BUSINESS OVERVIEW

Tower Semiconductor Ltd.

Nasdaq / TASE: TSEM

7 September 2026

Evidence base: Tower Semiconductor Forms 20-F for fiscal years 2016 through 2025 (most recent filed 30 April 2026); Forms 6-K containing the FY2025 annual and Q2/H1 2026 interim financial statements and management's discussion (2 March 2026 and 17 August 2026); Tower press releases and earnings-call remarks through August 2026. Industry structure, competitor economics and capacity data are drawn from independent sources — SEMI, TrendForce, and the reported filings of UMC, GlobalFoundries, X-FAB, Hua Hong, VIS, SMIC and SkyWater — and are attributed inline.

This document explains how the business works and where its economics come from. It is not a valuation and not a recommendation. It contains no price target, no multiple, and no view on the security.

1. Executive Snapshot

2. What the Company Does

A company that designs an analog or radio-frequency chip faces a problem a digital chip designer does not. Digital performance improves largely with transistor density, so a digital design can be ported to whichever foundry offers the smallest node. Analog performance depends on the physical behaviour of the devices themselves — noise, breakdown voltage, isolation, matching — and that behaviour is a property of the specific process that made them. The designer therefore needs a manufacturer that has built, characterised and qualified a process with those exact properties, and needs it to stay available for a decade. Tower Semiconductor sells that.

Tower is a pure-play foundry: it manufactures to customer designs and offers no products of its own (FY2025 20-F). It runs process geometries of 0.35, 0.18, 0.16 and 0.13 micron on 200mm wafers and 65 nanometer on 300mm wafers — nodes that are two decades behind the leading edge and deliberately so. What it sells is not smallness. It is a library of specialty platforms: silicon photonics, silicon-germanium BiCMOS, RF silicon-on-insulator, BCD and high-voltage power up to 700V, CMOS image sensors, MEMS and silicon backplanes for micro-displays.

The unit, and what happens to it

The unit of economics is one processed wafer. Tower states the pricing mechanism plainly: "We price our products on a per-wafer basis, taking into account the unique value of our technology, its ability to enable customers to differentiate their products, the complexity of the technology, prevailing market conditions, volume forecasts, the strength and history of our relationships with the customer and our current capacity utilization" (FY2025 20-F).

The path of one wafer runs as follows. A fabless company or IDM designs a device against Tower's process design kit — the device models and layout rules specific to that platform in that fab. Tower or a third party produces the photomasks. A raw silicon wafer then passes through several hundred sequential steps in one of six fabs: photolithography, oxidation, etch, ion implantation, thin-film deposition, chemical-mechanical polish, thermal processing. The finished wafer is probed to identify working dice and shipped to the customer, who almost always handles dicing, packaging and test elsewhere. Tower's revenue recognition ends at the wafer; the value added after it — packaging, test, the module, the system — belongs to somebody else.

Two features of that path govern the economics. First, the cycle from first customer contact to first shipment runs 9 to 24 months for a new customer and can exceed two years for a new technology, against 6 to 12 months for an existing one (FY2025 20-F). Revenue in any quarter was won two years earlier. Second, purchase orders are typically placed two to six months before shipment, so the order book is short even though the relationship is long — Tower has visibility on the customer, not on the quarter.

Who pays, and how concentrated they are

Customers are fabless analog and mixed-signal companies, IDMs using Tower as a second source or as outsourced capacity, and — a newer category the FY2025 20-F names explicitly — "module integrators for AI and data centers." Concentration has fallen sharply over the decade. In FY2016 four customers each contributed between 5% and 35% of revenue. In FY2025 the largest single customer, Nuvoton Technology Corporation Japan, was 11%; a further seven customers each contributed between 4% and 7%, for 39% together; and the remaining 50% came from many smaller accounts (FY2025 20-F). A business once anchored to one very large Japanese relationship is now spread across a wide book.

The geographic mix has moved with it. In FY2016, 49% of revenue was billed in the United States and 36% in Japan. In FY2025 the split was United States 42%, Asia excluding Japan 39%, Japan 13% and Europe 6% (FY2016 and FY2025 20-Fs). Japan has shrunk from over a third of revenue to an eighth, while non-Japan Asia has more than tripled its share.

