Intel Fab Watch, Issue 02: building and subtracting at the same time

Summary

Issue 01 left Intel in an awkward posture: the factories were producing genuine firsts while the customers those factories were built for remained mostly hypothetical. A quarter later that gap has not closed. What changed is that Intel stopped waiting for it to close before spending against it.

The decision that matters most was made inside the quarter and announced on the earnings call. Intel has “fully committed” to a high-volume manufacturing ramp of 14A — the process generation after the 18A node now running in Arizona — in 2028. As recently as June, Intel’s stated position was that 14A might be paused or cancelled outright if no major external customer signed up in time. No external customer has publicly signed up. Intel is proceeding anyway, on the strength of demand for its own products, and the pause-risk language remains in the filing even as management calls the node committed. Alongside it, Intel completed the buyout of Apollo’s stake in its Irish fab, taking full ownership of its most productive leading-edge site, and told investors capital spending will exceed $20 billion this year and run significantly higher in 2027.

Which makes the quarter’s two loudest numbers doubly unhelpful, because both point away from that story. Free cash flow swung from minus $3.9 billion to plus $4.4 billion — not because the buildout got cheaper, but because capital spending nearly halved in a quarter sitting directly before the heaviest spending Intel has ever guided to. And the reported bottom line was an $11.0 billion net loss, driven almost entirely by a non-cash revaluation of the US government’s stake, an item that grows larger precisely when Intel’s share price rises. Underneath both, the number this note actually tracks moved the way it has been moving: external foundry revenue reached $293 million, up from $174 million, and 1.8% of total revenue instead of 1.3%.

And then there is the part that a US-fab-only lens would have missed entirely. While Ohio’s two modules inch toward operations in 2030 and 2032, Intel is closing its Costa Rica test plant by the end of this year and consolidating that work into Malaysia and Vietnam. Germany and Poland were cancelled outright, with the Irish expansion absorbing capital that Magdeburg would have taken. New Mexico, once a wafer fab, is now Intel’s advanced packaging site. The network is not simply getting bigger. It is being rebalanced — fewer sites, more concentrated, with older ones exited or moved down the value chain. From this issue the whole register is published, because tracking only the additions was measuring half the transaction.

Free cash flow turned positive for the first time in this tracker, on a 46.6% quarter-on-quarter fall in capital spending. External foundry revenue — the measure most directly tied to this note’s question — rose from $174M and 1.3% of total revenue.

The network

What changedFirst full publication of the register. Six campuses have entries that changed this period: Arizona, New Mexico, Ohio, Santa Clara, Ireland and Costa Rica.

What the register makes visible

Three patterns are easier to see in the table than in any narrative. Sites move down the value chain rather than simply closing. New Mexico’s Rio Rancho campus was a wafer fab; Fab 9, opened in January 2024, makes it Intel’s advanced packaging site instead. Ownership is consolidating. Ireland went to 100% Intel in June after the Apollo buyout; Arizona remains 51/49 with Brookfield; Ohio has no partner at all and Intel is reported to be looking for one. And the assembly and test tier is being concentrated into Asia — Costa Rica’s work moving to Malaysia and Vietnam, with Kulim still building and its completion date undisclosed.

Fab register

Every Intel manufacturing campus, including those closed, cancelled and divested. The register holds to a stricter sourcing standard than the rest of this note: Intel filings and Intel investor-relations material only. Where Intel has not stated something, the cell reads “not disclosed” rather than being filled in from trade press. There are a lot of those cells, and that is itself the finding — Intel discloses relatively little about the composition of its own network, and this register will fill in as that changes.

Location and first-operations dates are folded into the campus name and notes below to keep the register readable in portrait. Rows marked ▲ changed this issue.

United States

International

Winding down, cancelled and divested

Sourced from Intel filings and Intel IR only; anything Intel has not stated shows as “not disclosed” rather than being filled from trade press.

Financials

What changedQ2 2026 landed on 23 July, bringing the first balance-sheet view of April’s $14.2bn Ireland buyout and $6.5bn note issuance. Revenue, margin and cash generation all rose; capital spending fell sharply in a quarter positioned ahead of a guided second-half surge.

