--- title: "Why Are There Only Three Leading-Edge Chip Foundries?" url: "https://goremy.ai/blog/articles/intel-foundry-vs-tsmc-economies-of-scale-capex" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-26T23:40:06.174Z" updated: "2026-10-26T23:40:06.174Z" topic: AI Industry & Strategy description: "Fixed costs near 80%, yield leverage, and a capex gap of roughly $20B vs $50B explain why three foundries remain and what Intel must do to catch TSMC." --- # Why Are There Only Three Leading-Edge Chip Foundries? > Fixed costs near 80%, yield leverage, and a capex gap of roughly $20B vs $50B explain why three foundries remain and what Intel must do to catch TSMC. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-10-26 ## Why are there only three leading-edge chip foundries? Because chipmaking at the leading edge is a fixed-cost business, and only the largest players can fill their fabs enough to pay for them. An Intel Foundry executive describes roughly 80% of a fab's cost structure as fixed. Once that is true, every extra point of utilization and every extra good die drops almost straight to the bottom line. Smaller players could not keep up, so the field narrowed. The executive notes a chart showing around 20 foundries 40 years ago, and only three at the advanced node today. ## TL;DR - About **80% of a fab's costs are fixed**, so utilization swings profit far more than in most manufacturing. - **Yield gains are not linear.** Moving die yield from 60% to 65% or 70% yields much more output from the same footprint. - Intel says it is **near parity with TSMC on technology**, within a few percent on performance, power, or area, but well behind on wafer scale. - The reported capex gap is **about $20B for Intel versus about $50B for TSMC** this year, though Intel calls the comparison unfair. - Tool lead times mean **supply cannot rise much before 2030**, even though marginal output can grow within about 12 months. - Memory, not compute, is now the tighter supply crunch, according to the Intel executive. ## How do fixed costs shape the foundry business? Fabs are enormously expensive to build, and that cost does not shrink when the fab runs half empty. With about 80% of costs fixed, the economics reward whoever keeps the line full and productive. Two levers matter most: - **Utilization.** Each additional percentage point of used capacity adds revenue against a cost base that is already paid for. The Intel executive says the company fights for every extra percent. - **Yield.** If a process produces 60% good die and an improvement lifts it to 65% or 70%, the gain is outsized. You get more sellable chips from the same wafers, tools, and floor space. This is why the executive calls scale economics in semiconductors probably the purest of any advanced manufacturing sector. Larger volume spreads fixed cost thinner and generates more learning on each process, which feeds back into yield. ## How big is TSMC's scale advantage over Intel? In the interview, the interviewer puts TSMC's leading-edge wafer shipments at upwards of 3x Intel's. The Intel executive does not dispute the point and says TSMC is also larger than Samsung. His framing is that the real gap is scale, not technology, and that closing it will be "a slow process, a gradual process." TSMC's N3 node is now majority AI accelerators, and possibly majority Nvidia alone, according to the interviewer. That leaves overflow demand, including a large influx of CPU demand driven by AI workloads. Intel's pitch is that this overflow is an opening. ## Is Intel really at parity with TSMC on technology? Intel's claim: it is already there, maybe a few percent off on performance, power, or area density. The executive argues that benchmarking only matters much when a foundry is a full generation or two ahead or behind. Within a few percent, he says, "you're splitting hairs." He gave several reasons to distrust single-number comparisons: - A process node is not one transistor. A node like 18A or 14A includes low-leakage and high-drive-current variants. - It also includes high-density and high-performance SRAM, plus analog and RF functions. - Different customers weight these differently, so competitiveness is "the sum of all of those." The interviewer cited SemiAnalysis numbers suggesting Intel's 14A, expected around 2028, would be comparable to TSMC's N2 on transistor density. The executive said he would dispute that benchmarking but did not offer replacement figures. Treat any density comparison as contested, and note that these are claims from a company with a stake in the answer. His stronger argument is about execution. Being slightly off on specs is survivable if you ship on time, with high yields that support very large and rapid ramps. He said that matters more than closing a few percent on performance or area. ## How do the capex numbers compare? | Company | Approximate capex this year (as cited) | |---|---| | TSMC | around $50 billion | | Intel | around $20 billion | Those figures came from the interviewer, not from the Intel executive. His response was that the comparison is "a little bit of an unfair" one. TSMC is a foundry for nearly everyone. Intel's foundry mostly serves Intel today. As it takes on outside customers, he said, Intel's capex will grow. That matters because the tools bought now decide who can serve the next wave of AI demand. A foundry that cannot fund capacity cannot win the customers who would fill it, and a foundry without those customers struggles to justify the capacity. That loop is the scale problem in miniature. ## Why won't chip supply ramp much before 2030? The interviewer summarized the executive's projections: marginal production can rise fast, on the scale of 12 months, but total increases through the end of the decade are limited. Next year's output could beat forecasts. Output in 2030 probably won't differ much from current forecasts, unless the whole supply chain accelerates quickly. The executive's reasoning is lead time. It is already mid-2026. If you place a huge tool order today and tools take about 18 months to arrive, you are into 2028. If you also need a new shell and a new fab, you are near the end of the decade. Squeezing more from existing fabs is quick. Building new ones is not. ## Is memory a bigger bottleneck than compute? The executive thinks so. He said the world "will eat up as much compute as can possibly be made" and as much memory as can be made, and argued that the supply crunch is larger for memory today than for compute. The interviewer raised Intel's history here. Intel began as a DRAM business in the late 1960s, later sold its DRAM operation, and sold its flash business a couple of years ago. Those are now very large industries because of AI demand. The executive conceded that, in hindsight, it was "probably not the smartest time to sell off a flash memory business," while saying there were good reasons at the time and that nobody foresaw the demand. ## Can Intel catch up to TSMC? On technology, Intel says it is close enough. On scale, it is not close, and scale is where the fixed-cost math bites. The path it describes depends on three things: 1. Hitting yield and schedule targets on its advanced nodes so ramps are rapid. 2. Winning external customers to raise utilization and justify more capex. 3. Using TSMC's saturation, especially the AI-driven CPU demand, as an entry point. None of this is guaranteed, and the evidence here is one executive's view. What the economics do make clear is why there are so few players: falling behind on volume raises unit cost, which makes it harder to fund the next node. ## Frequently Asked Questions ### What does "80% fixed cost" mean for a fab? It means most of a fab's expenses, such as buildings, tools, and depreciation, do not change with output. When utilization rises, revenue grows much faster than cost. When it falls, losses grow quickly. ### Why do yield improvements matter so much? Yield determines how many good die you get from each wafer. Going from 60% to 65% or 70% gives far more sellable chips from the same fixed footprint, so the benefit is larger than the percentage change suggests. ### How many leading-edge foundries exist? The Intel executive says three remain at the advanced node. The interview names TSMC, Intel, and Samsung as the relevant players. He cited a chart showing about 20 foundries 40 years ago. ### When could new chip capacity arrive? Marginal gains from existing fabs can come within about 12 months. New tool orders with roughly 18-month lead times land in 2028, and new fabs push toward the end of the decade.