The datacenter is tight as a drum. DRAM main memory and flash memory prices are going nuts as the hyperscalers, cloud builders, and AI model builders are trying to get hefty configurations to drive the performance of their expensive GPUs and XPUs. Which for the most part have been allocated for all of 2026, just like CPUs from Intel and AMD as well as the homegrown crowd.

HBM memory capacity, which is vital for most AI accelerators, has probably been largely allocated, although a case could be made that Micron Technology, Samsung, and SK Hynix have very little incentive to lock down HBM pricing now for future sales because, just like DRAM and flash, every time anyone looks, the price goes up.

My advice is to stop looking. . . .

Similarly, demand for the chip etching and packaging from the world’s largest, most advanced, and most consistent foundry exceeds supply, which gives Taiwan Semiconductor Manufacturing Co a lot of control over its pricing. But even with Intel trying to get its foundry act together – and there are hints that it is doing that – and Samsung getting better and better each year at manufacturing chips for others in its foundry, TSMC does not really have much in the way of competitive pressure.

The company can walk away from business if it chooses (we are not saying that it does, but it can certainly has to focus on the most profitable deals it can make), and still the revenue per wafer it can extract from its foundries in Taiwan, the United States, China, and Japan still keeps rising every quarter.

And the number of etched wafers it can produce each month has finally come out of the doldrums of 2023 and 2024. So revenues are rising and net income is rising faster, as it has been since Q3 2024, when the GenAI boom went from chemical to nuclear.

Measured in 12-inch (300 mm) wafer equivalents across all of its machines and all of the many different process nodes that TSMC still etches chips in, the foundry made 4.17 million wafers in Q 2026, which was a record for the company, besting even the 4.09 million wafers it produced in Q3 2025. That wafer rate in the March quarter was 28.1 percent higher than the year ago period and up 5.4 percent sequentially. Of course, chips come in all shapes and sizes, with dozens to hundreds per wafer, so it is hard to get a sense of how many chips were produced by TSMC, but it surely knows this number and tracks it very carefully as well.

The revenue it is getting per wafer is also growing, which suggests more complex chippery and the consequent packaging that is implied from getting a high-end chip fabbed by TSMC. Only four years ago, TSMC was averaging less than $5,000 per wafer of revenue (that’s total revenue divided by wafer output in 12-inch equivalents), but in the March quarter, the revenue per wafer is up 1.7X compared to four years ago to $8,600. That revenue per wafer was up 9.8 percent year on year from the $7,832 level set in Q1 2025, but only up 1 percent sequentially.

Add all those wafers and packaging up, and TSMC posted sales of $35.9 billion, up 40.6 percent year on year in the March quarter, and net income exploded by 65.2 percent to $18.13 billion, representing a crazy high 50.5 percent of revenues. TSMC is a capital-intensive, highly complex hardware manufacturer that has margins that rival any software stack in history, including the venerable IBM mainframe that is still keeping the company afloat because of the stickiness of the applications written over the decades on the System/360 and its progeny.

TSMC ended the quarter with $105.53 billion in cash and equivalents in the bank thanks to those booming profits, and it is only projecting to spend somewhere between $52 billion and $56 billion in capital expenses in 2026. During its call with Wall Street analysts, TSMC just raised that to the “upper end” of that range. So call it $56 billion. Capital spending will no doubt grow in the coming year and beyond, but it is handy that TSMC has nearly two years of capex in the bank and it is only the beginning of the current year.

But what does that amount of money really buy? TSMC says that somewhere between 70 percent to 80 percent of the capex this year will go for foundries, and that means a blend of expansion of the N3 node (3 nanometers), the buildout of the N2 node (2 nanometers), and the beginning of the A14 node (1.4 nanometers) in Taiwan plus the ongoing ramp of the foundry operations it started a few years back in Arizona. Depending on who you ask, for fab modules that can handle 50,000 wafer starts per month, a 3 nanometer fab module costs around $20 billion, a 3 nanometer module costs around $28 billion, and a 1.4 nanometer module costs on the order of $49 billion. Instead of moving to High NA processes, which will cut the reticle size in half but radically boost the density of transistors, TSMC is doing a lot of extra patterning with its N2 and A14 processess, which means more equipment but less grief and change for customers.

The total capacity of TSMC in 2025 in terms of wafer starts per month “exceeded 17 million 12-inch wafer equivalents,” which was a little higher than the level set in 2024. TSMC is going to have another growth spurt, to be sure, but will not and cannot add so much capacity to not be able to charge the scarcity premium it has enjoyed in 2025 and into 2026 and beyond.

