Skip to content
primarysourced Photonics sector Tower
TSEM
~11 min read · 2,441 words ·updated 2026-04-29 · confidence 46%

Confidence legend: ✓ verified-primary (SEC filing, Tower 6-K / earnings release, hyperscaler 10-K / 10-Q, LightCounting / Yole / Dell’Oro press release) · ◐ partial / aggregator · ⚠ inferred / estimate.

This page sizes the AI-capex tailwind that flows through to Tower Semiconductor via three structurally distinct pull-through channels. Tower does NOT participate in the advanced-CMOS layer of AI compute (Nvidia / AMD GPUs and hyperscaler-custom Trainium / Maia / TPU / MTIA all sit at TSMC’s leading-edge nodes). Instead, Tower captures AI-capex spend at three adjacent layers: the optical-interconnect layer (PH18 SiPh + SiGe driver IC), the RF-infrastructure layer (5G + 6G + satellite buildout that supports AI-edge connectivity), and the power-management layer (BCD ASICs for hyperscaler PSU + optical-engine drivers).

The single most important framing fact: the FY 2025 RF Infrastructure surge from 17% to 27% of revenue (+73% YoY, +$178M absolute) is overwhelmingly AI-photonics-driven — silicon photonics alone added $122M YoY ($106M → $228M, +115%), with the SiGe driver-IC layer adding the residual. Approximately 97% of Tower’s total FY 2025 revenue growth ($+126M) is attributable to the AI-optical-interconnect stack (Q4 2025 earnings call slides ✓).

1. Hyperscaler CapEx Trajectory (Context)

The aggregate “big five” hyperscaler capex (Amazon, Alphabet, Microsoft, Meta, Oracle) trajectory is the upstream driver:

YearAggregate CapExYoY GrowthSource
2024~$256B—◐ CreditSights — Hyperscaler Capex 2026 Estimates
2025~$443B+73%◐ same
2026~$602–690B+36% to +56%◐ MUFG — Hyperscalers’ Capex Above $600B in 2026 ; ◐ Futurum — AI Capex 2026: $690B
2027⚠ projection range $700–900B+15–30% (slowing)⚠ analyst-class projection

Industry consensus is that ~75% of 2026 hyperscaler capex is AI infrastructure = ~$450B at the $602B aggregate level (or ~$520B at the $690B level). ✓ Goldman Sachs — Why AI Companies May Invest More than $500B in 2026

Individual hyperscaler 2026 capex disclosures

  • Amazon: ~$200B targeting 2026 capex.
  • Alphabet: ~$175–185B targeting 2026 capex.
  • Microsoft: ~$120B+ targeting 2026 capex.
  • Meta: ~$115–135B targeting 2026 capex (with Broadcom MTIA partnership announced 2026-04-14).
  • Oracle: ~$50B targeting 2026 capex.

⚠ Hyperscalers raised $108B in debt during 2025; multi-year projections suggest $1.5T in cumulative debt issuance over the AI-buildout cycle (MUFG — Financing the AI Supercycle). The financing-side stress is important for the Tower thesis because a hyperscaler-capex slowdown driven by debt-market or AI-ROI concerns would compress Tower’s RF Infrastructure demand 6–12 months later.

2. AI CapEx Decomposition — Where Tower Captures Spend

The $602B 2026 hyperscaler capex doesn’t flow uniformly. Approximate breakdown:

Layer% of total$ at $602BWhat it buysTower exposure
AI accelerators (GPU + custom silicon)~40%~$240BNvidia GPUs, AMD MI300/MI325, hyperscaler-custom (Trainium, Maia, TPU, MTIA)Zero direct (TSMC leading-edge only)
Networking (switches, NICs, transceivers, cabling)~15%~$90BSwitches, DACs, AECs, pluggable optical transceivers, CPO-equipped switchesDirect via SiPh + SiGe driver IC
Memory (HBM)~12%~$72BSK Hynix, Micron, Samsung HBM3E / HBM4Zero
Power infrastructure (UPS, PDUs, transformers, PSU)~10%~$60BSchneider, Eaton, Vertiv; rack-level PSU + optical-engine driversIndirect via BCD power management
Cooling (liquid + air)~7%~$42BDLC vendors, immersion coolingMinimal
Storage (SSD, etc.)~6%~$36B—Zero
Real estate + buildout / shell~10%~$60BDC real estate, fit-outZero

⚠ The ratios are analyst-class estimates synthesized from LightCounting / Dell’Oro / 650 Group decomposition reports plus hyperscaler-specific 10-K capex-line-item disclosures. Tower’s pull-through is concentrated in networking (the silicon-photonics layer) and to a lesser extent power infrastructure (BCD power management) and 5G/6G RF infrastructure for AI-edge connectivity.

