# AI Chips Drive Around A Third Of TSMC Revenues > Source: > Published: 2026-07-16 21:13:48+00:00 # AI Chips Drive Around A Third Of TSMC Revenues There is an old saying in the IT sector that probably dates back to the mainframe era that when it comes to electronic component and overall system prices, when demand exceeds supply you can milk the cash cow but you can’t rip the udders off. Anyone peddling corporate computing products, be it for traditional systems or for AI clusters, are trying not to chap the tender bits of those Bessies but are milking as fast as they can as the supply of investment for AI systems is proving to be insatiable because everyone is counting on getting richer from this AI megatrend. It remains to be seen if these trillions of dollars of incremental investment in the IT sector will pan out, but if the budgetary actions by hyperscalers, cloud builders, model builders, neoclouds, and sovereigns are an indication, then AI will be responsible for doing as much work as people before the end of the decade. No one really knows what that means for the global economies and their cultures, at least not to my satisfaction and probably not yours. It is in this whirlwind of incremental spending driven by GenAI enthusiasm that Taiwan Semiconductor Manufacturing Corp has to do its capacity planning. The world’s largest foundry talks not only to its biggest customers to get capacity planning data, but goes one level deeper and talks to their biggest customers to get a read on what is happening. And then the top brass at TSMC make their best estimate of what the truth is. And as such, TSMC’s revenue and capital expense projections are probably the best gauge of what will happen. At the moment, CC Wei, the foundry’s chief executive officer and chairman, is not ready to update its forecast for its AI-related businesses. Back in January, TSMC said that it expected for its AI business would grow in the mid-to-high 50 percents at a compound annual growth rate between 2026 and 2030, inclusive, which is pretty impressive. Wei would not give an updated number during his call with Wall Street today, but he did say the number was increasing and the CAGR is getting “stronger and stronger and stronger.” Wei joked a bit about how fuzzy all of these numbers are even as TSMC has more visibility into the future than it usually has because customers want to lock in capacity at the foundry so they don’t miss out on AI opportunities in the future. “Now remember that I believe every customer tells me the truth – every one. You put all the truths together, it's not the truth. So we have to make some of the judgement. You know what I mean, since you are laughing. Because all the customers are very aggressive, right? That's the CEO's job. The CEO has got to be aggressive. So they give me the number for their demand, and I believe they try their best to tell me the truth. So I put all together, all the truths together is not a truth. Mark down that word. So yes, we do a very careful judgment. It may not be correct, but we did it carefully because this is a big money.” And it is big money indeed. TSMC announced during the call that it was adding another $100 billion to expand capacity in its Arizona fabs, which brings the total investment in its US operations to $265 billion over several years. (We don’t know if it is three, four, or five years because TSMC is not sure how much customers will want it to ramp or how fast.) We do know that this $100 billion will result on four more fabs in Arizona, and they will be focused on 2 nanometer and smaller transistor geometries. Arizona is currently getting a 3 nanometer foundry, which was already part of the plan, and mirroring new 3 nanometer foundries in Taiwan and Japan. The company is also converting 5 nanometer foundries to 3 nanometers to boost capacity of this node. As you can see below, TSMC’s revenues for the 3 nanometer node are ramping, but 5 nanometer processes still drive slightly more revenue: You will also not that this is the first quarter where 2 nanometer processes recognized revenue, which started out at $1.21 billion. A lot of this driven by AMD’s “Venice” Epyc 9006 CPUs and the MI450 GPUs that are expected later this year. (Perhaps next week, in fact, at AMD’s Advancing AI 2026 event. . . . ) Others will follow as capacity and budgets permit. I assume 2 nanometer chips are more expensive than 3 nanometer chips, and certainly are riskier to make. But someone has to drive the ramp to better yields and AMD is doing it now just as Nvidia did it with 4 nanometer TSMC processes a few years back. Wei reminded everyone that it takes five to seven years to build these fabs, by the way. So it does not happen overnight. Which is why the annual capex for TSMC is smaller than the commitments to spend. To give you some perspective, TSMC spent $29.8 billion in capex in 2024, and increased that by 37.4 percent to $40.9 billion in 2025. Earlier this year, TSMC expected to pump $52 billion to $56 