Three headwinds, at Intel’s own prices
Intel’s third quarter beat its guidance on revenue, gross margin and earnings, and the fourth has been guided about a billion dollars lower. Three things explain the distance; Intel has priced one of them and described the other two. This issue starts with them because they are the headlines, and then goes beneath them, link by link.
- External foundry revenue, Q3 2026
- $411M
- Share of total revenue
- 2.5%
- 14A customers Intel has confirmed
- 0
What this series watches, for the quarter: external foundry revenue, from $174M (1.3%) in Q1 and $293M (1.8%) in Q2. Intel attributes the rise “primarily” to a contract it calls non-recurring.
The first of the three headwinds is not a business story, and this issue does not treat it as one. On 14 October an explosion at Fab 28, Intel’s plant at Kiryat Gat in Israel, killed two workers and seriously injured a third, as news agencies reported that day; Intel’s filing says “employee casualties and injuries”. Fab 28 has appeared in this publication before only as a line in a register, and the people who worked there come before any figure that follows.
| Headwind | What Intel says happened | What Intel says it costs |
|---|---|---|
| Fab 28, Israel | Explosion on 14 October; two workers killed, a third seriously injured; production halted about two weeks; restart expected by end of October | Q4: about $300M of revenue and about 200 basis points of gross margin, “almost all” of it 10nm server products. Charges in Q4 for scrapped inventory, remediation and under-absorbed overhead; costs affected “through the first half of 2027”. The total is one Intel is “currently unable to reasonably estimate”; insurance recoveries are being assessed. Q1 2027 “likely to be similar” on revenue, on the call |
| The PC market | Memory shortages pushing up system prices, tight IT budgets, and PC makers cutting inventory “aggressively”. Intel ties the shortage to “the unprecedented demand for Super Intelligence”, its term for AI infrastructure | Client revenue (CCPG) $8.1B, −9% on Q2 and −5% on a year earlier, “below our expectations”. Q4 client revenue guided to fall “greater than 20%” sequentially. Intel’s view of the PC unit market: 2026 down about 12%, the second half down more than 20%, 2027 planned down about 10% |
| Supply | Shortages of substrates, the base layers chips are mounted on, for advanced packaging, and of memory, plus Intel 3 yield and cycle time; data-centre demand “continues to meaningfully exceed available supply” | Not quantified as lost revenue. Tan: “I am disappointed that we were not able to meet strong customer demand in this quarter.” The mitigation has a price: two substrate agreements signed after the quarter, $5.3–6.1B in total, about $3.3B of it paid by the end of 2027; supply prepayments of $1.2B so far this year |
Intel’s Q3 2026 Form 10-Q, earnings release, prepared remarks and call.
Put the three together and the fourth-quarter guide is what they add up to. Against the $16.5B just reported, Intel expects revenue of $15.0–16.0B, a midpoint about a billion dollars lower, with a non-GAAP gross margin of 38.0% against the 44.1% just delivered and non-GAAP earnings of $0.25 a share against $0.52. The Fab 28 effect sits inside those figures and is the one part of them Intel has sized; it has not split the rest.
Set against all of that is the part of Intel whose demand ran ahead of its supply. Data-centre revenue was $7.3B in the quarter, 78% more than a year earlier and 17% more than in the second, and Intel expects the division to finish the year “nearly $10 billion above expectations entering the year”, supply permitting.
None of this is what Fab Watch tracks. This series watches the US fabs Intel is building for outside customers: Fab 52 and Fab 62 in Arizona, owned with Brookfield, the first running 18A, Intel’s current leading process, and the second being tooled; Ohio One, under construction with no date before 2030; and the advanced-packaging plant in New Mexico. It reads them as a chain of four links: a process ready to sell, a customer who commits to it, a fab ready to make it, and external foundry revenue at the end. A link’s move lasts only if the one before it has moved, since capacity without a customer idles and revenue without a customer does not recur, so each issue reads the four in order and marks each against the last. The quarter moved the last link, and link 4 asks who bought, because the $411M does not say.
