Skip to content
primarysourced Photonics sector Coherent
COHR
~5 min read · 1,108 words ·updated 2026-04-29 · confidence 62%

AI capex cycle

The AI-datacenter capex super-cycle is the single load-bearing demand variable for Coherent’s Datacenter & Communications segment. The optical-interconnect layer sits at a structurally constrained chokepoint of that capex envelope: a hyperscaler can buy GPUs and DRAM on accelerated schedules, but InP EML wafers are a 4–6-month lead-time line item with merchant-supply concentration in two vendors. Coherent’s Q1 and Q2 FY2026 revenue trajectory — Datacenter & Communications +33.6% YoY in Q2 FY2026 — is best read as the supply curve catching up to a demand curve that was already inflected.

Hyperscaler capex trajectory

HyperscalerCY2024 capex (≈)CY2025 capex (≈)CY2026 guided (≈)Optics-relevant signal
Microsoft$55B$80B+$100B+First mover into 800G transceivers; Azure NVDA reservation visibility
Google$50B$75B+$90B+Internal Trillium/TPU + merchant NVIDIA; ROADM/wave-shaper procurement
Meta$38B$60B+$75B+NVIDIA H200/B200 deployment; AI Research SuperCluster expansion
Amazon AWS$77B$100B+$125B+Mix of NVIDIA + internal Trainium; merchant optics for both
Oracle$7B$25B+$35B+NVIDIA-Stargate joint-build; pure NVIDIA optics consumption
Top-5 aggregate~$227B~$340B~$425BOptical transceiver share ≈ 1–3% of total

Aggregator estimates from sell-side and trade press; ⚠ flag — exact 2026 capex guidance is moving target as quarterly disclosures update. The directional shape (≥30% YoY growth 2024 → 2026) is consistent across LightCounting, Dell’Oro, and Cignal AI.

Optical-component lead-time stretch

The structural symptom of supply tightness is lead-time extension. Coherent management commentary on the Q1 FY2026 call (November 2025) framed InP-laser capacity as the binding constraint:

Indium phosphide capacity, specifically EMLs, has been the primary supply constraint. We expect indium phosphide laser supply to increase from the current quarter into the next quarter and to continue improving sequentially throughout the next calendar year. — Q1 FY2026 earnings call, paraphrased from Motley Fool transcript Nov 6, 2025 ◐

The phrase book-to-bill above 4× in datacenter appeared in the Q2 FY2026 commentary (paraphrased), implying multi-quarter backlog visibility (Futurum Group Q2 FY2026 review) ◐.

Practical implications:

  1. ASP discipline holds — when capacity is sold out, new demand at the margin clears at higher pricing. Double-digit ASP increases on 200G EMLs in CY2026 are credible (per industry trade-press analyst commentary — ⚠).
  2. Customers prefer dual-source even at ASP premium — re-qualifying a non-incumbent EML supplier is an 18–24-month project; the cost of waiting on the duopoly is lower than the cost of qualifying a third source.
  3. Vertical-integration value rises — Coherent’s transceiver business captures more of the supply-constrained value-stack than a pure-component vendor.

NVDA-direct demand pull

The single most informative data point on AI-capex linkage to merchant optics is NVIDIA’s March 2, 2026 paired equity investment:

The bilateral structure — symmetric capacity dedication at both incumbents — is the most important signal of AI-capex shape:

  1. NVIDIA’s optical-supply view through 2028 is sufficiently capacity-constrained that NVIDIA chose to commit balance-sheet capital rather than rely on commercial-contract take-or-pay alone.
  2. NVIDIA’s CPO roadmap (Spectrum-X / Quantum-X Photonics) is dependent on merchant InP source-laser capacity that NVDA cannot easily build in-house on the relevant timeline.
  3. The duopoly-preserving structure tells against in-sourcing risk — a NVDA that planned to displace merchant optics by CY2028 would not simultaneously commit $4B of equity to merchant-vendor capacity expansion.

The instrument-design asymmetry (Coherent: common stock / immediate dilution; Lumentum: convertible preferred / HSR-gated) reflects relative bargaining position and balance-sheet considerations, not different strategic intent. See 03_ecosystem NVIDIA partnership for terms detail.

CY2026–CY2028 capex outlook

The base-case capacity-build path for the merchant duopoly:

PeriodCoherent actionLumentum actionIndustry effect
CY2025 H26-inch InP fab (Sherman TX) ramping; ~80% of target wafer-start rate (Coherent 6-inch InP press release) ✓Towcester UK + San Jose rampingCapacity 1.5–2× FY2024 baseline
CY2026 H1TX Semiconductor Innovation Fund $14M grant + $154M project (Semiconductor Today) ◐New NVDA-funded San Jose fab announced1.6T transceiver volume ramp
CY2026 H2Internal InP capacity targeted to double vs. mid-CY2025NVDA-funded new fab construction200G EML supply incrementally relaxed
CY2027Greenfield 6-inch capacity online; 1.6T module mainstreamLITE new fab partial productionPluggable-vs-CPO mix shift visible
CY2028CPO commercial-volume ramp at NVDA Spectrum-X / Quantum-X scaleSymmetric CPO rampPluggable+CPO TAM ≈ $40–60B (LightCounting baseline ⚠)

Demand sensitivity scenarios

ScenarioCY2025–28 capex CAGRMerchant optics rev CAGREML ASP path
Bear — AI capex digests10–15%10–15%Flat ASPs; no supply premium
Base25–30%25–35%Mid-single-digit ASP gains; duopoly discipline
Bull35%+35–50%Double-digit ASP premium; lead times stretch

The bull case assumes hyperscaler capex acceleration plus CPO transition adds (rather than replaces) chip-volume per unit of bandwidth — see CPO market for transition-architecture detail.

What would break the AI-capex cycle thesis

  1. Hyperscaler capex digestion — visible in ≥1 hyperscaler missing or cutting CY2026 guidance. Watch Microsoft/Meta/Google quarterly calls.
  2. AI workload unit-economics deterioration — if GPU FLOPs revenue stops scaling with capex, the merchant-optics demand growth flattens.
  3. Alternative-architecture displacement — if linear-pluggable optics (LPO) or onboard optics (OBO) prove out at lower cost, the EML-volume curve steepens less than expected.
  4. NVIDIA in-sources — least likely; the NVDA + Coherent + Lumentum partnership architecture suggests NVDA prefers merchant-supply continuity for CY2028 horizon.

Sources