Clean Compute Centers — True Photonic, Inc.

Four line items disappear.

No cooling plant

Chillers, air handlers, hot and cold aisle containment, and the floor area given over to all of it. A photonic compute floor does not generate the heat those systems exist to remove.

No cooling water

A conventional hyperscale campus can consume millions of gallons a day. A Clean Compute Center consumes none, which removes the objection that has stopped data center projects in water-stressed jurisdictions across the country.

No dedicated substation

The load is small enough, and divisible enough, to be served by the electrical service a commercial building already has.

No interconnect queue

Projects that do not require new transmission capacity do not wait years for it. This is usually the difference between a site that is live in months and one that is live at the end of the decade.

Divisibility, rather than total draw.

A conventional AI rack draws forty to a hundred and thirty kilowatts, and an enterprise rack five to fifteen. Neither divides below the rack. A building either has service built to match that load or it does not host the equipment.

A photonic module sits orders of magnitude below a conventional AI rack and divides at roughly rack scale. That means it can be fitted across the panels, floors, and risers a building already has, matched to whatever service exists rather than requiring service built to match it.

The divisibility is what makes the rest of it work. It is why compute can go into a downtown office tower, a campus building, or a facility already connected to the grid, rather than into a greenfield site waiting on a substation.

Three tiers, chosen by what the workload tolerates.

Tightly coupled

Training and other workloads that need processors close together run on a campus, or in a tower used as a vertical campus. Stacking floors keeps the machines near one another.

Within a mile

Workloads that tolerate tens of microseconds can be distributed across buildings inside a mile. Light in fiber travels about two hundred million meters a second, which puts a mile at roughly eight microseconds one way.

Regional

Inference sits near the users it serves. This is where distribution earns its keep, because the latency that matters is the distance to the person waiting for an answer.

Check it yourself

A million square feet on a single floor is roughly a thousand feet by a thousand feet. The longest path inside it runs corner to corner, about 1,414 feet, or 2.2 microseconds one way in fiber.

The same million square feet in a forty-story tower at twenty-five thousand square feet a floor is a footprint of about a hundred and fifty-eight feet, four hundred feet tall. The longest internal path is about 458 feet, or 0.7 microseconds.

The tower has roughly a third the maximum internal path of an equivalent single-story hall. Stacking shortens interconnect distance rather than merely tolerating it.

Roughly a tenth of the energy, at facility scale.

The figure is a facility-level number and we state the scale every time we use it, because the same figure applied to a switch or a single server would be wrong in both directions. It accounts for compute drawing less and for the cooling plant not existing at all.

It is a modeled figure. The program that will settle it measures energy to solution at equal throughput, which is the honest form of the question: the same work, measured against the electricity it took.

On build cost, we are waiting

A capital cost per gigawatt of compute depends on manufacturing cost at volume, and that number is ours to supply once production runs. Until it does, we would rather say so than publish a figure derived backwards from a power ratio.

What can be stated now is the avoided cost, and it can be checked without us: no substation, no transmission interconnect, no cooling plant, no water infrastructure, and buildings that are already standing.

Zoning and permitting.

The first question a city asks about a data center is what it will draw and what it will consume. The second is where it is permitted to go. Both questions were written for a facility with a substation and a cooling plant, and a photonic compute floor has neither.

We have worked that question through with an American city planning commission, following New Orleans' adoption of an interim zoning district for data centers in January 2026, and produced a brief setting out how a facility with no water draw and no dedicated substation should be classified. Municipalities and building owners weighing the same question can request it.

Compute that fits the building already standing.