Six fabs, and what changed at two of them

Tower processes wafers at six facilities: Fab 2 in Migdal Haemek, Israel; Fab 3 in Newport Beach, California; Fab 5 in Tonami and Fab 7 in Uozu, Japan, held through the 51%-owned TPSCo; Fab 9 in San Antonio, Texas; and Fab 10 in Agrate, Italy, where Tower has the right to use one third of the installed capacity in a 300mm cleanroom built and operated by STMicroelectronics (FY2025 20-F). Five are 200mm; Fab 7 and Fab 10 are 300mm at 65nm.

Two changes at the edges of that footprint explain more than any strategy statement. Fab 10 completed qualification during 2024 and began volume production in the fourth quarter of that year, so FY2025 was the first full year it contributed. In the same period, during the first quarter of 2025, Tower "discontinued the lower-margin legacy 150mm process flows, ceased operations in Fab 1, and consolidated certain flows into Fab 2" (FY2025 20-F). One 150mm fab of legacy low-margin work went out; one third of a 300mm cleanroom came in. That is the direction of travel stated in assets rather than in adjectives — and it is why FY2025 revenue growth and FY2025 margin moved in opposite directions.

A third change is scheduled. Under an agreement signed in March 2026, Tower will take full ownership of the 300mm Fab 7 while Nuvoton takes full ownership of the 200mm Fab 5, with mutual long-term supply agreements so that neither company's existing customers are disrupted. Nuvoton pays Tower $25 million on closing, targeted for 1 April 2027 (FY2025 20-F; Tower and Nuvoton joint release, 25 March 2026). Tower is trading its share of a 200mm fab for sole control of a 300mm one, and simultaneously converting its largest customer relationship into a supply contract.

3. Industry, Competitive Position and Moat

The specialty foundry industry exists because the reasons to outsource manufacturing are different at mature nodes than at the leading edge. At the leading edge, the product is the node: a fabless company outsources because it cannot afford a $20 billion fab. At mature nodes, the product is a qualified process: a company outsources because building and characterising a 700V high-voltage platform, or a silicon photonics platform with integrated modulators and photodiodes, takes years of process engineering that has nothing to do with lithography scaling.

Where the profit actually sits

It does not sit with the foundry. Reported fiscal 2025 gross margins across the specialty and mature-node group run from 11.8% at Hua Hong to 29.0% at UMC, with GlobalFoundries at 24.9%, X-FAB at 21.2%, SkyWater at 19.7% and SMIC at 19.2% in the fourth quarter; Vanguard International Semiconductor reported 27.1% for FY2024 (company results releases, January–February 2026; SMIC via TrendForce, 11 February 2026; VIS FY2024 annual report). Operating margins across the same group run from 18.5% at UMC and 16.0% at Vanguard down to 11.7% at GlobalFoundries, 8.8% at X-FAB and an operating loss at SkyWater and Hua Hong, against Tower's 12.4%. Tower's FY2025 gross margin of 23.2% sits in the middle of that range. For comparison, Texas Instruments — an analog designer that owns its own fabs — reported a 57.0% gross margin and a 34.1% operating margin for FY2025 (TI results release, 27 January 2026), and the outsourced assembly and test houses that sit downstream of the foundry reported 17.7% at ASE and 14.0% at Amkor. The specialty foundry captures a manufacturing margin roughly in line with packaging, well below the margin earned by the people who own the circuit design.

This is the single most important structural fact about Tower's industry, and it explains the shape of everything downstream. A foundry does not own the end product, cannot price against the value the chip creates in a phone or a switch, and prices instead against the cost of capacity — its own and everyone else's. Differentiation in this industry is not a route to designer-like margins. It is a route to being paid a manufacturing margin reliably instead of cyclically.

Structure, and what market share means here

On the standard measure Tower is small. In the first quarter of 2026 the global foundry market was $47.95 billion, of which TSMC took 72%, Samsung 6.5%, SMIC 5.1%, UMC 3.9%, GlobalFoundries 3.3% and Hua Hong 2.5%. Tower ranked seventh at $414 million, or 0.8% (TrendForce, 12 June 2026).

