Why Intel Foundry Is Losing Money: Fabs Without Customers

Intel Foundry is losing money because it has built leading-edge chipmaking capacity faster than it has found outside customers to fill it. A fab's cost is dominated by the depreciation of tools and cleanrooms, so a fab that is not full loses money quickly, whatever the quality of its process. In the third quarter of 2026, Intel Foundry lost $1,853 million on revenue of $6,697 million, and most of that revenue came from Intel's own product groups rather than outside designers.

Two Arizona fabs tell the story. TSMC Arizona announced its first Phoenix fab on 15 May 2020, ran into labor shortages and costs that Morris Chang put at about 50% above Taiwan's (reported), and still began volume production in the fourth quarter of 2024. Its US subsidiary went from a loss of about NT$14.3 billion in 2024 to a reported profit of about NT$16.1 billion in 2025. Intel Foundry declared its Fab 52 in Chandler fully operational on 9 October 2025, with its new 18A process running above target by mid-2026. One had customers waiting. The other had a working process and was still looking for them.

Why is Intel Foundry losing money?

A leading-edge fab is the most capital-intensive factory ever built. A single site costs $20 billion or more, and its cleanrooms hold hundreds of tools priced from a few million dollars to more than $350 million each (reported). Depreciation, not labor, is the largest part of a wafer's cost. That gives fabs the most extreme operating leverage in manufacturing: full, they are very profitable; half-empty, they bleed.

Intel's fabs were built for a world in which Intel's own processors filled every one of them. That world ended when Intel lost its process lead. Its 10 nm process slipped for years from 2015, and TSMC took the lead. In March 2021, under chief executive Pat Gelsinger, Intel launched IDM 2.0: it would regain process leadership with "five nodes in four years", build fabs in Arizona, Ohio, Germany and Poland, and sell foundry capacity to other chip designers. Capital spending ran at roughly $25 billion a year in 2022–24, by reported estimates, while revenue fell.

Company figures as reported.

The losses are narrowing as 18A ramps and revenue grows, up 58% year on year in the third quarter of 2026. But the core problem remains. The capacity was built at the scale of a foundry, and the demand that fills it is still mostly Intel's own.

How does a foundry fill its fabs?

The industry has two ways to fill a fab. An integrated device manufacturer designs and makes its own chips, as Intel historically did, along with Samsung, Texas Instruments and Micron. A pure-play foundry makes chips designed by others, as TSMC, GlobalFoundries and SMIC do.

TSMC's model has been the decisive manufacturing innovation of the past three decades because it pools demand. One customer's slump is offset by another's growth, and its volume lets it learn yields faster than anyone. The book's chapter on chips as a manufacturing business shows the effect with invented numbers. Five chip designers whose demand rises and falls at different times would need 61,000 wafer starts a month of capacity if each owned its own fab, running at about 73% on average. One foundry serving all five needs 48,000 and runs at about 93%.

Stylized example from Chapter 8 of The Production Equation; invented numbers.

At a stylized $500,000 of capital per monthly wafer start, pooling saves $6.5 billion of capital for the same output. The advantage holds only if customers trust the foundry with their designs. An integrated maker that opens a foundry must earn that trust while competing with its customers' products. Until it does, it carries the capital cost of separate fabs with the utilization of a single customer: itself.

What happened to Intel's Ohio fab?

In January 2022 Intel announced a new site at New Albany, outside Columbus, Ohio: an initial investment of about $20 billion in two leading-edge fabs. It broke ground in September 2022, weeks after the CHIPS and Science Act was signed. Ohio offered grants of about $300 million per fab, tied to completing construction by the end of 2028 (reported). There was no anchor customer.

In February 2025 Intel moved the first fab's completion to 2030, with operations in 2030–31, and the second's to 2031 and 2032. By early 2026 it had spent about $5 billion on construction (Axios, reported), and in October 2026 the first major roof truss went up. Chief executive Lip-Bu Tan, who took over on 18 March 2025, said projects like Ohio would proceed only when enough customer demand was secured (reported).