The scorecard
Three links moved up and one stayed flat, and the one that moved most carries a condition: the revenue arrived before the customer did.
The scorecard runs the four links in chain order and then the context rows, each marked against Issue 02. Up, flat or down is a fact about the reading and never a verdict on it; how each row is read was fixed before the results came, and is in Sources and method.
Figure 1. The scorecard as a chain: four links linked in order, each with its move against Issue 02, and the context rows that set the pace.
Links 1 and 2: a process to sell, and who has signed
Is there a process to sell?
A foundry customer buys a process before it buys a wafer, and the thing it buys first is the process design kit, or PDK: the rulebook and the models its engineers design against, of which 0.9 is the near-final version and 1.0 the one designs are normally committed to.
The one dated milestone due this quarter was met. Intel released the 14A 0.9 PDK in October, “as committed” and “to select customers”, and called it “our highest-quality 0.9 PDK”. On the node already in production, 18A yields rose “at a double-digit rate” to “industry standard levels ahead of our internal projections”, and 18A client processors have gone from under 5% of client CPU revenue in the first quarter to nearly a third. 18A-P, the variant designed for Intel’s products and for outside customers, is “on track for launch by year-end across our internal product portfolio”, and 14A is being prepared “for risk production next year”, a sentence that this quarter carries no internal-only qualifier where the second quarter’s did.
Intel has still given no date for 14A for external customers; the next PDK has one only from the call, where the CEO said “We should be able to do the 1 PDK by Q1 next year”.
Figure 2. Intel's dates for 14A, in order, not to scale. Filled: stated in a filing or the prepared remarks; ticked: met. Hollow: from the call.
Who has signed?
The whole list of 14A customers confirmed by Intel, as at 30 October 2026, is the same as it was in Issues 01 and 02: nobody. On other nodes the confirmed roster is as Issue 02 left it, Microsoft, AWS and the US Department of Defense on 18A, MediaTek and Tower on mature nodes, and the Q3 documents name no customer, new or old.
Intel’s policy, in the CEO’s words
“As is standard practice in the foundry business, I have consistently maintained that we will not disclose customer details.”
Lip-Bu Tan, prepared remarks, 29 October 2026
That policy fixes how this row can ever move. On the call the CEO put the corollary the same way: “If they want to announce, we will respond to that.” So the row moves when a customer announces, or when Intel responds to one, and not before. The filing’s wording is “potential significant customers to evaluate Intel 14A”, and it still ties the “scale and pace” of the fab build-out to “committed demand for Intel 14A” from Intel’s own roadmap and “design wins with potential significant external customers”. The commitment to 14A and the condition on it sit in the same paragraph, as they did in August, and the fabs remain, in Intel’s words, “fully agnostic to the underlying ISA”, the instruction set a chip runs.
What Intel does say about new external customers names nobody, and it comes in three sizes. On 14A there are “multiple potential customers”, “very actively working with us” since the PDK went out. On 18A-P there are “a couple of highly strategic customers”, with the work now on die size and IP readiness. On advanced packaging, Intel’s EMIB and EMIB-T bridge technologies “continue to see significant external customer traction”, and Intel is adding capacity in New Mexico and Malaysia “to support strong customer ramps for advanced packaging next year”. Packaging is the one place Intel claims external traction in the present tense, and, as Issue 02 noted, a customer can buy it without ever porting a design to an Intel node. The CEO added on the call that Intel will “have a couple of customers ready to announce in the next 2 quarters”, a remark that sits inside his answer on packaging, so whether it means packaging customers or any foundry customer is not clear from the transcript.
Tesla, Terafab and TSMC
In April, on Tesla’s first-quarter call, Elon Musk said “We plan to use Intel’s 14A process”; Intel has never commented, and October’s recirculation of the remark had nothing new behind it. Tesla, Terafab, SpaceX and Musk appear nowhere in Intel’s Q3 release, remarks, deck, 10-Q or call. What did move in October was Terafab’s other conversation: Musk confirmed on 3 October that Terafab is in discussions with TSMC, and TSMC’s chief executive said on 15 October that any Texas fab it joined would have to be one TSMC runs. Intel said in April that it had joined Terafab as a “strategic partner”, and has not repeated the sentence in a release since.
