The switch industry has spent a decade promising that optics would move onto the package. In 2026 it is finally happening, in volume, at 51.2 terabits per second per chassis. The interesting question is no longer whether co-packaged optics works. It is who absorbs the operational consequences, because those consequences are not symmetric between a hyperscaler and everyone else.
Pluggable optics are a triumph of modular design. A failed transceiver is a field-replaceable unit: a technician pulls it, inserts a spare, and the port comes back in ninety seconds. Co-packaged optics integrates the photonic engine next to the switch ASIC, which removes the electrical channel that dominated the power budget, and in doing so removes the thing that made failures cheap.
Part One: The power argument, honestly stated
At 800G, a pluggable transceiver consumes roughly 12 to 15 watts, depending on reach and whether it uses digital signal processing at both ends. A 51.2T switch with 64 ports of 800G therefore spends close to a kilowatt on optics alone, which is comparable to the ASIC's own dissipation and roughly a third of the total chassis power.
Co-packaged optics removes the long electrical trace between the ASIC and the optical engine, and removes the retimer that exists only to survive that trace. Published demonstrations put the optical engine at 5 to 7 watts for the same 800G, and the system-level saving is larger than the module-level saving because the retimer and the package-to-front-panel path also disappear.
The saving is real, and it is the reason hyperscalers are deploying. But power per bit is a metric that only matters when you are power-constrained at the rack level. A colocation tenant buying a 42U cage with 8 kilowatts of power cares about the ratio of usable ports to watts, and yes, co-packaged optics helps. A enterprise network that replaces a core switch every seven years cares about mean time between failures and replacement cost, and there co-packaged optics is a much harder sell.
Part Two: The failure model is the whole argument
A pluggable module has a laser, a driver, a photodiode, a DSP and a connector in a metal shell. It fails, on average, at a rate that vendors express in FIT: failures in time, or failures per billion device-hours. A mature 400G module might run at a few hundred FIT, which in a 10,000-module fabric means a handful of failures per month. That is an operational rhythm, not an incident.
Co-packaged optics changes the denominator. The laser is usually still discrete, moved to a remote laser source, because lasers degrade and you want to replace them without touching the package. But the optical engine itself is now soldered to the same substrate as a switch ASIC that costs tens of thousands of dollars. A failure in the engine is not a ninety-second swap. It is a chassis replacement, or a board swap inside a chassis, with the associated downtime and the associated spare-parts logistics.
Vendors answer this with reliability data: external laser sources are field-replaceable, the engine is qualified to telecom-grade lifetimes, the package is sealed. All of that is true. It is also true that the failure rate of a new integrated photonic engine is unknown until you have a few hundred million device-hours of field data, and nobody has that yet. The first adopters are paying for the learning curve, and they know it.
Part Three: The economics of the spare parts tray
There is a mundane operational detail that decides more deployments than any datasheet. Pluggable optics are interoperable. A 400G DR4 from vendor A works in vendor B's port, because the standard exists and the market is competitive. That interoperability is what keeps module prices falling and what lets a network team carry one spare type instead of nine.
Co-packaged optics is, by construction, a proprietary interface. The optical engine is part of the switch. You do not buy a cheaper engine from a second source; you buy the switch, and you buy the vendor's support contract, and you accept the vendor's roadmap. The lock-in is not a side effect. It is the business model.
This is why the deployment pattern looks the way it does. Hyperscalers with their own hardware teams and their own software stacks adopt first, because they can absorb lock-in and they capture the power savings at a scale where a kilowatt per chassis is a megawatt per hall. Everyone else waits for the second generation, when the standard bodies have defined a compliant interface and a second vendor has a part.
Part Four: What 1.6T does to the timing
The transition to 1.6T per port compresses the timeline. At 1.6T, a pluggable module is projected to draw 25 to 30 watts, and the front-panel power budget of a 1U switch runs out of room. There is a physical limit here: a front panel can only dissipate so much before the airflow required to cool it becomes the dominant acoustic and power cost of the rack.
That limit is what makes co-packaged optics inevitable rather than optional at some point in the next three years. The only question is whether the industry gets a standardised, multi-vendor optical engine before it gets forced into proprietary ones. The history of networking suggests it will get both, in that order, and that the proprietary phase will last about one product cycle.
Part Five: The everyday consequence
It is easy to treat this as a data centre plumbing question. It is not, because the cost of moving a bit sets the price of every service that moves bits. Video calls, remote surgery, cloud gaming, model inference at the edge: all of them are priced against a bandwidth cost curve that has fallen by roughly a factor of ten per decade.
If co-packaged optics delivers the promised halving of power per bit, that curve continues. If it delivers a proprietary, hard-to-service interconnect with a five-year replacement cycle, the curve flattens for everyone who is not a hyperscaler, and the services that depend on cheap bandwidth become a little more expensive for everyone else. That is the real stake in a connector standard.
Market sizing
The optical transceiver market passed 15 billion dollars in 2025 and is forecast to reach 25 to 30 billion by 2030, with the 800G and 1.6T segments growing fastest. Co-packaged optics, starting from a small base, is projected to take 10 to 15% of that by 2030 in the most aggressive forecasts and 5% in the conservative ones. The spread between those forecasts is not about technology. It is about whether the failure model turns out to be acceptable.
The comparison that matters
Pluggables optimise for replaceability and interoperability, at the cost of power. Co-packaged optics optimises for power and density, at the cost of serviceability and vendor choice. Neither is universally better; they are answers to different constraints.
The practical rule: if your binding constraint is megawatts per hall and you have a hardware team that can qualify a new failure mode, co-packaged optics pays for itself in eighteen months. If your binding constraint is uptime and a small operations team, the pluggable tray is not legacy technology. It is the correct engineering choice, and it will remain so until the failure data exists.
SOL: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS
Creating this takes real effort — thank you for the support.
Top comments (0)