Headline figures, Q1 2026 to Q2 2026. Source: Intel 10-Q and earnings 8-K filings.

The capex number is a trough, not a trend

Capital spending fell 46.6% quarter on quarter, and free cash flow — operating cash flow minus capital spending — turned positive for the first time in this tracker. Read in isolation, that looks like a buildout winding down. The guidance says otherwise. Intel spent $7,615M across the first half against full-year guidance of more than $20 billion given on the same call, which implies roughly 62% of the year’s capital spending falls in the second half, and the CFO said 2027 will run significantly above 2026. Quarterly capital spending in a fab programme is lumpy by nature — it tracks when tools are delivered and paid for, not steady construction progress — so a single quarter tells the reader little. The mechanically useful reading is that Q2’s positive free cash flow is a timing artefact sitting immediately before the largest spending period Intel has guided to.

The $11 billion loss has no cash in it

Intel reported a GAAP net loss of $11.0 billion and diluted earnings per share of −$2.16, against non-GAAP EPS of $0.42. The gap is almost entirely one line: a $12,529M non-cash mark-to-market loss on the Escrowed Shares derivative — the accounting treatment of the warrant and escrow features in Intel’s agreement with the US government, which took a 9.9% equity stake in August 2025. That derivative is revalued every quarter, and the mechanism inverts the usual reading of an earnings statement: when Intel’s share price rises, the government’s rights become more valuable, and that registers on Intel’s income statement as a larger loss. No cash moves. Nothing about fab utilisation changes. For a reader tracking whether capital spending converts into durable cash generation, this line contains no information — but it will keep distorting headline GAAP figures for as long as the agreement is outstanding. Confirmed

Positive free cash flow, and net debt up $8.6bn anyway

These two facts sit together because the Ireland buyout does not appear in the free cash flow calculation. Intel paid Apollo $14.2 billion in April; in the cash flow statement that sits within partner distributions, a financing item, while free cash flow measures operating cash flow minus capital spending. So the quarter can show $4.4 billion of free cash flow generated and $8.6 billion of net debt added without contradiction. Total debt rose $5,506M, reflecting the $6.5 billion of senior notes issued on 30 April in five tranches maturing between 2031 and 2066. Confirmed

Foundry segment

Intel Foundry — the manufacturing arm, whose revenue is still substantially Intel manufacturing for itself — posted revenue of $5,765M, up 31% year on year and 6% sequentially. Its operating loss narrowed to $(2,089)M from $(2,437)M in Q1 and $(3,168)M a year earlier. This is the first quarter in this tracker where the Foundry loss narrowed both year on year and sequentially rather than widening. Confirmed

Q3 2026 guidance: revenue $15.8–16.8B; gross margin 41.0% GAAP / 42.0% non-GAAP; diluted EPS $0.31 GAAP / $0.38 non-GAAP at the revenue midpoint. Full-year operating expenses approximately $23.0B GAAP.

Nodes and customers

What changedThe node roadmap gained a commitment it did not have a month ago, and lost none of its conditionality in the filings. The confirmed customer list did not change at all.

14A: committed, and still uncommitted

On the Q2 call the CEO said Intel decided during the quarter to fully commit to a 14A high-volume ramp in 2028, with risk production — the early stage where a process is frozen on production tooling so customers can commit designs before full yields are proven — targeted for the second half of 2027. Set against Intel’s June position, that is a reversal: the company had told investors 14A could be paused or cancelled without sufficient committed external demand.

Two qualifications matter. The commitment as stated covers Intel’s own products, and reporting indicates no external customer has publicly committed volume to 14A, which remains the gating factor for tooling Fab 62 at Ocotillo. Speculation And the pause-risk language did not leave the filing — the Q2 10-Q retains it unchanged from Q1. The commitment therefore lives in management commentary rather than in disclosed risk, which is a meaningful distinction about where it sits between intention and obligation.