That number should be read with care, because it is not a share of a market Tower competes in. TSMC's 72% is overwhelmingly leading-edge AI and HPC logic — demand Tower never bids for and could not serve. The relevant market is segmented by process platform and by qualification relationship, not by wafers in aggregate. A customer choosing a 700V high-voltage platform or a 1.6 terabit-per-second silicon photonics platform is choosing among a handful of suppliers who have that platform qualified, not among everyone who owns a fab. Tower names its own direct competitors as GlobalFoundries, chiefly in RF, plus Vanguard, DongBu, X-FAB and Hua Hong, with TSMC, UMC and SMIC competing in certain areas (FY2025 20-F). That is a defensible characterisation: those are the companies with comparable specialty platform depth.

Which barriers actually bind

Capital is not the binding barrier at mature nodes. A new 200mm fab costs between roughly $450 million and $1.3 billion depending on the process, against more than $20 billion for a leading-edge facility (SurplusGlobal data reported by Semiconductor Engineering). SMIC alone spent $8.1 billion of capital expenditure in FY2025 and guided to roughly the same for 2026 (TrendForce, 11 February 2026). Money can build mature-node capacity, and in China it is doing so.

Two barriers do bind. The first is equipment. Fewer than 250 used 200mm core tools are available worldwide against an estimated need of 1,500 to 3,000, and buyers should expect to pay close to new-tool prices for used ones (SurplusGlobal, reported by Semiconductor Engineering). An installed base of qualified 200mm tooling is genuinely hard to reproduce quickly. The second is qualification. Tower's own account is that "customers that use our specialty process technologies cannot easily transfer designs to another foundry because the analog characteristics of the design are dependent upon the specific process technology used" and that the design kits and device models are specific to the fab in which the process is implemented (FY2025 20-F).

The second claim is supported by Tower's own operating evidence rather than merely asserted. Sales cycles running two years or more, purchase orders placed only two to six months ahead of shipment, and a customer book that has stayed intact while its largest relationship fell from 35% to 11% of revenue are all consistent with customers who stay because moving is expensive. What that stickiness has not bought is pricing power: gross margin averaged roughly 23% across FY2016 to FY2025 without trending upward. The switching cost defends the relationship. It does not lift the price.

What the outside evidence supports, partially supports, and contradicts

Company claims are quoted from the FY2025 20-F. Verdicts rest on the independent sources named in the basis column.

What kind of company wins here

The evidence points to one answer: the foundry whose revenue cannot be treated as a fungible mature wafer. Hua Hong ran above 100% of rated capacity in FY2025 and still earned an 11.8% gross margin and a net loss, because near-full utilization of commodity-adjacent capacity does not confer pricing power. UMC, the fourth-largest foundry in the world, watched gross margin fall from 45.1% in 2022 to 27.7% in the first half of 2025 as its 22 and 28 nanometer book — 37% of revenue — met Chinese capacity on price (TrendForce, October 2025). Scale did not protect it.

Tower is the right kind of company by construction: it has no commodity digital book to defend, and its 300mm capacity is 65nm BCD and RF SOI rather than a node anyone competes on. Whether it is a good version of that kind of company has, until very recently, been an open question — because for nine years the differentiated platform book produced no margin expansion at all. The silicon photonics ramp is the first evidence that a platform position can translate into price.

4. Growth Engine

Tower has made no material acquisitions during the period reviewed, so reported growth is organic growth. What complicates the comparison instead is capacity moving in and out of the consolidated group — and the filings decompose the two most recent periods differently, which is the most useful thing in them.

Two consecutive periods, two different engines

FY2025 revenue rose $130.0 million, or 9.0%, to $1,566.1 million. Management attributes it "mostly due to higher wafers shipments, including revenue and shipments from the 300mm facility in Agrate, Italy, which commenced volume production and operations in the fourth quarter of 2024, and therefore its revenue was included in our statement of operations for the entire 2025 calendar year, as compared to only a limited portion of 2024" (FY2025 20-F). That is volume growth, and a material part of it is the annualisation of newly consolidated capacity rather than more demand at existing fabs. Working against it, Fab 1's 150mm output left the group in the first quarter of 2025. Tower does not quantify either effect separately; the Agrate contribution to FY2025 revenue is not disclosed.