Whether a Tesla chip on 14A would be made in Intel’s Arizona fabs, and so reach link 4, or run as Intel’s process inside a fab someone else operates, nobody has said, and the question stands.
Link 3: are the fabs ready to take the order?
Arizona put another $0.9B of plant into service, Fab 62 turns out to be an 18A fab rather than a 14A one, and Intel says its own 18A demand will not yet fill the capacity it has.
A customer can sign whenever it likes, but it can only be served once the fab is built, tooled and running, and the number that measures that is in the 10-Q rather than the press release. For the Arizona joint venture with Brookfield, Arizona SCIP, Intel reports how much property has been placed in service, which counts as working capacity, against how much is still construction in progress.
Arizona: what moved into service
Figure 3. Arizona SCIP, the Brookfield joint venture that owns Fabs 52 and 62: plant placed in service, and plant still under construction above it, $B, at each quarter end.
Plant moving from the upper band to the lower is the charter’s test of capital spending becoming capacity, and in the third quarter plant in service rose $0.9B. Brookfield’s share of income was $197M in the quarter, “based on the placement of their manufacturing assets into service”, and the production contract obliges Intel to “operate Arizona SCIP at minimum production levels and limit excess inventory held on site” or pay damages, at volumes Intel does not disclose.
Fab 62, corrected
Issue 02 described Arizona’s Fab 62 as “intended for 14A” in its register, and in its text called external 14A demand the gating factor for tooling it. That was this publication’s own reading, not an Intel statement, and Intel has now said otherwise: the CEO on 29 October called it “Fab 62 our second high volume 18A fab”, and the CFO said Intel is “accelerating clean-room buildouts, including Fab 62 in Arizona”. Issue 02 had this wrong; this issue corrects it. Intel’s most recent date for the fab is unchanged, with tool installation “fully underway” and operations “by the end of 2028”.
Ohio, the order book and the counter-signal
Elsewhere on the US sites, Ohio has its first major structural truss, and the CEO said Intel is evaluating “unique and highly differentiated opportunities to accelerate our Ohio build-out” and hopes “to share more on this in the near future”; there is no new date, and the schedule is still the February 2025 reset. New Mexico and Malaysia are adding packaging capacity for next year. The order book behind all of it has grown: commitments for capital expenditure stood at $19.6B at the quarter’s end against $12.8B in December, the “substantial majority” due within twelve months, and firm equipment orders rose “nearly 40 percent sequentially”.
The counter-signal is that capacity is coming on faster than Intel’s own 18A products fill it. The CFO said on the call that 18A loading “isn’t going to be where we want it”, that Intel is “fixing a little bit of that” by moving Intel 3 into Oregon, and that “there will be a little bit of a underloading scenario for 18A”. Intel is short of Intel 3 wafers and substrates and long of 18A capacity, which is why 18A tools are being turned to Intel 3, and no new external 18A customer has been named since Issue 02. A fab placed in service and under-loaded is capacity, but it is also cost, and the Arizona contract’s minimum-production clause applies to it; whether the under-loading reaches Arizona is not disclosed.
Link 4: is any of it money yet?
Intel Foundry’s revenue rose by nearly a billion dollars in a single quarter and its operating loss narrowed by $236M, and both statements are true as Intel reports them. What they are made of is the question, and Intel’s segment note and its related-party note together answer most of it.
What the $6.7B is made of
Figure 4. Intel Foundry revenue by who bought it, Q1 to Q3 2026, $M, on an axis cut at $5,000M. Beneath, the Foundry operating loss, drawn downward from zero.
