Co-Packaged Optics Goes Mainstream: The Optics Bill Gets Rewritten
NVIDIA's Spectrum-X Photonics hit full production in August. Why moving the optics onto the switch package matters for power, reliability, and the supply chain, and the CPO-vs-NPO fight coming for Rubin Ultra scale-up.
For years, co-packaged optics (CPO) was always "two years away." In 2026 it actually shipped. NVIDIA said in August that Spectrum-X Ethernet Photonics switches had entered full production, with first systems going to CoreWeave, Lambda, and Oracle, and Azure, IBM Cloud, and Nebius named as early adopters (Macroplane summary). Quantum-X InfiniBand Photonics began low-volume shipments earlier in the year.
This matters because in a big AI fabric, the optics are the network bill, in dollars and in watts.
TL;DR
- A pluggable 800G-class transceiver burns ~30 W per interface. NVIDIA's CPO engines get that to ~9 W, and electrical loss from ASIC to optics drops from ~22 dB to ~4 dB (NVIDIA).
- On a 100K-GPU three-tier fabric, that's on the order of 10 MW of switch-side optics power. That power can run GPUs instead.
- One sell-side forecast puts NVIDIA CPO switches at 15K units in 4Q26 alone and ~100K in 2027. The next fight is CPO vs NPO for Rubin Ultra scale-up.
What CPO actually changes
In a normal switch, the signal goes from the ASIC across the PCB, through a connector, and into a pluggable module with its own DSP before it becomes light. NVIDIA puts that electrical path at up to 22 dB of loss on 200G-per-lane channels, which is why each module needs a power-hungry DSP. CPO moves the optical engine right next to the switch ASIC on the same package. Fiber attaches directly, loss drops to ~4 dB, and per-port power falls to as low as ~9 W (NVIDIA Technical Blog).
NVIDIA's claimed system-level gains: 3.5x power efficiency, 10x resiliency (far fewer discrete active parts to fail), and 1.3x faster time-to-turn-on.
The platforms:
| Switch | Fabric | Capacity | Ports |
|---|---|---|---|
| Quantum-X Photonics (Q3450) | InfiniBand | 115 Tb/s | 144 × 800G |
| Spectrum-X Photonics SN6810 | Ethernet | 102.4 Tb/s | 128 × 800G |
| Spectrum-X Photonics SN6800 | Ethernet | 409.6 Tb/s | 512 × 800G |
All of them are liquid-cooled. That 512-port SN6800 matters for topology too: a radix of 512 means a two-tier fabric can reach 131K endpoints (k²/2). That removes a whole tier of switches and optics compared to 64- or 144-port boxes. Background math in Fat Trees.
The power math at cluster scale
Take a 100K-GPU, non-blocking, three-tier fabric with one 800G NIC per GPU. There are roughly N links at each tier, so about 5N switch-side port-ends (one at the leaf for GPU links, two per leaf–spine link, two per spine–core link).
| Pluggable (~30 W) | CPO (~9 W) | |
|---|---|---|
| Switch-side port-ends | ~500,000 | ~500,000 |
| Optics power | ~15 MW | ~4.5 MW |
| Savings | ~10 MW |
Treat this as an order-of-magnitude estimate. Real designs mix in copper, run fewer tiers, and use different module wattages. But ~10 MW at a fixed interconnect is roughly 80 NVL72 racks of GPUs you can now power. (NIC-side optics stay pluggable for now. CPO is a switch-side story this generation.)
The ramp and the supply chain
NVIDIA named its CPO supply chain publicly, which it rarely does: TSMC (silicon photonics wafers), SPIL/ASE (packaging and test), Lumentum (laser chips), TFC Optical (laser sub-assemblies), Foxconn (system assembly) (Macroplane).
Data points worth tracking:
- Units: one late-September sell-side note forecasts 15K CPO scale-out switches in 4Q26, 16K in 1Q27, 30K in 2Q27, ~100K for 2027. In August, another firm had 15K for all of 2026.
- Lasers are the choke point. CPO moves the laser out of the module into an external light source, so ultra-high-power lasers become the critical part. Lumentum guided UHP laser revenue for CPO to ~$50M/quarter by end of 2026 and >$100M/quarter by the March 2027 quarter, and said it was well behind its ability to supply.
- Ethernet speeds: Dell'Oro says 1.6 Tb/s switches began sampling and should ramp in 2H26 (Dell'Oro, Sep 2026). In the same report, AI back-end switch sales passed front-end sales for the first time.
The next fight: CPO vs NPO for scale-up
Scale-out was the easy win. The bigger opportunity is scale-up: Rubin Ultra NVL576 links eight 72-GPU racks into one NVLink domain, and copper can't span eight racks at NVLink speeds. Per the same coverage, NVIDIA is developing two options in parallel:
- CPO: optical engines bonded onto the NVLink switch package. Densest and most efficient, but if an engine dies, the whole package goes.
- NPO (near-packaged optics): engines in sockets next to the package. Slightly less efficient, but you can swap them.
SemiAnalysis reportedly expects NPO to reach the market first. Scale-up optical engine forecasts in that note go from 6.5M in 2027 to 42M in 2028. Whichever design wins decides which suppliers capture the value: fiber-harness and shuffle vendors gain under CPO, while TIA/driver vendors gain under NPO.
What it means if you're deploying
- Serviceability changes. No more swapping a $1K pluggable when a link flaps. Ask the vendor about replacement procedures, laser-module field replacement, and how failures are isolated.
- Liquid cooling for switches. These boxes are liquid-cooled. Your network rows need TCS loops now, not just compute rows.
- Fiber management inside and outside the box. Higher fiber density at the switch faceplate needs a cabling plan.
- Power budget wins are real. If you're power-capped (and you are), a few MW of optics savings shows up directly as GPU count.
Bottom line
CPO moved from slideware to shipping in 2026, and the ramp is running ahead of forecasts. The direct win is watts. The bigger second-order effects are higher-radix switches, fewer tiers, and a supply chain shift toward lasers and silicon photonics. If you're modeling a 2027 cluster with pluggables everywhere, you're probably overestimating both its power budget and its failure rate.