Reading the XPO MSA, the pluggable fights back

An obituary, returned to sender

For several years the industry told the story of pluggable optics as an obituary. The arguments were these:

• Energy per bit is too high.
• The faceplate reached its density limit.
• The future belongs to co-packaging.
• The module had a good run.

The people who told that story had good arithmetic, and this site repeated much of it.

Then, at OFC 2026, a group led by Arista announced the XPO MSA (multi-source agreement). Coherent was among the founding members. The MSA defines a pluggable module with these properties:

• 12.8 terabits per second
• liquid cooling through an integrated cold plate
• up to 400 watts of dissipation
• four times the front-panel density of the 1.6T OSFP modules that ship today

The obituary was premature. Read the numbers carefully, because each one is a step in an argument.

Decoding 12.8, 204.8 and 400

Start with the density claim: 204.8 terabits of front panel per OCP rack unit. Divide by the module, and the result is sixteen XPO modules per rack unit. Each module has 64 electrical lanes at 200G, on two paddle cards of 32 lanes each. So sixteen modules are 1,024 lanes at 200G.

Now recall the switch lane count. The shipping 102.4T Tomahawk 6 has 512 lanes at 200G. XPO keeps that lane rate and doubles the lane count, and 1,024 × 200G = 204.8T. So sixteen modules in one rack unit serve one entire 204.8T switch.

The designers sized the module from the silicon backwards. In the same way, 64 × 1.6T modules serve a 102.4T switch today. XPO is not a module that happens to be big. It is the faceplate of the 204.8T generation, designed as one unit.

Now the power. Four hundred watts over 12.8 terabits is 31 picojoules per bit. That number looks alarming next to the 15-to-20 band of today’s modules. But it is a thermal ceiling, not a target. The spec explicitly accommodates linear (no DSP), half-retimed (DSP on the transmit side only) and fully retimed electronics. It also accommodates reach classes from SR and DR inside the data center out to ZR and ZR+ between data centers.

A fully retimed ZR+ engine with coherent-class electronics uses a lot of power. The 400-watt envelope exists so that even that engine fits. A linear DR variant of the same shell would use far less. So one mechanical and thermal platform holds the whole electronics ladder. The generous ceiling is what gives the flexibility.

The liquid is the least optional part. Sixteen modules at up to 400 watts is 6.4 kilowatts of optics on one rack unit of faceplate, before the switch beneath it. Air across a faceplate cannot carry that heat. The cooling walls essay puts the limit of air near 100 kW for a whole rack. Here 6.4 kW sits in one rack unit. The integrated cold plate simply extends the liquid loop, which already reached the rack, to the last inch of the front panel.

What the faceplate is actually defending

Why go to these lengths to keep the optics in a cage at all, when the energy arithmetic favors the package? The reason is that the cage defends two properties that the package cannot yet match. Both are operational, not photonic.

First, a dead XPO module at two in the morning is still a pull and a push, hot, in the aisle. That is the one-minute repair that co-packaging gives up, and that OpenCPX works to partially recover.

Second, a pluggable faceplate remains a marketplace. One switch can hold DR from one vendor, ZR+ from another, and linear from a third when the platform qualifies it. Multi-sourcing at the moment of purchase is where the optics layer already has choice. It is a property of cages, not substrates.

Read against the other two agreements of that week, XPO completes the picture. The three agreements divide the work more than they compete:

• OCI concedes the deepest scale-up links to silicon-centric optics.
• OpenCPX standardizes the middle rungs.
• XPO declares that, for everything else, the faceplate intends to match the switch roadmap generation for generation. That includes scale-out fabrics, DCI (data center interconnect) and the long tail of reach classes.

Liquid cooling is the condition for entry. The pluggable is not dying. It is specializing.

What to take away

1. XPO defines a 12.8T, liquid-cooled, up-to-400 W pluggable at four times today’s faceplate density. Arista leads it, with Coherent among the founders.
2. The numbers match the silicon. Sixteen modules per rack unit is 1,024 lanes at 200G, or 204.8T. That is exactly the next switch generation: one RU of faceplate per ASIC.
3. 400 W ÷ 12.8T = a 31 pJ/bit ceiling. It is deliberately generous, so one shell spans linear to fully retimed electronics and SR to ZR+ reach.
4. Liquid is an arithmetic requirement: 6.4 kW of optics per rack unit of faceplate is beyond any airflow.
5. The strategic reading: the faceplate cedes deep scale-up to OCI-class optics. It defends everything else with density, liquid, and the two properties that packages lack: the one-minute repair and the multi-vendor cage.