How Figure 4 is built
The axis starts at $5,000M, not zero, so bar heights are differences from $5,000M and not proportions; the contract’s rule is bars from zero, and this chart departs from it so the external slices can be seen. Sales to Intel’s own product groups are segment revenue less external revenue, both as Intel reports them. Altera’s wafer revenue is from the related-party note, the Q2 figure by difference from the nine-month total. Other external customers are the remainder, on the inference that Altera sits inside the external figure, which Intel has not said in terms. A narrowing loss is a shorter bar.
| Q1 2026 | Q2 2026 | Q3 2026 | |
|---|---|---|---|
| External share of revenue | 1.3% | 1.8% | 2.5% |
| Operating loss, $M | (2,437) | (2,089) | (1,853) |
The two measures the chart does not carry. The share is the firm’s calculation; the loss as reported.
Three things follow from the chart.
Most of the rise is Intel selling to itself. Segment revenue rose $932M on the second quarter, and $814M of that, 87%, was sales to Intel’s own product groups; external customers accounted for the remaining $118M, 13%. Intel’s explanation is “higher wafer volumes from Intel 18A, Intel 3 and Intel 4 process nodes, which carry higher ASPs”, which is to say more and dearer wafers for Intel’s own chips, and on the firm’s reading that is the main reason the segment’s loss narrowed. The filing’s own words: “substantially all of our Intel Foundry business supports internal manufacturing for Intel Products”.
Of the $411M external, $126M is Altera; the rise is mainly one contract. Altera, 49%-owned since Intel sold the majority in September 2025, counts as an external customer, and the filing credits the year’s rise in external revenue “primarily” to that change; whether its wafers sit inside the external figure is implied rather than stated, and this note treats it as an inference until the annual report confirms it. Altera’s own figure fell from about $181M in the second quarter to $126M, and Intel has not said why. The $118M rise on the quarter Intel’s remarks attribute “primarily” to “a non-recurring customer program”; the $379M rise on a year earlier the filing attributes “primarily” to “revenue recognized under a non-recurring foundry contract”; neither names customer, size or node. On the firm’s arithmetic the contract is at least half the quarter’s rise and at most the $285M that is not Altera, so somewhere between about $60M and $285M, and the recurring external base after Altera and the one-off is not knowable from the filings.
The loss is narrowing on cost, not on outside customers. Intel 18A wafer costs are down about 40% since the first quarter and Intel 3 costs about 15%, on better yields and shorter cycle times. Against a year earlier the loss narrowed $468M, and $793M of lower period charges did most of that work. The counter-pressure is loading: the CFO said on the call that 18A “loading isn’t going to be where we want it”, and that some 18A capacity in Oregon is being given over to Intel 3.
The word “non-recurring”
A foundry’s external revenue is supposed to be the slow kind, because a customer ports a design to the process, finalises it for manufacture, runs it through risk production, the first wafers on production tools, and only then buys wafers, for years. That is why this publication reads a chain rather than a number: there is no recurring revenue at link 4 without a customer at link 2, and no customer at link 2 without a process to sell at link 1, so a gain at link 4 with no match at link 2 does not last.
A non-recurring contract is a different animal. Intel has not said what it was, whether a one-off wafer run, a packaging job, an engineering service or a licence, but whatever it was, it is revenue that arrived at link 4 without a new name appearing at link 2; it may have come from one of the customers Issue 02 listed, and Intel does not say. Intel’s own label says it will not arrive again. It counts as external foundry revenue, correctly, and it is not evidence of a recurring customer.
What Q4 would show. If the contract was a one-off, the fourth quarter’s external figure falls back towards Altera plus whatever else recurs, and it is the first thing this publication will read in January.
The context rows: who pays, and the government
Operating cash still covered the quarter’s capital spending, by less than before; the August share sale turned net debt into net cash; the government’s terms did not move.
Who is paying for the wait?
The longer the customers take, the more this row matters. It asks two questions every issue.
| $M | Q1 2026 | Q2 2026 | Q3 2026 |
|---|---|---|---|
| Operating cash flow | 1,096 | 7,006 | 5,082 |
| Gross capex | (4,963) | (2,652) | (4,219) |
| Cash less capex | (3,867) | 4,354 | 863 |
| Cash and investments | 32,789 | 29,727 | 54,005 |
| Total debt | 45,031 | 50,537 | 50,134 |
| Net cash (net debt) | (12,242) | (20,810) | 3,871 |
Quarter-end balances; the third and sixth rows are the firm’s subtractions. Intel’s “adjusted free cash flow” for Q3, $1,639M, also counts $847M from Brookfield and $48M of incentives, less $119M of lease payments.