To be precise: Even if TSMC could move a lot faster, given the state of the world economy and the concentration of GenAI spending among the tech titans, it would not. There is very little benefit to TSMC in creating its own bust cycle – especially after it just lived through one when PC, smartphone, and general purpose server spending all hit the skids in 2022 and 2023.

TSMC has always talked about its business in terms of revenues by platform and revenues by process node, and we will review them quickly. But what everyone wants to know is what is the concentration of TSMC’s business on chips for AI compute and networking, and that is something that the company only hints about. We have taken a stab at estimating it based on what little TSMC has said in recent quarters.

First, revenues by process:

Volume production of the N5 process started nearly six years ago, and this is still the dominant manufacturing process (along with its N4 variations) in the TSMC fleet, at least reckoned by revenue, which was $12.92 billion in the first quarter, up 40.6 percent year on year and up 9.5 sequentially. The N5/N4 node sure doesn’t look like it is peaking. These are relatively inexpensive and high-yield manufacturing nodes, and there are plenty of chip companies that hang back exactly for these reasons. Nvidia and AMD can pay for the bleeding edge nodes because they can charge the big bucks for CPUs and GPUs.

The N3 node, which is a few years younger, is finding its feet, with sales of just under $9 billion in Q1 2026, up 59.8 percent but down 5 percent sequentially. The much older N7 node (with its N6 variant) is still not only kicking, but growing, with sales up 21.9 percent to $4.67 billion. All other nodes – and there is a lot of older stuff being made in the world – comprised $9.33 billion in sales, and was up 35.4 percent in the March quarter. This kind of revenue is what the old Intel foundry model walked away from, year after year, because it had a captive customer on the bleeding edge and was not a merchant foundry. Intel would kill for such a legacy business, and someday it may even build it.

With the revenue split by products, we can start zooming in on the AI piece of the action at TSMC. As always, we have to be careful about terms here. When TSMC says HPC, it means all types of high performance chip, which includes CPUs used in PCs and servers, GPUs and XPUs, various kinds of switch and routing ASICs, and FPGAs.

It has been nine quarters since the HPC segment at TSMC was tied with the smartphone segment in terms of revenues generated. Now, the HPC segment, which posted $21.9 billion in sales in Q1 2026, is more than twice as large as the smartphone segment, which had $9.33 billion in sales. As you can see, the HPC segment is growing faster, up 45.4 percent year on year and up 18 percent sequentially – and I do not think all of that growth is driven by AI. 

There are plenty of high-end PC and server CPUs being made these days, and sold out, too. Even though the smartphone chip business grew a very healthy 30.6 percent, unless the GenAI bubble bursts, I don’t think smartphones can ever catch HPC again at TSMC. It is far more likely that TSMC will split HPC into an AI and non-AI segment, since this is in fact a material distinction that investors deserve to know about.

Which brings us to this lovely little chart:

There is a lot of witchcraft in this one, make no mistake, but out model is consistent with the long-term compound annual growth rate that TSMC has set of “mid to high 50 percent” between 2024 and 2029. That forecast was for mid 40 percent CAGR in early 2025.

I have said that eventually the GenAI boom will stop growing explosively but it will still grow very fast and get very large. And out model reflects this believe because this is the curve that new and explosive technologies follow. Moreover, TSMC’s revenues are gated much more by supply than demand, and as I said, it cannot overbuild capacity even if it wanted to. And TSMC most certainly does not want to do that.

Back in 2022, we think TSMC’s AI business was a mere $1.52 billion, but it grew at triple digits quarterly for a lot of 2023, 2024, and the first three quarters of 2025 to hit a stunning $33.38 billion for all of 2025. I think that the growth rate for Q4 2025 and Q1 2026 was a lot slower, 79 percent and 84.9 percent year on year, respectively. But in those two quarters, the AI chip business drove a tad bit more than half of overall TSMC HPC chip revenues and by next quarter it will probably be more than a third of overall TSMC revenues. (And I am not counting AI PC CPUs in that number, or CPUs with matrix or vector engines in them. They may do AI on the side, but that is not their main purpose.)

To put numbers on it, I think the non-AI portion of the TSMC HPC business drove around $10.8 billion, up 19.2 percent, and the AI portion drove around $11.1 billion, up that 84.9 percent.

That’s pretty amazing for a business that was interesting but not overwhelming only four years ago.