That is a rational decision, not a failure of nerve. A fab's shell can stand unequipped until demand justifies installing tools, which is the cheapest way to keep an option on capacity. The tools are most of the capital, and they depreciate from the day they are placed in service. Building them before customers commit is how a fab becomes a loss. In July 2025 Intel went further and cancelled its planned Magdeburg fab in Germany, a project of about €30 billion, and an assembly and test plant in Wrocław, Poland (reported).

The book's plants and ramps chapter lists the stages a plant passes through before it earns: announced, incentive agreement, permits, groundbreaking, building complete, first production, customer qualification, ramp and volume. Ohio, in late 2026, was a building. Phoenix, announced two years before Ohio's groundbreaking, had been in volume production since late 2024.

How did the CHIPS Act become a government stake in Intel?

Intel's final CHIPS award, $7.865 billion, was finalized in November 2024 as a grant paid in tranches against construction, tool-in and production milestones, with clawbacks if commitments were missed. On 22 August 2025 it was converted. The government agreed to buy 433.3 million new Intel shares at $20.47, about $8.9 billion, making it the owner of about 9.9% of the company. It paid with about $5.7 billion of CHIPS grants not yet received and about $3.2 billion from the Secure Enclave program for defense chips. It also took a five-year warrant for another 5% at $20 a share, exercisable only if Intel ceased to own at least 51% of its foundry.

The conversion removed the unachieved project milestones, so money that might have been withheld as Ohio slipped became shares that could not be reopened. Other investors followed. SoftBank bought $2 billion of shares at $23 in August 2025, and Nvidia's roughly $5 billion investment at $23.28 closed on 26 December 2025. Intel then raised about $23 billion of equity in 2026.

The book's section on industrial policy, CHIPS and subsidies draws the lesson plainly: a subsidy can lower the cost of capital, but it cannot supply a customer. Public money belongs to the fourth factor in the production equation, capital matched to the cycle. It does not create the third, utilization and scale.

Disbursement tells the same story across the program. By April 2026 about $13.1 billion of roughly $31.2 billion of direct CHIPS funding awarded had been paid out, about 42% (secondary summary of a Government Accountability Office report, reported).

Is Intel's 18A process working?

The manufacturing evidence has improved. 18A, with gate-all-around transistors and backside power delivery, went into volume at Fab 52 in Arizona, and Intel's second-quarter 2026 call said output was above target and 18A-P was in risk production. Equipment installation began at Fab 62 in Arizona in October 2026, with operations expected in 2028, and Intel released version 0.9 of the design kit for its 14A process the same month.

For a foundry, though, the product is the process, and customers buy confidence that it will ramp on schedule. Intel has named Microsoft and AWS as 18A-related customers (2024) and the US Defense Department. A reported Apple foundry deal was unconfirmed by the companies as of November 2026. Tan said customers would begin making firm supplier decisions between the second half of 2026 and the first half of 2027 (reported). Those decisions, not the roof truss in Ohio, will decide whether the losses end.

What Wolfspeed shows about the same risk

Intel is not the only US chipmaker caught by capacity ahead of demand. Wolfspeed sold its LED business in 2021 to focus on silicon-carbide power chips for electric vehicles and opened the world's first 200 mm silicon-carbide fab in Mohawk Valley, New York. Then EV demand growth slowed, Chinese substrate makers cut prices and the ramp lagged. Its CHIPS preliminary terms of up to about $750 million were never finalized. It filed a prepackaged Chapter 11 on 30 June 2025 and emerged on 29 September with about 70%, about $4.6 billion, of its debt eliminated (reported).

Intel had a balance sheet, a government shareholder and new equity to absorb the same mismatch. Wolfspeed did not. The list of the biggest manufacturing failures puts Wolfspeed among the clearest cases of utilization turning a sound process into a failed balance sheet.

The takeaway

Intel Foundry's losses turn on utilization and scale. Intel rebuilt much of its process capability with 18A and secured capital from the government, SoftBank, Nvidia and its own equity raise, but a leading-edge fab network built at foundry scale cannot carry its depreciation on one customer's demand. Slowing Ohio to keep the option without buying the tools was the right call; the open question for 2027 is whether outside designers commit volume. The full comparison of Phoenix and New Albany, with the pooled-demand example, is in Chapter 8 of The Production Equation, available to preview on the book page.