Is operating cash paying for capex? In the third quarter it was, by $863M on the firm’s subtraction, and over nine months by $1,350M. The fourth quarter is where the year’s spending lands: Intel’s 2026 gross capital spending “remain[s] on track to be $20 billion”, which on the same arithmetic leaves about $8.2B for the final quarter after $11.8B in the first nine months. For 2027, “gross capital spending” will be “up strongly”, with no number attached, and on the call the CFO said next year is “probably not a dynamic where we’re going to have positive free cash flow”. The commitments made since the quarter ended, the substrate agreements at $5.3–6.1B and the acquisition of Classiq, an Israeli quantum-software company, at about $1.9B payable in cash or shares, sit on top of that.
Will more shares be needed? Intel has not said, and what it states is this: $54.0B of cash and short-term investments, $15.0B of undrawn revolving credit, and “we remain committed to deleveraging in 2027”. The swing from net debt of $20.8B in June to net cash of $3.9B in September is mostly the August offering, about 242 million shares at $95.00 for $22.6B net, and the CFO gave the reason, speaking of the spending planned through 2027: “that is why we raised the money that we did at the beginning of August, was to get ourselves positioned to do this.” The fourth quarter carries a full quarter of those shares, a diluted count of 5.44 billion against 5.31 billion in the third.
Government and policy
The terms have not moved. The US government’s warrants over 241 million shares at $20.00 are stated after the August offering exactly as they were before it, and remain “neither currently nor expected to become exercisable”. They would adjust if Intel sold shares below 95% of a trailing average price, and the 10-Q discloses no adjustment, from which this note infers that August’s sale did not trigger one; Intel has not said so.
The accounting line swung the other way: the Escrowed Shares derivative, the mark-to-market of the shares Intel holds in escrow for the government, produced a $1.0B non-cash gain after a $12.5B loss in the second quarter, “driven by a decrease in our stock price”. It follows the share price rather than policy, so it does not set the row’s direction. The government’s percentage is not restated in the filing; the 9.9% figure is carried from the August 2025 agreement, and the 242 million new shares will have diluted it by an amount Intel has not stated.
Two watch points stay open. The Delaware stockholder suit seeking to invalidate the agreement had its motions to dismiss filed in May, and no ruling is disclosed. The second phase of the Section 232 semiconductor tariffs, the US national-security tariff provision, with its data-centre review, had not arrived by 23 October, when an industry group said the proclamation “has not yet been issued”, and nothing on the call touched it.
The watch list: Issue 02’s, marked, and the next
What Issue 02 said to watch
Issue 02 closed on eight things to watch, marked here against the record; the ones outside the chain this series now reads were not re-checked, and say so.
| Watched | Now | What happened |
|---|---|---|
| Q3 results | Resolved | Issue 02 pointed to late October as the first quarter that could show the guided second-half capex surge, and the next external foundry reading. Reported 29 October: gross capital spending $4,219M against Q2’s $2,652M; external foundry revenue $411M, 2.5% of total, “primarily” a non-recurring contract |
| 14A 0.9 PDK | Resolved | Due in October. Released in October “as committed”, to select customers |
| Customer window | Open | The second half of 2026 was the window the CEO pointed to for external commitments, with Crescent Island due in it. No customer named by Intel as at 30 October; “a couple of customers ready to announce in the next 2 quarters”, on the call. Crescent Island is now the Intel GPU D-series, “upcoming” in the release |
| 18A yields | Resolved | Industry-standard levels were due in early 2027. Intel’s remarks now say yields have risen “to industry standard levels ahead of our internal projections” |
| 14A dates | Open | Risk production in the second half of 2027, then the 2028 ramp. “Risk production next year” in the remarks; “begin production in ’28” on the call. Unchanged |
| Costa Rica | Open | Closure by the end of 2026. Not re-checked: outside the chain since this issue |
| Ireland and UMC | Developing | Fab 34 completing in 2027, and the UMC 12nm first production. Ireland: Intel 3 tool installation “progressing well”, per the remarks. UMC not re-checked |
| Six undated items | Open | An Ohio operating partner; a Brookfield buyout; Synopsys on 14A; the tariff review; Kulim’s date; Clearwater Forest benchmarks. Ohio: Intel is evaluating “opportunities to accelerate”, no partner named. The tariff review: Phase 2 not issued at 23 October. The other four not re-checked |
Issue 02’s list against the Q3 record. Items not re-checked are outside the chain this series now reads.