The first half of 2026 was different in kind. Revenue rose $143.5 million, or 19.6%, to $873.7 million, and management attributes the increase "mostly due to a higher average selling price associated with a different product mix shipped in the six months ended June 30, 2026 as compared to the six months ended June 30, 2025 (especially more SiPho related products shipped), as well as a higher quantity of products shipped from our fabs" (6-K, 17 August 2026). Price and mix led; volume followed. Gross margin moved from 21.0% to 28.5% across the same comparison — a 7.5 point swing that volume growth alone could not produce.

That is the decomposition that matters. FY2025 was capacity growth at unchanged economics: revenue up 9%, gross margin down from 23.6% to 23.2%. H1 2026 was mix growth: revenue up 19.6% and gross margin up 7.5 points. The business changed engines between the two periods.

The drivers, ranked

• Silicon photonics for AI optical interconnect — structural, and by a wide margin the largest. The annualised run-rate went from $180 million in the second quarter of 2025 to $680 million in the second quarter of 2026, and management targets more than $1 billion annualised by the fourth quarter of 2026 (Q2 2026 release and call, August 2026). In May 2026 Tower disclosed $1.3 billion of signed customer contracts for 2027 silicon photonics revenue, with $290 million of prepayments received for capacity reservation and more than 50 active silicon photonics customers (Tower release, 13 May 2026). This is a genuine platform position: Tower runs a 200mm platform and a 300mm platform in high-volume production across 400Gb/s to 1.6Tb/s, is ramping 1.6Tb/s, and is developing 3.2Tb/s and co-packaged optics (FY2025 20-F).
• Capacity additions converting backlog into shipments — management-driven. The $920 million programme across Fabs 2, 3, 7 and 9 targets December 2026 silicon-photonics wafer-start capacity more than five times the fourth quarter 2025 shipment level (FY2025 20-F; Q4 2025 release). Roughly half had been paid by the second quarter of 2026. Growth here is a function of installation and qualification schedules, not of demand.
• Mix shift away from legacy — management-driven. Closing Fab 1's 150mm flows removed revenue and raised the average margin of what remained. Tower has been trading volume for margin at the low end.
• RF infrastructure, power management and image sensors — cyclical. On the first-quarter 2026 call management cited year-on-year growth of 12% in RF SOI, 10% in power management, 24% in SiGe and 9% in CMOS imagers — recovery-rate growth off a corrected base, not step changes.
• RF mobile — cyclical, and currently negative. Down 14% year on year in the second quarter of 2026 and now 12% of revenue (Q2 2026 call). Smartphone RF, once a core Tower franchise, is shrinking as a share of the business while infrastructure grows.

What the company says the engine is worth

Management has published three successive capacity-and-earnings frames, and the sequence is informative. In November 2025 the chief financial officer described $2.7 billion of annual revenue and roughly $500 million of net profit at full loading of existing, already-qualified capacity. In May 2026 a 2028 model of $2.8 billion of revenue and $750 million of net profit was set. In July 2026, alongside the Japan expansion announcement, that model was raised to $3.6 billion of revenue and $1.2 billion of net profit, described as fully spoken for by customers, with a 45% gross margin assumption. These are company targets, not results, and the first is a different quantity from the other two — a ceiling on existing capacity rather than a forecast for expanded capacity. They are recorded here because the gap between the FY2025 outcome and the 2028 frame is entirely a function of capacity that is not yet built.

5. Margin, Cash and Capital Allocation

Why margin moves the way it does

Tower states the mechanism directly: "As is common in our industry, a large portion of our total cost is comprised of fixed costs. Therefore, while during periods in which we operate at high utilization rates we are able to cover our costs, at times when the utilization rate is low, the reduced revenues may not cover all of the costs since a large portion are fixed costs which remain constant, irrespective of our capacity utilization" (FY2025 20-F). Depreciation and amortisation ran $166.5 million in the first half of 2026, or 19.1% of revenue. Almost a fifth of the cost base does not move with volume at all.

