An AI cluster's bill of materials is mostly fixed. The switches, the GPUs, the fabric, all specced and locked. Optics is different. It's the one component where a better choice quietly changes your operating cost for the life of the deployment.

Power is where that choice shows up. Most 800G 2xFR4 modules are specified at 15W and above. Ours run below 11W typical, with a 13W maximum published on the datasheet. Four watts a port doesn't sound like much. Across a cluster, running continuously, it compounds into real money and real carbon.

A single module saving four watts is easy to dismiss. A deployment of several thousand ports running around the clock is not.

At ten thousand modules, the difference is roughly half a gigawatt-hour a year, tens of thousands of dollars, and over a hundred tonnes of CO2. Every year the cluster runs. The number that matters isn't the annual figure, it's the five-year one, because that's the life you're buying.

Set your own deployment below and see what the gap is worth to you.

The 11W figure isn't inherited from a reference design. We design and build these modules ourselves, so the thermal and power work is ours.

That has one practical consequence worth caring about. What you qualify from a sample is what arrives in the shipment. The power performance doesn't drift between the unit you test and the ten thousand you deploy, because the same lines build both. Consistency at scale is the part that's hard to buy from a distributor, and it's the part we control.