To the next issue
Scheduled or stated events only, never predicted outcomes.
Figure 5. Events to the next issue, in order, not to scale. Solid: scheduled, or a date Intel has stated. Hollow: a date only from the call, or none.
Sources and method
Sources, method, scope and what was not reached
Sources. Intel’s Q3 2026 earnings release (Form 8-K, 29 October 2026), the CEO and CFO prepared remarks and earnings deck of the same day, and the Q3 2026 Form 10-Q filed on 30 October 2026, read on Intel’s investor site because EDGAR could not be reached from the research environment. Issue 02’s readings are from the 8 August 2026 note and Intel’s Q2 filings; Fab 62’s status from an Intel newsroom piece of 21 October 2026; the Terafab items from Reuters, Tom’s Hardware and Fortune; the tariff status from CCIA; the deaths at Fab 28 from news-agency reports carried by Ynet News and Data Center Dynamics, Intel’s filing saying “employee casualties and injuries”. Intel had not published a transcript of the question-and-answer session when this issue went out, so its quotations from that session are from a third-party transcript.
Method. Each row’s direction is a fact against Issue 02, read a fixed way: node readiness by dated milestones met; customers by the count Intel has confirmed; capacity by Arizona assets placed in service; foundry revenue by the external figure and its share of the total, the loss reading up when it narrows; who pays by operating cash against gross capex and by cash against debt; government by the terms. A figure that is the firm’s arithmetic or an inference says so where it is used. The five figures are drawn from one data file in the research record; Figure 4’s axis is cut at $5,000M by decision, against the house rule of bars from zero.
Scope. From this issue Fab Watch narrows to Intel’s US fabs and the chain from customer to revenue. The whole-network register and the competitor charts Issue 02 carried are dropped; sites outside the US appear only where they bear on the chain, as Fab 28 does here.
Not reached. Intel’s Q2 10-Q figure for Altera’s wafer revenue was not on record, so the Q2 Altera figure is the nine-month figure less the first and third quarters.
What this note is. It presents developments and explains their mechanics. It does not characterise them as favourable or unfavourable, offers no view on the price or value of any security, and makes no investment recommendation.
Position disclosure. Spider Eyes Research and/or its principals hold a long position in Intel Corporation (INTC).
Important notice. This document is provided for information and discussion only. It is not investment advice, not a personal recommendation, and not an offer, invitation or inducement to engage in any investment activity. The firm is not authorised or regulated by the Financial Conduct Authority or any other regulator, and holds no licence to provide investment advice.
The firm issues no ratings, recommendations or price targets, and this note does not tell anyone to buy or sell anything. It is written once and published unchanged to everyone who receives it, takes no account of any reader’s circumstances, objectives or tolerance for risk, and creates no advisory relationship with any reader or subscriber.
Figures and estimates in this document are illustrative consequences of the assumptions stated alongside them. They are not forecasts, targets or predictions. Nothing here should be relied upon for any investment decision.
Information is drawn from public filings believed accurate at the date of writing. No representation is made as to accuracy or completeness, and no obligation is accepted to update it. The value of investments can fall as well as rise and you may get back less than you invest. Past performance is not a guide to future performance. Anyone considering an investment should carry out their own research and seek advice from a suitably qualified and regulated adviser.