The filing adds the qualification that matters most for the next two years: "This issue may be further exacerbated with the ramp-up of new capacity, where we incur fixed costs upon the start of operations while gradually ramping up utilization, causing a decrease in our profit margins." Tower is mid-way through a $920 million equipment programme and has announced roughly $3 billion more in Japan. Each tranche of capacity depresses margin before it lifts it, and the depression arrives on a schedule Tower controls while the lift arrives on a schedule its customers control.

Operating leverage runs in both directions and the recent record shows the range. Between the first half of 2025 and the first half of 2026, revenue rose 19.6% and gross profit rose 62.4% — $143.5 million of additional revenue produced $95.5 million of additional gross profit, an incremental gross margin of 67%. Below the gross line the leverage is milder because research and development has been held at a constant 5.4% of revenue and selling and administrative expense at 5.3–5.4%, so operating profit rose 113% on that 19.6% of revenue growth.

The financial spine

Sources: FY2016, FY2022 and FY2025 Forms 20-F; 6-K of 17 August 2026. Comparability: FY2016 includes Fab 1 (150mm) and three Japanese fabs including Arai, and FY2016 net profit includes a $50.5 million net gain on the San Antonio fab acquisition. Fab 10 in Agrate entered the consolidated results only from the fourth quarter of 2024 and Fab 1 ceased operations in the first quarter of 2025, so FY2025 is not comparable to FY2016 or FY2022 on a like-for-like asset base. FY2023 is omitted from the table because its 38.5% operating margin includes a $313.5 million net Intel merger-termination fee and a $52.2 million restructuring gain; the underlying operating margin that year was in line with FY2024. H1 2026 is a six-month period and its capex ratio is not annualised. Margin percentages and the capex ratio are computed from the disclosed dollar figures.

Cash conversion

Reported profit and cash have diverged in both directions, and for a reason worth separating. In FY2025, operating cash flow of $395.5 million against net capital expenditure of $436.5 million produced negative free cash flow of $41.1 million — the business generated cash and spent more than all of it on capacity. In the first half of 2026, operating cash flow was $686.9 million against $343.0 million of capital expenditure, apparently transforming the picture. But $282.6 million of that operating cash flow was an increase in customers' advances, disclosed as such in a footnote to the cash flow statement (6-K, 17 August 2026). Excluding it, operating cash flow was $404.4 million and free cash flow $61.4 million.

Those advances are the silicon photonics capacity prepayments, and the balance sheet shows their scale. Deferred revenue and customers' advances, current and long-term combined, went from $27.5 million at 31 December 2025 to $321.4 million at 30 June 2026. This is customer money, not shareholder money, and it will be credited against future purchases rather than repaid — which means it converts into revenue at a zero incremental cash margin when the wafers ship. It is a genuinely favourable funding structure and it is not free cash flow.

Where the cash has gone, ranked

• Capital expenditure, by an enormous margin. Roughly $2.9 billion of net capital expenditure across FY2016 to H1 2026, running above operating cash flow in the most recent full year.
• Debt reduction. Total debt was $161.5 million at the end of FY2025 and $141.7 million at 30 June 2026, against $1,481.4 million of cash and short-term deposits — a net cash position of $1,339.7 million. Tower carries essentially no financial leverage.
• Dividends and buybacks: none. No dividend and no repurchase programme appears in any filing or release reviewed across the period. Shares outstanding were 113.0 million at 30 June 2026 against 114.4 million on a diluted basis, and the count has been broadly flat — the company has neither returned capital nor funded itself by issuing equity.

The pattern is unusually consistent for a company of this size: everything the business earns, plus the accumulated cash, is directed at capacity, and nothing is returned. Three non-shareholder sources are being used alongside it — $290 million and rising of customer prepayments, $1 billion of Japanese government METI grants against the announced $3 billion Japan programme, and a $25 million payment from Nuvoton on the TPSCo closing. Management is deliberately structuring the expansion so that shareholders fund the smaller share of it.

