A typical 800G OSFP 2xFR4 transceiver is specified at around 15W. Lower-power modules exist that run below 11W typical, with a maximum near 13W at full operating temperature. That difference looks minor on a single module. Across a cluster of several thousand ports running continuously, it becomes one of the few operating costs in the build that you can still change.

Optics power tends to get overlooked because each module's draw is small next to a GPU or a switch. But there are a lot of modules, they run around the clock, and they run for years. The maths compounds in a way that a single spec line doesn't show.

What drives power consumption in an 800G module

An 800G 2xFR4 module carries eight lanes of 100G PAM4, converts them between electrical and optical, and does the digital signal processing that keeps the link clean over distance. The DSP, the lasers and the drive electronics are where the power goes.

Silicon photonics designs integrate much of the optical path onto a single chip, which reduces component count and the power lost between discrete parts. That integration is a large part of why a well-designed 800G 2xFR4 module can sit meaningfully below the market-standard 15W rather than at it.

Power also rises with temperature. This is why the figure that matters for planning is the typical consumption under normal operating conditions, with the maximum at full temperature as the ceiling you design cooling around. A module quoted below 11W typical and 13W maximum tells you both numbers.

Why four watts a port matters in an AI cluster

Four watts saved on one module is 4W. The same four watts across ten thousand modules, running every hour of the year, is a different quantity.

A single port saving 4W and running continuously saves roughly 35 kWh a year at the module itself. Data centres don't only pay for the module, though. They pay to cool it too, so the real figure scales with power usage effectiveness. At a PUE of 1.4, that 35 kWh becomes closer to 49 kWh per port per year once cooling is included.

Multiply that across a real deployment and the numbers stop being rounding errors. Ten thousand ports at that rate is around half a gigawatt-hour a year, tens of thousands of dollars at typical commercial electricity prices, and over a hundred tonnes of CO2. Not once. Every year the cluster runs.

What the difference actually costs

The honest answer is that it depends on your electricity price, your PUE and your grid's carbon factor, because all three vary by site. That is exactly why a fixed figure in a blog post is less useful than your own.

We built a calculator that takes those three inputs and your deployment size and returns the annual energy, cost and CO2 difference, plus the five-year total. The five-year number is usually the one worth looking at, because it reflects the life you are actually buying, not a single year of it. You can run your own deployment through it here.

For most operators the result is the same shape: a saving that is easy to ignore per unit and hard to ignore per cluster.

How to capture it without changing supplier

The reason this cost usually goes unaddressed is that changing an optics supplier feels like more trouble than the saving is worth, especially when an existing relationship works well.

The useful thing about optics is that it is genuinely dual-sourceable. Modules that comply with the same MSA share form factor, interface and behaviour, so qualifying a second source does not disrupt what you already run. You keep buying what you buy now, qualify a lower-power module alongside it, and choose port by port where the power and the lead time work in your favour. Nothing gets ripped out, and the saving lands on the ports where it counts.

For a NeoCloud scaling a build, a reseller fulfilling against a contract, or an enterprise standing up a cluster, that is a low-risk way to act on a number that would otherwise stay theoretical.

Consistency is the part that is hard to buy

One caveat worth naming. A power figure from a sample is only useful if it holds across the shipment. Where modules are designed and built by the same manufacturer on their own lines, the thermal and power performance stays consistent from the first unit to volume, because the same process produces both. That consistency at scale is the part of a power claim that is genuinely difficult to verify from a distributor, and it is worth asking about during qualification.

The short version

A typical 800G 2xFR4 module is specified at 15W. Modules running below 11W typical exist, and across a cluster the gap is real money and real carbon, every year of the deployment. You can act on it without switching supplier by qualifying a lower-power module as a second source. The size of the prize depends on your own numbers, which is worth working out before your next optics order.

See what it is worth for your deployment and download the datasheets while you are there.