billion into capex, and now that has been boosted to somewhere between $60 billion to $64 billion. At the high end of that range, that is a 56.5 percent increase in capex for the year, and would burn more than half of the $110.22 billion in cash and equivalents that TSMC has in the bank and in the stock markets. Some of this money is for increased capacity for 2026, but some of it is also driven by price increases on facilities and equipment, which are inflating as demand exceeds supply all over the world. My guess is that slightly less than half of the increase in capex is from inflation, and the rest is for actually capacity expansion beyond the original plan. In general, about 70 percent of this capex is for front-end chip wafer production for all kinds of chips (CPUs, GPUs, DPUs, smartphones, embedded devices, etc), with 10 percent to 20 percent being allocated for chip packaging facilities. The remaining 10 to 20 percent is for specialty products. Wei said that going forward those ratios will stay roughly the same, and added that he was pleased that Intel was getting traction on its EMIB-T interposer technology because it took some of the heat of its CoWoS interposer packaging supply, which is even together than its chip etching supply. This would seem to argue that TSMC should spend more money on CoWoS and the now-ramping COUPE co-packaged optics packaging. But TSMC wants the cow to give big milk, and having supply shortages in high demand means that TSMC can drive up prices. In the second quarter ended in June, wafer shipments in 12-inch wafer equivalents (some older processes are still on 8-inch wafers) were 4.34 million wafers, up 16.6 percent year on year and up 3.9 percent sequentially from Q1 2026. Revenue per wafer came to $9,271, up 14.6 percent year on year but up 7.8 percent sequentially. Four years ago, the revenue per wafer was half as much, if you can believe it, and seven years ago, revenue per wafer was a third of what it is today. Clearly, advanced processes are more expensive and also more profitable, particularly with demand exceeding supply. In the June quarter, TSMC reported record revenues of $40.2 billion, up 33.7 percent year on year and up 12 percent sequentially from Q1 2026. But net income grew 2.2X faster, up 74.7 percent year on year to $22.37 billion. This amounted to 55.6 percent of revenues. You can see why TSMC might not be in such a hurry to have demand met perfectly by supply. Shortages are more profitable. The trick is to not be *too short*. That will only embolden Intel Foundry and Samsung Foundry, the latter of which has plenty of money thanks to the DRAM and HBM memory booms to ramp up its CPU and GPU business and the former which can take a bite out of TSMC’s packaging operations. China’s Semiconductor Manufacturing International Corp is backed by the Chinese government and has 6 percent of the foundry business at the moment and will grow its share as China is basically shut out of using TSMC for advanced chips. TSMC will eventually have competition, just as Nvidia will. Take a very deep breath before you hold it waiting for those days because it may take another four or five years. In the meantime, that “HPC” business at TSMC, which includes all of the various datacenter XPUs as well as switch ASICs and FPGAs plus desktop and laptop CPUs and APUs, is growing like gangbusters and now far exceeds the revenues that TSMC derives from smartphone circuits, which was the big driver of its business for many years. Specifically, the HPC segment drove $26.53 billion in sales, up 47.1 percent year on year and dominated by server CPUs and datacenter XPUs. The smartphone business cooled by 5.2 percent sequentially but was still up 8.9 percent year on year to $8.84 billion in sales. All other segments – IoT, Automotive, DCE, and others like military/aerospace – brought in $4.82 billion, up 23.4 percent year on year. What everyone wants to know is how much AI is driving sales at TSMC. The company used to hint enough about this that you could make an educated guess, but stopped doing that many quarters ago, Just for fun, we have been taking a stab at estimating it, Take a gander: My best guess is that AI training and inference chips manufactured by TSMC represented $13.31 billion in Q2 2026, up 68.3 percent and way faster than that mid-to-high 50 percent CAGR in the forecast that TSMC provided earlier this year. Non-AI HPC chips (using TSMC’s definition of HPC) represented maybe $13.23 billion, up 30.5 percent. If my model is right, AI products represented about a third of overall TSMC revenues and about half of HPC segment revenues. I would love to know how much AI drove profits, but I reckon AI chips are almost certainly more profitable than any other segment with perhaps truly ancient processes used for obscure purposes that have extremely high yields. And that is a relatively small part of the TSMC business, and even if unit profits are high, profit dollars are likely smaller than for AI products.