Announced but not in the reported figures

On 14 July 2026 Tower announced a Japanese expansion of approximately $3 billion supported by roughly $1 billion of METI grants, leaving a net cost of about $2 billion. It has two tracks: converting the Arai facility to 300mm silicon photonics and advanced packaging with production readiness targeted for the fourth quarter of 2027, and a greenfield 300mm facility adjacent to Fab 7 contributing revenue from 2029, contingent on further approvals. The 2028 target model was raised to $3.6 billion of revenue and $1.2 billion of net profit on the same day. None of this appears in the FY2025 20-F or in the H1 2026 accounts, and the second track is explicitly conditional. The TPSCo restructuring, targeted to close on 1 April 2027, likewise post-dates every reported figure in this document.

One item cuts the other way. The 9% effective tax rate Tower paid in FY2025 rested on a 7.5% preferred Israeli rate applied to eligible Israeli income "through December 31, 2025" (FY2025 20-F). The first half of 2026 was taxed at 12.7%. The low-tax period as previously constituted has ended, and the filing does not state what replaces it.

6. Cyclicality, Constraints and What to Monitor

Tower's own description is the correct starting point: the semiconductor industry "fluctuates over time, cycling through periods of weak demand, excess capacity, excess inventory, and price pressure, as well as periods of strong demand, full capacity utilization, and wafer shortages, which command higher selling prices" (FY2025 20-F). The ten-year record shows exactly that shape and shows how deep it runs.

Where the business sits against its own peaks and troughs

Revenue peaked at $1,677.6 million in FY2022, fell 15.2% to $1,422.7 million in FY2023, and had recovered only to $1,566.1 million by FY2025 — still 6.6% below the prior peak three years after it. Gross margin traced a wider arc: 27.8% at the FY2022 peak, a trough of 18.4% in FY2020 and 18.6% in FY2019, and 23.2% in FY2025.

The full ten-year series, taken from the Forms 20-F for FY2016 through FY2025, runs: FY2016 $1,249.6 million at a 24.3% gross margin; FY2017 $1,387 million at 25.5%; FY2018 $1,304 million at 22.5%; FY2019 $1,234.0 million at 18.6%; FY2020 $1,265.7 million at 18.4%; FY2021 $1,508.2 million at 21.8%; FY2022 $1,677.6 million at 27.8%; FY2023 $1,422.7 million at 24.8%; FY2024 $1,436.1 million at 23.6%; FY2025 $1,566.1 million at 23.2%. Nine years produced 25% revenue growth and no margin expansion whatsoever. That is the base against which the last six quarters should be read.

Those six quarters run: first quarter 2025 $358.2 million at a 20.4% gross margin; second quarter 2025 $372.1 million at 21.5%; third quarter 2025 approximately $395.7 million at approximately 23.6%, derived by subtracting the reported half-year and fourth-quarter figures from the reported full-year totals; fourth quarter 2025 $440.2 million at 26.7%; first quarter 2026 $414.0 million at 26.8%; second quarter 2026 $460.1 million at 30.0%. Revenue rose 28% across the six and gross margin rose 9.6 points. The second quarter of 2026 produced $90.8 million of net profit attributable to the company and $0.79 of diluted earnings per share on 114.4 million diluted shares. Guidance for the third quarter of 2026 is $520 million, plus or minus 5%.

The current position is unusual and worth stating precisely, because it is the fact most likely to be misread. On volume, Tower is only now clearing its prior peak: the first half of 2026 ran at $873.7 million, an annualised pace above FY2022 but achieved with Fab 10 added and Fab 1 removed. On margin, it has already cleared it: the second-quarter 2026 gross margin of 30.0% is above the FY2022 cycle peak of 27.8% and the highest in the period reviewed. Margin is running ahead of volume. That is the signature of a mix change, not of a cyclical upswing — a record margin at a mid-cycle volume is a different fact from a record margin at a volume peak, and this is the former.

Utilization confirms there is room left. On the second-quarter 2026 call management put Fabs 2, 3 and 9 at 80–85%, Fab 5 at 75%, and Fab 7 well above its 85% target model. Full loading of existing qualified capacity is the $2.7 billion revenue frame management described in November 2025, against $1,566.1 million actually delivered in FY2025.

End-market exposure

Source: Tower second-quarter 2026 earnings call, August 2026. Tower does not publish an end-market split in its financial statements and does not disclose profitability by end market, so these shares are revenue only and are not audited segment data. The residual is derived by subtraction.

