Skip to content
primarysourced Photonics sector Lumentum
LITE
~5 min read · 1,066 words ·updated 2026-04-29 · confidence 54%

AI capex cycle and the optical-component pull

The single most decisive variable for the Lumentum thesis is the slope and durability of hyperscaler AI capex. Lumentum’s revenue beat its FY2024 trough ($1,359.2M) by 21% in FY2025 ($1,645.0M) and is on a run-rate that — at the Q2 FY2026 print of $665.5M and Q3 FY2026 guide midpoint of $805M — implies a $3.0B+ FY2026 with continuing acceleration into FY2027–FY2028. Nearly all of that growth lands in Cloud & Networking, and inside that segment the load-bearing component is the InP electro-absorption modulated laser (EML) at 200G/lane, the source-laser layer for 800G and 1.6T pluggable optical transceivers and (eventually) co-packaged optics.

Hyperscaler capex trajectory

The four US hyperscalers plus Oracle drive ≥80% of merchant datacenter optical-transceiver demand. Public capex disclosures (each company’s most recent 10-K / 10-Q / earnings releases — verified-primary ✓ for absolute figures, ⚠ for forward periods) have inflected sharply upward since the 2024 AI-cluster build-out began.

HyperscalerCY2023 capexCY2024 capexCY2025 capex (actual/guided)CY2026 trajectory (mgmt commentary)
Microsoft (incl. Azure)$28B$55.7B~$80B+Continuing build (no public guide)
Alphabet (Google)$32B$52B~$75B$90–100B+ implied
Amazon (AWS)$48B$77B~$100B+$130B+ implied
Meta$28B$39B~$70B$115B+ planned (per CY2026 guide)
Oracle$7B$21B~$25B+OpenAI $300B 5-year contract drives sustained build

Sources: each hyperscaler’s most recent 10-K and earnings releases. Aggregator estimates (Visible Alpha, Bloomberg consensus) for forward periods are ⚠.

The qualitative picture is unambiguous: cumulative hyperscaler capex 2024–2026 is north of $1T, with the marginal dollar increasingly directed to AI training and inference clusters, which are dramatically more optics-intensive per unit of spend than traditional cloud workloads. NVIDIA’s GB200 / GB300 NVL72 architectures, Meta’s Llama clusters, OpenAI / Oracle’s Stargate build-out, and xAI’s Colossus expansions all assume escalating bandwidth-per-GPU at 800G → 1.6T → 3.2T cadence on roughly the 2024 → 2026 → 2028 timeline.

Optical-component lead-time stretch

By Q2 FY2026, Lumentum management framed demand as booked through CY2027 and visible through 2028 for 200G/lane EMLs (Bloomberg April 10, 2026). Coherent Corp echoed the supply-constrained framing on its Q1 FY2026 call (datacenter revenue +23% YoY but “supply‑constrained by InP lasers”). Industry trade press (TrendForce December 2025) reported lead times for 200G/lane EML chips stretched to 40+ weeks across the merchant pool. ⚠ aggregator-level on the exact lead-time figure.

The supply-constrained backdrop is the structural underpinning of:

  1. Pricing discipline — analyst commentary points to double-digit ASP increases on 200G EML in CY2026, with the lack of a viable third source (NTT/Sumitomo do not service merchant datacom) sustaining the duopoly’s pricing power.
  2. NVDA’s strategic-supply concern — the March 2026 NVIDIA $2B Series A Convertible Preferred placement into Lumentum is a direct response to the supply tightness. The fact that NVIDIA committed balance-sheet capital — not just a long-term purchase order — is the strongest publicly-disclosed signal that the EML supply constraint is binding through 2028.
  3. Capacity investment scramble — both Lumentum (San Jose new fab + Towcester UK expansion) and Coherent (Sherman TX expansion) are deploying capex to multi-year lead-time fab capacity. The 200G/lane EML capacity build is dimensioned in 12–24 month construction-and-qual timelines, which means the supply tightness persists structurally for at least the next 6–8 quarters.

NVIDIA-direct demand pull on EML supply

The NVDA partnership (announced March 2, 2026) is the most concrete data point on this dynamic and is documented in Lumentum’s 8-K dated March 2, 2026 and the NVIDIA press release:

  • $2.0B equity investment via 2,876,415 shares of new Series A Convertible Preferred Stock at $695.31/share — a private-placement floor for NVDA’s effective entry point.
  • Multibillion-dollar purchase commitment from NVDA to Lumentum (exact total not publicly quantified — ⚠).
  • Future capacity access rights for advanced laser components dedicated to NVDA AI infrastructure.
  • Multiyear strategic agreements spanning R&D collaboration on next-generation AI optics (the language “state-of-the-art optics technology” implies CPO co-development).

NVDA simultaneously announced a parallel $2B Series A Convertible Preferred Stock investment in Coherent Corp (CNBC March 2, 2026) — the dual-source structure preserves duopoly economics rather than tilting NVDA toward a captive single-supplier strategy. The combined $4B is the largest publicly-disclosed AI-photonics supply-chain commitment to date.

The mechanic is that NVDA — by investing equity rather than only purchasing components — locks in capacity allocation ahead of competing hyperscalers. AWS, Azure, Google, and Meta all source 200G EML-based transceivers from the same merchant pool; if Lumentum and Coherent’s NVDA-allocated capacity is contractually carved out, the residual capacity available for the rest of the hyperscaler stack tightens further. This is the mechanism by which AI capex translates from a top-line demand variable into ASP and gross-margin uplift for LITE.

Sensitivity: what the bull case requires

The Lumentum bull thesis assumes hyperscaler AI capex grows ≥30% annually 2026 → 2028, with the optics-intensity-per-capex-dollar rising as the network shifts from 400G → 800G → 1.6T → CPO. A capex pause — driven by GPU oversupply, model commoditization, monetization plateau, or sovereign-AI funding pullback — would compress LITE revenue back toward the FY2024 trough faster than the EML capacity ramp could be re-allocated to non-AI applications.

The bear inflection-monitor signals are: (1) hyperscaler capex guides cut materially in Q3/Q4 CY2026 prints; (2) inventory-cycle warning signs in module pricing (CWDM4 / DR4 / FR4 ASPs); (3) NVDA GPU-shipment deceleration as a leading indicator. None of these have appeared as of April 2026; the inflection-monitor remains “no signal.”

Sources