Pluggable vs Embedded — When the CPO Hype Actually Pays Off

Every conference deck about co-packaged optics shows the same number: switching ASIC SerDes drives 30 to 40 percent of the system power, and CPO can eliminate most of it. The number is correct. The conclusion that follows from it — that CPO is the future of every high-density switch — does not follow.

CPO solves one problem and creates four new ones. The math works for two deployment shapes. Outside those, pluggable wins on TCO, and will keep winning until at least 2027.

The problem CPO actually solves

Driving signals from a switch ASIC to an external pluggable cage costs power. On a high-radix 51.2T switch, the SerDes blocks driving 256 lanes at 112 GBaud burn somewhere between 200 and 350 watts on their own. That number scales upward with each generation. Past about 102.4T, the SerDes power becomes the limiting factor for what you can cool in a 1RU box.

CPO moves the optics next to the ASIC. The signal travels millimeters instead of centimeters. The SerDes can run shorter, lower-power reach modes. The savings on a 51.2T switch are around 100 to 140 watts, which is the difference between a box that runs hot but works and a box that needs liquid cooling.

That's the win. It's real.

The four problems CPO creates

Yield. A pluggable module that fails comes out of the cage and gets replaced in five minutes. A CPO module that fails takes the whole switch down. The yield economics of a thousand-port switch built around CPO are not the same as the yield economics of a thousand pluggable modules. Even a small per-module failure rate compounds badly when the modules are soldered into the ASIC carrier.

Repair. Field-replaceable doesn't survive in CPO. The module comes out only with the whole package. Spares strategy, RMA flow, maintenance windows — all of these get rewritten. Hyperscalers can build that flow. Most enterprises cannot.

Qualification cycle. Pluggable optics qualify against a switch independently. CPO optics qualify only as a complete switch SKU. Every line speed change, every fiber type change, every reach class becomes a new switch qualification event. The qualification matrix for a generation of CPO products is meaningfully larger than for the equivalent pluggable family.

Supplier diversity. With pluggable, the switch vendor and the optics vendor are independent businesses. You can dual-source the optics, take the cheaper bidder, swap on price pressure. With CPO, the switch and the optics are one bill of materials. The single-source risk becomes structural rather than commercial.

Where the math actually works

Two deployment shapes break even or better on CPO.

Hyperscaler AI fabrics inside a single data hall. The customer is one buyer with a multi-thousand-unit volume commitment. The qualification matrix is bounded — same reach, same speed, same fiber. The repair model is centralised maintenance teams who can swap whole switches as the unit of replacement. The yield economics are absorbed by the volume. This is the deployment shape that paid for the development of CPO in the first place.

Carrier core platforms with extremely long service lives. A vendor selling into a Tier-1 carrier for a 10-year service contract can amortise the yield and qualification cost across the contract. The carrier values the power saving for its operational TCO. The vendor accepts the qualification overhead because they're already running long-cycle hardware programs.

Outside these two shapes, the cost surfaces of CPO don't get amortised. Enterprise data centers, regional carriers, telco edge — all of them have qualification cycles too tight and volume too low to swallow the yield and repair overhead. The TCO calculation tips back toward pluggable.

What the timeline actually looks like

The vendors who can ship CPO at scale are the ones already in the hyperscaler AI fabric supply chain. They've been sampling for two years and have meaningful pre-production volume now. First production deployments are inside specific hyperscaler clusters in 2026. Volume production for that segment is 2027 to 2028.

The second wave — carrier core, hyperscaler general-purpose data center — is 2028 to 2030. Enterprise CPO, where the qualification and repair stories matter most, is not on the public roadmap of any vendor I've read. Pluggable will dominate that segment through at least 2030.

The third wave, never, is on the roadmap of vendors talking about CPO for enterprise switching. It's a marketing story. There is no path that makes the qualification economics work for the enterprise refresh cycle.

How to read vendor CPO slides

Three filters cut through most of the noise.

First — what's the deployment shape they're targeting? If they don't say "hyperscaler AI fabric" or "carrier core platform" explicitly, the technology might be real but the segment isn't bought.

Second — what's the field replacement model? If the vendor can't answer cleanly, the operational TCO has been waved away rather than designed.

Third — what's the qualification matrix for the first generation? If it's one reach and one speed, that's a hyperscaler-only product. If they claim it covers everything, they haven't built the qualification rigs yet.

CPO is not the future of high-radix switching. It's the present of a specific high-radix switching segment, and the future of one more. Pluggable is the future of everything else. The TCO math, for now, is not close.