Output and yields

The primary-sourced statement is the CFO’s: Q2 volume upside came from “higher factory yields and improved cycle times.” Confirmed Trade-press summaries of the call add that 18A output ran roughly 25% above internal target and rose more than 50% sequentially; those figures were not verified against filing text, and analyst supply-chain checks putting 18A yield near 85% remain unconfirmed by Intel. Speculation Intel’s own public position is unchanged — yields improving roughly 7–8% per month, reaching industry-standard levels in early 2027. Mechanically, output rising faster than plan on a node whose yields were the binding constraint through 2025 is the supply-side precondition for serving external customers at volume. It says nothing about whether the demand exists to absorb it.

Two execution markers landed on schedule: 18A-P entered risk production, and the Q2 8-K confirms Intel Foundry entered high-volume manufacturing for a subset of Core Ultra Series 3 using ASML’s High-NA EUV tools — moving July’s newsroom shipping claim into company financial disclosure. Confirmed

Customers, and a new way in

External third-party foundry and assembly/test revenue was $293M in Q2 against $22M a year earlier, and $467M for the first half against $53M — growth of roughly 13× year on year from a base small enough that the multiple carries limited information. Against $16,128M of quarterly revenue it remains under 2% of the total. The confirmed roster is unchanged: Microsoft, AWS, the US Department of Defense, and MediaTek and Tower on mature nodes. The widely circulated list of rumoured design wins — AMD, Nvidia, OpenAI, Marvell, Micron, and an Apple evaluation — remains unconfirmed by Intel or by any company named. Speculation

One genuinely new disclosure: on 29 July Intel said some foundry customers are engaging it for advanced packaging before any wafer business. That is a different customer path from the one this note has been tracking. Packaging does not require a customer to port a design onto an Intel process node, which is the expensive, slow, high-commitment step. It is a lower-barrier entry point that can precede a wafer relationship — and can also exist without ever becoming one. Intel did not name the customers. Confirmed

A gap worth naming

Clearwater Forest, Intel’s first 18A server processor, launched on 1 June and still has no published independent laboratory benchmarks — the available performance figures are Intel’s own. That gap was flagged in Issue 01 and has not been re-checked since 17 July, so it is recorded here as open rather than as unchanged. Carried forward

Competitive position and ecosystem trust

What changedThe foundry chart has a second data point and immediately shows Intel and TSMC crossing in opposite directions. One methodology change is disclosed below. Product share has no new data.

Ecosystem and policy

No new material this period. Both major chip-design software vendors, Synopsys and Cadence, remain certified on 18A and 18A-P, and Cadence’s multi-year 14A agreement from 8 June stands; no equivalent Synopsys commitment on 14A has been announced. The process design kit for 14A — the technical rulebook an external customer needs before committing a chip design to a node, which Intel has called the milestone gating customer sign-off — is still targeted for October 2026 and is still a future event. On policy, the Section 232 regime imposing a 25% duty on certain advanced chips, with tariff-free quotas tied to US fab construction, remains in force; the Commerce review of data-centre chip tariffs due on 1 July has no reported outcome in the sources this note tracks.

Worth watching next

Scheduled or expected events, not predictions of outcome.

Sources and method

Primary sources are Intel’s Q2 2026 Form 10-Q and Q2 2026 earnings Form 8-K, both filed 23 July 2026, the Q1 2026 filings, the 30 April 2026 Form 8-K on the senior notes issuance, and excerpts from the FY2025 Form 10-K filed 23 January 2026. Fab and node developments are drawn from Intel Newsroom releases and named-source trade press including Tom’s Hardware, Semafor, EE Times and Tech Times. The fab register is built from Intel filings and Intel IR material only. Competitor capital spending figures were obtained by targeted lookup on 8 August 2026: TSMC capital expenditure from its cash flow statement and revenue from its Q1 and Q2 2026 results; Samsung Device Solutions division figures from its Q2 2026 results reporting. CPU share figures are Mercury Research via trade press.

Confidence labels. Confirmed means stated in an SEC filing, company release, or on the record by a named party. Speculation means analyst-sourced, anonymously sourced, or reported but not confirmed by the companies involved. Carried forward means the most recent primary-sourced statement is reused because it was not re-verified this cycle.