Co-packaged optics (CPO) market

Co-packaged optics is the architecture inflection that sits at the 2027–2028 boundary for AI-cluster networking. CPO replaces front-panel pluggable transceivers with optical engines mounted in-package alongside the network ASIC (switch silicon), eliminating the long electrical traces that drive most of the power and signal-integrity overhead in 1.6T-and-beyond pluggable-module designs. For Lumentum, CPO is both an opportunity (the 200G/lane EML chip remains the source-laser layer in CPO modules, as in pluggables) and a risk (CPO disrupts the Cloud Light pluggable-transceiver assembly business).

NVIDIA’s CPO roadmap (the load-bearing reference)

NVIDIA’s CPO roadmap is the sequence the rest of the industry calibrates against, given NVDA’s market position in AI-training silicon. The publicly-disclosed milestones (NVIDIA GTC March 2025 Spectrum-X / Quantum-X Photonics announcement and NVIDIA Technical Blog):

Product	Architecture	Timing	Application
Quantum-X Photonics	InfiniBand CPO switch	Late 2025 / early 2026	Scale-out InfiniBand fabric in AI clusters
Spectrum-X Photonics SN6810 / SN6800	Ethernet CPO switch	2026 (per NVDA disclosure)	Scale-out Ethernet fabric
NVLink CPO (scale-up)	GPU-to-GPU CPO	2028	Scale-up GPU domain (replaces copper)

The progression is scale-out (network) CPO first, scale-up (GPU domain) CPO second, with the scale-up transition roughly 2–3 years behind the scale-out transition. NVIDIA is reportedly “going to extreme lengths to avoid going to optics” for scale-up out to 2027–2028 (SemiAnalysis CPO research) — the copper-to-optics inflection in the GPU domain is genuinely 2028+.

For Lumentum, the relevant CPO market segments split as:

Scale-out CPO switch optical engines — early-volume opportunity 2026–2027; Lumentum 200G/lane EML feeds the laser source layer of these modules
Scale-up NVLink CPO — 2028+ opportunity; very large unit volume per AI cluster (every GPU is a node) but still tightly bound to NVDA’s roadmap

Hyperscaler CPO posture

Outside NVIDIA’s switch products, the hyperscalers each have their own CPO programs with differentiated architectures:

Hyperscaler	CPO posture	Status
Microsoft (Azure)	Co-developing with merchant suppliers	Pluggable focus through 2026; CPO eval
Google	Internal silicon-photonics-led approach	TPU-coupled photonics; less reliant on merchant EML
Amazon (AWS)	Multi-vendor; in-house Annapurna efforts	Pluggable focus; CPO selective
Meta	Aggressive CPO eval; OCP contributor	Pluggable focus; CPO 2027+
NVIDIA-customer hyperscalers	Adopt NVDA’s CPO with NVDA’s switch products	Implicit — they buy NVDA’s SuperPOD / Spectrum-X kit

The market splits into “NVIDIA-stack CPO” (Quantum-X/Spectrum-X) and “hyperscaler-internal CPO” — and the InP EML source layer is required in both. Lumentum’s 200G/lane EML position is upstream of either route.

Alternative architectures: LPO and OBO

The 2026–2028 transition will not be CPO-only. Two intermediate architectures compete:

Linear-Pluggable Optics (LPO) — removes the DSP/PHY silicon from the pluggable module and pushes signal-recovery responsibility to the ASIC. Less power, lower cost than DSP-equipped pluggables, but more limited reach. Industry adoption: meaningful 800G LPO traction in CY2025 in datacenter spine; uncertain 1.6T trajectory. Implication for LITE: LPO modules still consume 200G/lane EML; reduces module-assembly value (Cloud Light pressure) but does not pressure the EML chip layer. Net neutral-to-positive at the LITE component level.

On-Board Optics (OBO) — optical engines mounted on the linecard (board-level) rather than in the ASIC package. Bridge architecture between pluggable and CPO. Some industry momentum but less than LPO or CPO. Implication for LITE: similar to LPO — EML chip required regardless.

The architectural battle is at the module-assembly layer (DSP-equipped pluggable vs. LPO vs. OBO vs. CPO), not at the source-laser layer (where InP EML is the consensus winner across all four).

Why CPO is structurally favorable for the EML chip layer

CPO modules consume more InP EML chips per unit of bandwidth than pluggables. Three reasons:

More lanes per CPO module — CPO engines integrate higher-radix optical interfaces; an 8-lane or 16-lane CPO engine has more EML-die content than a typical 8-lane pluggable.
Tighter optical-power budgets — CPO modules push closer to the link-budget limit (because the package thermal envelope is constrained), which favors reliable single-mode EML chips over alternatives.
External-laser-source (ELS) architectures — some CPO designs use a remote external laser source (a dedicated InP laser-array module separate from the CPO engine itself). This is incremental EML chip volume on top of the in-engine count.

Net effect: CPO transition is EML-chip-volume-positive even if module-ASP economics shift.

Risk to Lumentum’s pluggable-module business

The Cloud Light heritage transceiver-assembly business is at structural risk in a CPO transition. Pluggable-module assemblers (Innolight, Eoptolink, Cloud Light, Coherent’s Finisar legacy) are the layer most directly disrupted:

CPO modules are assembled in switch-OEM-led ecosystems (NVIDIA, Cisco, Arista, Juniper) rather than module-vendor-led
The CPO assembly value-chain may shift to OSAT providers or switch-ASIC-vendor in-house assembly rather than independent transceiver vendors
Pluggable-module ASPs and unit volumes face long-term decline as CPO captures premium AI-spine deployment

Lumentum’s defensive position is its chip-layer franchise (which CPO does not threaten and may amplify), not its module-assembly franchise. The Cloud Light acquisition’s strategic logic shifts in the CPO era from “be in the finished-module business” to “be a credible CPO module assembler too” — Lumentum has the heterogeneous packaging skills from Cloud Light to compete in CPO assembly, but so does Coherent (more deeply, given Finisar legacy).

Timing risk

The principal CPO timing risk for the LITE thesis is slip. If NVDA’s NVLink CPO scale-up delays from 2028 to 2029–2030 (plausible given the engineering complexity and copper’s continuing reach), Lumentum’s investment in CPO-relevant capacity gets a longer payback period. The opposite risk — CPO accelerates faster than expected — is the bull case.

The visible early-warning signals:

NVIDIA Quantum-X / Spectrum-X Photonics shipment volume disclosures (does the scale-out CPO ramp materialize on schedule?)
Hyperscaler CPO qualification announcements (who’s first to qual a non-NVDA CPO?)
Module-vendor CPO assembly capacity commitments (where is the assembly volume forming?)

Cross-link

AI capex cycle
Datacenter optics TAM — TAM mix-shift mechanic
InP EML duopoly
02_technology CPO roadmap — Lumentum’s product-side CPO position
03_ecosystem NVIDIA partnership
07_thesis bear case — CPO timeline-slip risk