMC will have pricing pressure if demand for wafers and etching does not exceed supply. Thus far, demand seems to be higher than supply. If HBM memory can ramp faster than it has, we will get to test this theory.

TSMC has run ahead of our financial model in 2025 by about $10 billion in revenues, which is pretty good, and is a bit more profitable, too. (A monopoly is a terrible thing, unless you happen to have one.) There is no reason to believe that in 2026, given the incredible advantages that TSMC has, it cannot bring 50 cents on the dollar to the bottom line. (It was 45.1 cents to the dollar in 2025, and 40.5 cents on the dollar in 2024.) TSMC has more of a monopoly than does Nvidia, and its profitability should reflect that but doesn’t. (Nvidia has delivered 55.8 cents on the dollar as net income in its trailing twelve months ended in October.)

The question everyone wants to answer is how the AI business is driving the TSMC business, and based on things the company has said and our own estimates, we have taken a stab at it:

In the past, TSMC gave us a sense of overall chip and packaging revenues driven by AI chips, including GPUs, XPUs, and networking stuff. In the latest call, the company said that sales of AI accelerators accounted for “high teens percent” of total revenues in 2025. OK, so pick a number. We pick 19.2 percent because that’s what Spock would do. That works out to $23.51 billion in revenues from AI XPUs and GPUs. Based on past data, past guesses, and reasonable extrapolation, we think overall AI revenues in 2025 were around $33.4 billion, which works out to 27.3 percent of overall revenues. We think in 2024, AI revenues for TSMC were $13.13 billion, representing 14.6 percent of revenues. That is a factor of 3.54X growth between 2025 and 2025.

Looking ahead, TSMC is now forecasting that “AI accelerators” will grow at a compound annual growth rate in the middle to high 50 percent range for the five years between 2024 and 2029. (We listened to that many times to make sure we were hearing it right.) Let’s take 57.5 percent as the midpoint. Working backwards, we think AI accelerator revenue etching and packaging drove maybe $10.2 billion in revenues for TSMC in 2024, and if you plug that into the CAGR calculator, that means in 2029 TSMC thinks AI accelerators will drive $98.5 billion in revenues. With AI networking chips, the AI revenue five years hence will probably be larger than all of TSMC’s revenues in 2025. This is a bit more aggressive but consistent with the forecast TSMC provided back in October 2025, which was considerably larger than forecasts we made based on what TSMC was saying back in April 2024. In that short time of forecasting, TSMC’s forecast for AI product sales for 2029 has more than doubled.