3. Channel A — Optical Interconnect (the Dominant Pull-Through)

Optical-interconnect within networking spend

Of the ~$90B “networking” line item in 2026:

  • ~50% switching ASICs + line cards (~$45B): Broadcom Tomahawk, Cisco / Arista internal silicon, NVIDIA Spectrum, Marvell Teralynx.
  • ~30% optical transceivers and cables (~$27B): Pluggable 800G / 1.6T optics, AECs, DACs.
  • ~10% CPO modules and switches (~$9B at the early-2026 ramp).
  • ~10% other networking (~$9B): NICs, switch chassis.

Silicon-photonics share within optical interconnect

Of the ~$36B 2026 optical-interconnect total ($27B transceivers + $9B CPO):

  • ~30% is silicon-photonics-based ($11B). The rest is VCSEL multimode (low-end short-reach), externally-modulated lasers (EML, single-mode), and discrete passive optics. The silicon-photonics share is rising as 1.6T module shipments ramp.
  • ~10% of the SiPh module spend goes to silicon-foundry wafers and back-end (~$1.1B at the foundry-wafer layer in 2026). The bulk of the SiPh module ASP is value-add (DSP, packaging, optics, lasers, module-test, distribution).

⚠ Decomposition source: synthesis from Yole / Dell’Oro / LightCounting; specific allocations vary by analyst.

Tower’s capture of the SiPh foundry-wafer pool

If the silicon-photonics-foundry-wafer pool is approximately $1.1B in 2026:

  • Tower SiPh revenue FY 2025: $228M (Q4 2025 earnings call ✓) — ~20–25% of the (forward) 2026 foundry-wafer pool, assuming Tower captures incremental share through capacity expansion.
  • Tower SiPh capacity target Dec 2026: >5× FY 2025 actual wafer-starts — implying >$1B+ run-rate exit-2026 if utilization fills out at premium pricing. ⚠ Tower’s >5× capacity target is wafer-starts, not revenue; revenue ramp lags capacity by 6–12 months. (TipRanks ✓)
  • Tower SiPh capacity reservation status: Over 70% of total SiPh capacity reserved or in process of being reserved through 2028, with customer prepayments. ✓ (Q4 2025 release)

The capacity-prepayment-revenue conversion is the most visible-revenue-trajectory disclosure in the foundry industry. Most foundries don’t disclose customer prepayment levels at all; Tower’s 70%+ disclosed reservation rate at the Q4 2025 release implies an unusually high-conviction revenue-realization profile through FY 2028.

SiGe driver IC complementary revenue

Within RF Infrastructure (27% of FY 2025 = $423M), the SiGe driver-IC layer is the second component beyond pure SiPh wafers. SiGe BiCMOS driver ICs pair with SiPh dies to drive the high-speed modulators in 800G / 1.6T pluggable transceivers. The SiGe-only portion of the FY 2025 RF Infrastructure mix is implied at approximately $193M (from the $421M SiPh+SiGe combined disclosure minus the $228M SiPh-only disclosure). Tower management claims #1 position in SiGe and SiPho technologies for optical transceivers (Q3 2025 release ✓).

The SiGe driver-IC revenue is NOT a separate forecast bucket — it tracks SiPh wafer volumes 1-for-1 in the typical transceiver bill-of-materials. As Tower’s SiPh capacity ramps 5×, the SiGe driver-IC revenue ramps at a similar rate (subject to share-erosion if a SiGe driver IC could be replaced by CMOS-based driver ICs as data rates push higher). ⚠ The SiGe-driver-vs-CMOS-driver tradeoff at 200G/lane is a specific watch item — Coherent’s OFC 2026 silicon-MZM 400Gbps/lane demonstration (Tower-Coherent ✓) provides production-ready evidence that SiGe driver ICs remain capable at the 400G/lane envelope.

