◆ INVEST DEEPER / Company Research · Tower Semiconductor (TSEM) · Foundry
Thesis. Customers have wired $290M of prepayments to reserve silicon photonics capacity and signed $1.3B for 2027 — the purest scarcity signal on the watchlist — but a ~30% gross margin says Tower has not converted that scarcity into pricing power yet.
Reviewed 19 Sep 2026, after Q2 2026 results on 4 Aug and the August capacity and partnership news. Next review after Q3 2026 earnings on 10 Nov 2026.
What it does
Tower is a specialty foundry — a factory-for-hire that manufactures other companies' chip designs. Unlike TSM it does not chase the leading edge; it owns niche processes. One of those niches has become critical: silicon photonics (SiPho), building optical components on ordinary silicon wafers instead of the exotic indium phosphide laser makers use. It also makes image sensors, power management, MEMS and RF. Israel-headquartered.
When Coherent or Marvell builds a silicon-photonics transceiver, Tower often manufactures the chip. It sits directly underneath the Photonics category in this file.
One physical constraint shapes everything: silicon cannot emit light, so silicon photonics always requires an external laser. Tower captures the passive optical layer, not the scarce laser layer.
$290M of prepayments say Tower's capacity is scarce. A ~30% gross margin says its manufacturing is not. Which of those is true in 2027 decides this name.
Bull case
- Q2 2026 record revenue $460M, +24% YoY, at a record 30% gross margin. Gross profit $138M, operating profit $90M. Q3 guided to a record $520M, +31% YoY
- $1.3B of signed silicon photonics contracts for 2027 revenue — and $290M of customer prepayments to reserve capacity. Customers wiring cash a year ahead is the purest scarcity signal that exists; nobody prepays for something available on demand. Larger contracted wafer commitments for 2028, with further prepayments due January 2027
- 50+ active silicon photonics customers
- 2028 revenue target raised to $3.6B, up from the $2.8B model previously cited
- $3B Japan expansion announced July 2026 backed by ~$1B of Japanese government grants — the largest capital commitment in Tower's 30-year history. Converting the Arai facility in Niigata into a 12-inch SiPho and advanced packaging platform, expanding Uozu Fab 7 in Toyama, and preparing a new 12-inch fab. SiPho mass production targeted Q4 2027
- Expanded cooperation with Innolight, the world's largest optical transceiver maker, on next-generation mass production
- CPO foundry technology now offered on the SiPho and EIC platforms, using wafer-scale 3D-IC derived from years of stacked BSI sensor production, with full Cadence design tool support
Bear case
- Gross margin is only ~30% — foundry economics, not monopoly economics. Compare LITE, whose margins rose with scarcity; Tower's have not yet
- The capacity race is now explicit. UMC's Singapore 12-inch fab has shipped its first mass-produced SiPho wafers; TSM's COUPE platform is in production and scaling toward 25K wafers/month by 2028; GFS bought Advanced Micro Foundry to get there; China's CanSemi has a 12-inch SiPho line. Four-plus foundries building into the same demand is how the margin question resolves the wrong way
- GlobalFoundries has three active patent infringement claims against Tower over specialty process technologies, still pending
- Competes with TSMC, Intel, Samsung Foundry, UMC and AMF — and several customers, including Broadcom and Intel, have in-house silicon photonics capability
- The $3B Japan expansion is a large commitment with a 2027 completion date
- It trades with the AI-financing mood, not its own fundamentals — it fell 10% on 18 August, the largest single-day drop in the foundry group, on a WSJ report about ~$3 trillion of off-balance-sheet AI commitments across nine tech companies. Nothing company-specific
- Small relative to the foundries it competes against
Major customers
- NVDA — 1.6T silicon photonics built for NVIDIA networking protocols, with a claimed ~70% power saving versus electrical interconnects
- COHR — jointly demonstrated 400Gbps per lane using a silicon modulator, targeting 3.2T transceivers and co-packaged optics
- MRVL — multi-million-unit photonic chip shipments
- Innolight — the world's largest optical transceiver maker, on Tower's latest SiPho platform
- 50+ silicon photonics customers across the transceiver and optical component industry
- Broad specialty base in image sensors, RF, power management and MEMS across automotive, industrial and consumer
- IQE supplies indium phosphide epiwafers into the flow — a supplier, not a customer
Read-through. Tower is the manufacturing layer beneath the Photonics category, so its wafer volumes are a leading indicator for silicon-photonics transceivers from COHR, and its NVIDIA and MRVL relationships tie it directly to the AI interconnect roadmap. The structural point worth holding onto: silicon photonics is the technology that could eventually disintermediate indium-phosphide optics — but it always needs an external laser, which is exactly what LITE and COHR sell. So Tower's growth is not straightforwardly bearish for them; it changes what they sell rather than whether they sell. This also corroborates TSM's statement that lasers, optical fibres, fibre connectors and testing are the four remaining bottlenecks — Tower is building the part that is not on that list.
What would change the view
- Gross margin through 2027 as the $1.3B of contracts converts — the single most important unknown. Expansion means genuine pricing power; flat at ~30% means well-booked commodity manufacturing
- Prepayment balance and the January 2027 tranche
- Silicon photonics revenue disclosed as a standalone figure
- Japan expansion milestones toward Q4 2027 readiness
- GlobalFoundries patent litigation developments
- Whether it starts trading with COHR and LITE rather than with the foundry cycle — and whether AI-financing sentiment keeps overriding its fundamentals
Update log
19 Sep 2026 — Converted to the standard template. Corrected a stale target: the Backlog property still carried the $2.8B / $750M 2028 model, which was raised to $3.6B at the August update. Property rewritten. Also removed the "may belong in Photonics instead" rank note — the Stack Layer property reads Foundry and the read-through now explains the relationship to the photonics names directly.
31 Aug 2026 — Q2 detail and the capacity race documented. 2028 revenue target raised to $3.6B; Japan expansion firmed up; Innolight cooperation expanded; CPO foundry technology offered with Cadence support.
18 Aug 2026 — Fell 10%, the largest single-day drop in the foundry group (UMC −7%, GFS −7%), on a WSJ report about ~$3 trillion of off-balance-sheet AI commitments. Nothing company-specific; it was still up ~125% year to date, which is why profit-taking hit hardest.
4 Aug 2026 — Q2 2026 reported. Record revenue $460M (+24%) at a record 30% gross margin; gross profit $138M; operating profit $90M. Q3 guided to a record $520M (+31%).
Jul 2026 — $3B Japan expansion announced, backed by ~$1B of Japanese government grants.
Probabilities: strengthen ~55% / weaken ~30%.
Invest Deeper · Understand the business behind the stock · Research and education only, not investment advice.
Made with Bullet