Where the watts go.
A first-principles model from one switch port to a whole facility: electricity, heat, water, carbon. Every input is sourced. Change the inputs and the numbers follow.
The chain, not a number.
Claims like "100x less energy" are meaningless until you say per what. The honest answer is a chain of four measured links.
Per traversal. Crossing a 51.2 Tb/s electrical switch with its 64 pluggable optics costs about 40 picojoules per bit, from the box's own power rating. Crossing a mirror-based optical circuit switch costs about one picojoule, because the light is only steered.
Per fabric. When Google replaced the spine of its data center network with optical switching, network power fell 41%, measured in production.
Per facility. Networking was 4.5% of US data center electricity in 2024, up from 3.4% in 2018 as accelerator fabrics spread, and it is 5 to 8% of an AI cluster's IT load. Forty-one percent of that is 2 to 3% of IT. LBNL 2025 update
Per watt of IT. Every watt removed inside the hall saves PUE watts at the meter, the heat the cooling plant would have rejected, the water evaporated on site, and the water and carbon at the power plant.
One traversal, energy per bit.
| Traversal | pJ per bit | Basis |
|---|---|---|
| Electrical, 51.2T switch with 64 DSP pluggables, typical | 40.6 | Nvidia SN5610, 2,080 W with optics |
| Same, decomposed at datasheet maxima | 52 | 20 RX module + 2 SerDes + 10 ASIC + 20 TX module |
| Google Palomar OCS, 3D MEMS, 136 ports | 0.99 | 108 W, Apollo paper |
| Calient S320, 3D MEMS, 320 ports | 0.18 | 45 W datasheet |
| Polatis 7000n, piezo, 384 ports | 0.46 | 140 W datasheet |
| PQT bench prototype, DLP, 32 ports, 16 dB, one 3 W amplifier per fiber | 4.2 | 12 W electronics plus 96 W of EDFAs |
| PQT next chip, phase MEMS, 1,024 ports, 3 dB, no amplifier | 0.01 | design target, not measured |
Loss is an energy term. A binary hologram on a DLP cannot exceed 10% diffraction efficiency, so today's prototype loses about 16 dB end to end. No standard link budget absorbs that: DR allows 3 dB of channel loss, FR 4 dB, LR 6.3 dB. So every fiber through the prototype needs 15 to 20 dB of amplification, and the amplifiers outweigh the switch eight to one. Even amplified, it is ten times better per bit than an electrical traversal; it is only worse than a mirror-based OCS. At 3 dB the amplifiers go away and the switch's own 12 W, spread across a thousand ports, would be, on paper, a hundred times better than the best switch shipping today. That is the whole energy case, and it is the engineering the next chip buys. One caution we take from the inventor: none of the switch's own power figures has been measured. The 12 W is a sum of part datasheets, the per-bit rows above are arithmetic from it, and the next-chip row is a design target. The switch is certainly more efficient than an electronic one, because it converts nothing; by how much is a measurement still to be made, and we do not put a multiplier on it. Two further savings are structural: the switch has no servo loop, since the mirrors land on physical stops rather than being located by feedback, and a single input can multicast to several outputs in one pass, so a broadcast step in a training job is one transmission instead of many.
Your facility.
Set the inputs for a data center you know. The outputs are what replacing its spine with optical switching would save each year, using Google's measured 41%.
The ceiling for the whole country.
US data centers used 192 TWh in 2024, 4.7% of the nation's electricity, and networking was 4.5% of that. If every spine were optical and saved 41%, the fleet would save about 3.5 TWh a year, 1.2 million tonnes of CO₂, about 15 million cubic meters of water at power plants, and about $300 million. That is 1.8% of data center electricity, against a fleet projected to reach 649 TWh by 2030.
Real, and worth doing. Not a revolution in the grid, and we will not pretend it is one. The larger prize is what microsecond switching lets a cluster do: run the same work on fewer, busier switches, and stop over-provisioning for traffic that an optical fabric can simply follow.