CPO ramp implications

Tower announced a CPO Foundry Technology program on 2025-11-12 (Tower-CPO release ✓), making Tower’s CPO heterogeneous-integration capacity available on the 200mm + 300mm SiPh + EIC platforms. This is a strategic response to the broader CPO transition:

  • 2024: ~$46M CPO market; essentially zero meaningful 2024 deployment.
  • 2025: First commercial CPO ramp (Broadcom Tomahawk-Ultra references; ~50,000 CPO switches shipped).
  • 2026–2027: Hyperscaler-scale CPO deployments expected.
  • 2030: $0.75B (Mordor 35% CAGR) to $8B+ (aggressive Mordor 137% CAGR) range.

For Tower, CPO is positive net-net because each CPO switch contains 8–16 SiPh dies vs ~1 SiPh die per equivalent pluggable transceiver. CPO ramp expands per-switch SiPh wafer demand at a rate that more than offsets pluggable-displacement risk.

4. Channel B — RF Infrastructure for AI-Ready 5G / 6G / LEO Satellite

The “RF Infrastructure” segment in Tower’s mix includes more than just SiPh + SiGe for optical interconnect. It also includes:

  • 5G base-station RF front-end wafers — SiGe BiCMOS for higher-frequency transceivers; SOI for power amplifiers at sub-6 GHz and mmWave bands. AI-driven 5G traffic (cloud-rendered AR, AI-assisted edge inference, hyperscaler edge-points-of-presence) creates incremental demand for 5G base-station wafer volumes.
  • LEO satellite payload silicon — SpaceX Starlink, Amazon Kuiper, OneWeb, Telesat all drive incremental SiGe / RF SOI / mixed-signal volumes for satellite phased-array beamforming silicon. Several of these programs are publicly tracked; Tower’s specific named-customer relationships in this segment are aggregator-level (⚠ not primary-source confirmed).
  • 6G research-tier silicon — research customers and early-commercial customers exploring 6G (above-100 GHz / sub-THz) communications use Tower’s SBC18S5 (250 GHz fT SiGe BiCMOS) + 65nm RF CMOS platforms. 6G commercial deployment is 2030+; current volumes are R&D-tier.

⚠ Quantitative breakout of “AI-adjacent vs base-cycle” within the non-photonics portion of RF Infrastructure is not publicly disclosed. Analyst-class estimate: the AI-photonics stack (SiPh + SiGe driver IC) accounts for ~85–90% of the FY 2025 RF Infrastructure revenue, with 5G / 6G / satellite RF accounting for the remaining 10–15%. The 5G / 6G / satellite component grows at a slower rate (single-digit YoY, ⚠) but provides cyclical-diversification buffer.

5. Channel C — BCD Power Management for Hyperscaler PSU + Optical-Engine Drivers

Tower’s Bipolar-CMOS-DMOS (BCD) power-management process portfolio at 0.18 µm and below — and the Tower-ST Agrate 300mm BCD capacity — supports:

  • Hyperscaler rack-level PSU silicon — power-management ASICs for AI-server PSUs (typically 48V / 12V conversion + intermediate bus voltage regulation). This is an indirect AI-capex pull-through; the BCD wafers supply ASIC fabless customers (TI, ON Semi, Infineon, ⚠ aggregator-class names; not primary-source confirmed) whose power-IC products end up in hyperscaler racks.
  • Optical-engine driver ICs — power management for the multiplied optical-engine count in CPO + on-board-optics deployments.
  • Smartphone PMIC + envelope tracker (the Apple iPhone-related design win ⚠) — separate handset-PA cycle, NOT directly AI-capex driven.

The BCD-AI-capex pull-through is structurally smaller than the SiPh / SiGe pull-through but provides additional revenue tailwind. The Power Management segment grew +20% YoY in FY 2025 (16% of revenue, ~$251M); the second-highest YoY growth rate after RF Infrastructure (segment revenue mix ✓).

⚠ Quantitative breakout of “AI-capex-adjacent BCD revenue vs handset-cycle BCD revenue” is not publicly disclosed; analyst-class estimate is that ~30–40% of BCD growth in FY 2025 is AI-server-related and ~60–70% is the Agrate handset-PMIC / automotive-48V cycle. Confirm in next 20-F MD&A.