The downside mechanism

The relevant downside for Tower is not a general semiconductor recession. It is a pause in AI data-centre optical spending arriving while the capacity built to serve it is being switched on. Silicon photonics moved from roughly 12% of a $1.5 billion revenue base to a $680 million annualised run-rate in four quarters, and 49% of second-quarter 2026 revenue now sits in the RF infrastructure line that contains it. The $920 million equipment programme and the $3 billion Japan programme are both sized against that demand, and the fixed cost they create arrives on installation, not on shipment. A customer base that deconcentrated impressively at the account level has re-concentrated at the application level.

The industry backdrop offers no clear read on the other side of the business. Mature-node pricing signals reversed within a single year — UMC pressed its own suppliers for a 15% price cut in October 2025 to defend a gross margin that had fallen from 45.1% in 2022 to 27.7% in the first half of 2025, and by March and April 2026 SMIC, Hua Hong, Nexchip and UMC were all raising mature-node prices by 4% to 20% as 8-inch supply tightened and TSMC and Samsung retired legacy lines (TrendForce, October 2025 and March–April 2026). Whether that tightening is durable or a spike ahead of the next Chinese capacity wave is genuinely unresolved on the evidence available.

Durable versus borrowed

Leading indicators, and where each is published

• Gross margin against the 45% target, quarterly — Tower results releases and 6-K filings.
• Silicon photonics annualised run-rate against the stated $1 billion fourth-quarter 2026 goal — Tower earnings calls, which are where this is disclosed; it does not appear in the financial statements.
• Deferred revenue and customers' advances on the balance sheet — 6-K interim statements. A stalling balance would signal that 2027 and 2028 capacity reservations have stopped growing.
• Capital expenditure against the $920 million programme and the $3 billion Japan programme — Tower cash flow statements and releases.
• Mature-node foundry utilization and wafer pricing at UMC, SMIC, Hua Hong and VIS — company quarterly releases; UMC, SMIC and VIS disclose a utilization rate, GlobalFoundries and X-FAB do not. TrendForce publishes the quarterly foundry share and pricing series.
• 200mm capacity additions — SEMI World Fab Forecast and 200mm Fab Outlook.
• The GlobalFoundries ITC investigation docket — US International Trade Commission. An exclusion order, unlike a damages award, would take effect as an import ban.

7. Risks, Unknowns and Questions for Deeper Work

The risks below are those cyclicality does not capture. Cycle exposure and the AI-spending downside are covered in Section 6 and are not repeated.

• An ITC exclusion order would be an import ban, not a damages award. On 26 March 2026 GlobalFoundries filed three infringement actions against Tower — at the US International Trade Commission and in the Western District of Texas — asserting eleven US patents and seeking to block Tower's US import and sale of allegedly infringing products as well as lost-profits damages. Tower "disputes these claims" (FY2025 20-F; GlobalFoundries release, 26 March 2026). The mechanism matters: the ITC's remedy is exclusion. An adverse determination would sever US-bound shipments from Tower's Israeli, Japanese and Italian fabs at the point in the ramp when US-billed revenue is 42% of the total and the silicon photonics contracts are being fulfilled. Tower has disclosed no reserve and no assessment of likely outcome.
• The second Intel relationship has failed, and the first one produced the cash pile. Under a September 2023 agreement Intel undertook to build a capacity corridor for Tower's customers at its 300mm New Mexico facility, against up to $300 million of Tower investment. "Intel expressed its intention not to perform under the agreement, and the parties are presently in a mediation process" (FY2025 20-F). This follows the collapse of Intel's $5.4 billion acquisition of Tower in August 2023 for want of Chinese regulatory approval, which produced the $313.5 million net termination fee recorded in FY2023 operating profit. The consequence is not the mediation outcome but the capacity: 300mm US capacity Tower had planned around must now come from elsewhere, and the Japan programme is where it is coming from — at Tower's cost rather than Intel's, and two years later.
• Regional conflict is already touching the capital programme, not just the risk factors. The FY2025 20-F states that delays in equipment delivery, installation and qualification "may also result from the recent conflicts between the United States and Iran, Israel and Iran, and Israel and Hezbollah that commenced in February 2026, which has prevented vendors from arriving in Israel to install the equipment, which may impact the timely execution of the $920 million SiPho and SiGE capital expenditure plan." This is a stated operational effect on a named programme, not a generic geopolitical caveat. No financial impact has been quantified, and the subsequent quarterly filings do not return to it.
• The Newport Beach lease is disputed at the same site that runs silicon photonics. Fab 3's landlord has requested a judicial declaration that the company has committed a material, non-curable breach of the lease over noise abatement, which Tower disputes; separately a third party asserts collateral or other rights over the site. The lease was extended in 2025 through 2030 (FY2025 20-F). Fab 3 is one of Tower's silicon photonics fabs, and a 200mm specialty fab cannot be relocated on a commercial timescale.
• Prepayments are a claim on future margin, not a cushion. The $321.4 million of customer advances on the balance sheet at 30 June 2026 will be credited against future wafer purchases. When those wafers ship, they generate revenue and cost but no incremental cash. The cash benefit is being taken now; the margin obligation lands in 2027 and 2028.
• The two risks that compound. An ITC exclusion order arriving during the capacity ramp would combine the worst version of two separate problems: fixed costs from newly installed equipment landing in the income statement while the US route to market — 42% of revenue — is closed. Each is survivable alone. Together they would strand the capital programme.

