Co-Packaged Optics
▶ See it in 3D Photonics region
Instead of plugging fibre-optic cables into the edge of a switch, you build the light engine right into the chip package — less power, far more bandwidth. Broadcom and NVIDIA are first out; startups Ayar and Lightmatter push it even closer to the GPU. Real technology, but mostly a 2028-and-beyond story at scale.
What it is
Move the optics into the switch/GPU package — millimetres of silicon instead of centimetres of lossy copper. Broadcom's Davisson (102.4 Tb/s) is shipping, NVIDIA's Quantum-X Photonics went commercial in early 2026 with Spectrum-X due H2; Ayar and Lightmatter push optics right beside the compute die.
The business
CPO helps Broadcom, NVIDIA, the laser makers, TSMC (Si-photonics foundry) and pure-plays Ayar/Lightmatter. It threatens pluggable-module volume (InnoLight, Fabrinet) — but only from ~2028. Marvell paid $3.25B for Celestial AI, the valuation anchor for the whole theme.
Bottleneck severity
Manufacturing yield, not demand: attaching fibres to silicon at scale is still unsolved, there's no interoperable standard yet, and every CPO design needs the same scarce lasers as everything else.
Who makes it
BroadcomAVGO
Tomahawk switching + CPO (Davisson 102.4 Tbps) + custom XPUs for hyperscalers. Anchor of the UALink/merchant-silicon camp.
NVIDIANVDA
Buys the most HBM, NVLink and optics on earth and pre-buys supply to lock rivals out: ~60% of CoWoS, $2B each into Coherent and Lumentum (Mar 2026), $2B into Marvell. B300 (288 GB HBM3E) is the shipping reference GPU; Rubin (288 GB HBM4, 22 TB/s) samples Q4 2026, volume Q1 2027.
MarvellMRVL
~70% optical-DSP share + custom AI silicon + Celestial AI photonic memory fabric ($3.25B). Most complete optical stack in public equities.
Ayar LabsPRIVATE
In-package optical I/O chiplets (TeraPHY). $3.75B Series E; NVIDIA/AMD/Intel backed. IPO signaled.
LightmatterPRIVATE
Passage photonic interposer (114 Tbps). $4.4B valuation; T. Rowe / Google / HPE backed.
Celestial AI→ MRVL
Photonic Fabric for optical memory disaggregation. Acquired by Marvell for $3.25B — the valuation anchor.
TSMCTSM
Makes the GPU die AND the CoWoS packaging that bonds HBM to it — and now the logic base die under SK Hynix's HBM4 too. CoWoS is heading for ~130k wafers/month by end-2026 and is still sold out into 2027 (NVIDIA alone holds ~60%); it is farming 240–270k wafers/year of front-end steps out to Amkor and SPIL. A Taiwan disruption stops the whole stack.
Metrics that move this layer
- CPO attach ratehyperscaler deployments
- Ayar/Lightmatter IPOprivate-market temperature
- yield / fiber-attachthe manufacturing gate