6. AI-CapEx Risks for Tower

The bull-case relies on the AI-capex cycle running through 2027–2028 minimum. The bear-case scenarios:

Bear Case 1 — AI capex plateau in 2027

Hyperscaler ROI on the 2025–2026 AI-buildout disappoints; capex growth flatlines or contracts in 2027. SiPh wafer demand softens; the Dec 2026 5× capacity ramp under-utilizes; capacity-prepayment economics deteriorate. Tower SiPh revenue would still grow (the prepaid contracts are committed) but the post-2028 trajectory degrades. ⚠ The 70%+ prepayment disclosure provides material cushion against this scenario through 2028.

Bear Case 2 — Hyperscaler vertical integration

Nvidia / Broadcom / AMD vertically-integrate the optical engine layer (CPO with internally-fabricated SiPh at TSMC). Merchant photonics customers (Marvell, Coherent, Lumentum, Innolight) lose share to captive offerings. Tower SiPh wafer demand still grows because Broadcom’s CPO uses TSMC SiPh wafers for top-tier hyperscalers, but the merchant module-OEM tier compresses. The LWLG polymer-modulator integration (Tower-LWLG agreement, 2026-03-11 ✓) targets the next-gen 200G+ per-lane envelope which would extend Tower’s competitive position into Tier-1 hyperscaler displacement scenarios.

Bear Case 3 — China-Taiwan flashpoint

A military / political crisis in the Taiwan Strait disrupts TSMC’s leading-edge fabs. Short-term: panic-buying surge of Tower specialty wafers (positive). Medium-term: AI accelerator supply collapse since Nvidia / AMD are TSMC-dependent; AI capex craters; downstream optical / SiPh demand follows. Net effect: probably negative for Tower in a 12–24-month horizon despite short-term tailwind. Tower’s Israel-based footprint provides some geographic diversification but is itself exposed to regional geopolitical risk — see regulatory landscape for Israel-conflict-period disclosure considerations.

Bear Case 4 — Polymer / next-gen modulator displacement

If LWLG’s Perkinamine polymer modulators reach production faster on GFS Fotonix (live PDK as of 2026-03-16 GDSFactory PDK ✓) vs Tower PH18 (development agreement 2026-03-11; multiple engineering tape-outs during 2026 LWLG-Tower agreement ✓), the next-gen 200G/lane and 400G/lane sockets disproportionately route to GFS. This is a competitive-positioning risk; mitigated by Coherent’s OFC 2026 silicon-MZM 400G/lane demonstration on Tower’s production-ready PH18 (which suggests Tower can compete on silicon-only modulators at higher data rates than originally feared).

7. Reading the AI-CapEx Signals for Tower

For an analyst tracking Tower’s AI-photonics tailwind, the highest-value signals to monitor:

  • Hyperscaler quarterly capex prints (AWS, Microsoft, Google, Meta, Oracle Q4 2025 + Q1 2026 commentary) — direct leading indicator.
  • Nvidia / Broadcom / AMD AI accelerator unit-shipment commentary — sets the demand for optical interconnect at the rack-level.
  • Module-maker quarterly commentary (Coherent, Lumentum, Innolight, Eoptolink) — front-line view of hyperscaler optical-procurement trends.
  • CPO ramp commentary (Broadcom, Nvidia) — accelerating CPO timelines compress the merchant-pluggable monetization window but extend the foundry-wafer demand window for the CPO modules themselves.
  • Yole / LightCounting / Dell’Oro quarterly photonics-market updates — independent measurement of SiPh module / chip volumes.
  • Tower SiPh capacity-utilization commentary on quarterly earnings calls — the single most-revealing forward indicator. Q4 2025 disclosed >70% reservation through 2028; subsequent quarterly updates on the reservation rate provide direct visibility.
  • Tower SiPh design-win press releases — Coherent (OFC 2026 production-ready 400G/lane), Xscape (on-chip multi-wavelength laser), LWLG (Mar 2026 development agreement), OpenLight (PIC integration partner). Cadence of new design-win announcements = indirect demand signal.
  • GFS Fotonix design-win press releases — competitive-dynamic signal; LWLG PDK live at GFS as of Mar 2026 vs in-development at Tower.

The single most important leading indicator: Tower’s Q1 2026 / Q2 2026 SiPh revenue prints — whether the >$50M/quarter SiPh run-rate from H2 2025 sustains into the >5× capacity ramp through 2026. Currently markets are pricing in continued AI-capex acceleration; the first inflection-point disclosure (Q1 2026 release ⚠ scheduled May 2026) is the next thesis-validating event.

Cross-section pointers