What the sources could not answer

• Wafer volumes and average selling price. Tower disclosed both in the FY2016 20-F — "a 27% increase in the number of wafers shipped and a 6% increase in the average selling price per wafer" — and has not disclosed either since. The unit economics of a per-wafer pricing business therefore cannot be reconstructed from the filings, and price and volume cannot be separated except where a narrative sentence happens to attribute them.
• The maintenance versus growth split of capital expenditure. Not disclosed in any year, which makes it impossible to establish what Tower must spend simply to keep six fabs qualified.
• Profitability by end market, by platform or by fab. Tower reports as a single segment. The margin on a silicon photonics wafer relative to a power management wafer — the central question raised by the H1 2026 mix shift — is not disclosed anywhere.
• Fab 10 and the Agrate arrangement. Tower's committed capital, the capacity actually secured and the current ramp status do not appear in any 2025 or 2026 disclosure reviewed, despite Fab 10 being the largest single contributor to FY2025 revenue growth. Nor is total company capital expenditure guidance for 2026 given.
• The pro-forma effect of the TPSCo restructuring. Nuvoton was 11% of FY2025 revenue and Fab 5's output leaves the group on closing. Neither the revenue nor the margin effect is quantified.
• Named silicon photonics customers. Consistent with long-standing practice, none is disclosed — so the concentration inside the 49% RF infrastructure line, and inside the $1.3 billion of 2027 contracts, cannot be assessed.

Three questions follow for anyone doing deeper work. What is the gross margin on a silicon photonics wafer, and does the 45% company target imply a mix in which photonics is most of the book? How many customers stand behind the $1.3 billion of 2027 contracts, and what happens to the capacity if the largest one re-plans? And what does Tower's Israeli tax position look like from 2026 onward, given the preferred rate as described applied through the end of 2025?

8. Investor Takeaways

• What this business really is. A contract manufacturer of qualified analog process platforms — small in the foundry industry at 0.8% of wafer revenue, but genuinely differentiated within the handful of segments it serves, and structurally unable to earn designer-like margins.
• The core economic engine. Fab utilization against a fixed cost base of which depreciation alone is 19% of revenue. Incremental gross margin ran 67% in the most recent half-year comparison, which is what makes the mix shift so visible so quickly.
• The main growth lever. Silicon photonics: from a $180 million annualised run-rate to $680 million in four quarters, with $1.3 billion of 2027 revenue under signed contract. It is why gross margin cleared its ten-year peak while revenue has only just cleared its FY2022 level.
• What could break the story. A pause in AI optical spending landing while $920 million of installed equipment and a $3 billion Japan programme are switching on — or an ITC exclusion order closing the US, which is 42% of revenue.
• What to monitor. Gross margin against the 45% target, the silicon photonics run-rate, and the customers' advances line on the balance sheet. The third is the earliest of the three: capacity reservations are booked before the wafers are made.