Compute

The frame

A Flow With Neither

Compute cannot be banked and cannot be stockpiled. The only hedge against an outage is a decision computed in advance, which makes the discipline siting.

Compute is produced and consumed on the spot, and that is the whole of its difference from the other two substrates. Energy has a buffer and water has an inventory; compute has neither. A rack that loses power does not draw down a reserve of thinking - it simply stops, and the only protection is to have already computed the decisions the site will need. So compute is managed by siting rather than by storage: put it where it can run and reject heat, because once it is running there is nothing to hold back.

One more fact sets it apart. Energy and water are made and consumed locally, but the hardware compute runs on is always an import - chips are the hardest thing in the Primitives to make off Earth. The service is local; the silicon is shipped. A site’s compute economy is therefore set by resupply cadence and failure rate, not by what the ground holds. Building

Thermal siting

Rejection Temperature Is the Discriminator, Not Altitude

Where compute belongs is decided by the temperature it can dump heat to, which is why a cold crater floor is a different market from a warm orbit.

The case for moving a data center off the ground is thermal, and it is governed by rejection temperature, not by how high the orbit is. Radiator performance scales with the fourth power of temperature, so a modest change in the sink is a large change in what a given radiator can shed. The gain from leaving low Earth orbit is about 10.7% at a 345 K sink and is spent by geostationary orbit; at 250 K it rises to 26 to 54%; below the passive radiative floor it is effectively unbounded. That is why a permanently shadowed lunar crater floor at 40 K is a different commodity market from a solar rack in LEO - the two price compute on entirely different thermal terms. Observed

The cold trap

Using the Cold Is Consuming It

A cold sink is a resource that a radiator spends rather than taps, so cold compute competes for the coldest ground instead of getting it free.

The coldest terrain looks like a free thermal sink, but using it consumes it. A surface at 25 K radiates only about 0.022 W/m², so rejecting 10 kW into that regime demands on the order of 83,000 m² of radiator, and the radiator itself can never sit inside the cold trap it depends on. Cold compute is therefore a claimant on the 25 to 40 K ground, not a free rider on it, and it competes there with supervolatile cold storage and the mass-driver pulse store for the same scarce real estate. The cold is a resource with a budget. Building

Train vs infer

Inference Disperses; Training Does Not

The two halves of the compute workload want opposite geographies, and the split, not the silicon, is what sites a cluster.

Training and inference pull in opposite directions. Training couples thousands of processors in a tight all-reduce that wants about 3.2 Tb/s per node against the 1.6 demonstrated today, and optical links lose roughly 40 dB going from 1 km to 100 km of separation, so a training cluster wants everything packed close. Inference disaggregates - it tolerates distance and runs happily spread out. The gate on packing training close is not silicon or sunlight; it is microradian pointing and centimeter-scale relative navigation between free-flying modules. That 1 km co-location ball is the tightest spatial requirement in the whole model, and it sits in exactly the regime where spreading out is the survivability argument - a tension the design has to resolve, not wish away. Projected

Co-location

Co-Located Means Sharing a Generator

The useful map of a compute site is electrical, not spatial, and what leaves it is decisions, not data.

Two capabilities are co-located, in the sense that matters, when they share a power generator - whatever the distance between them. The relevant map is the electrical one, not the spatial one, because the binding shared resource is the bus, and anything on the same bus rises and falls together. What a compute site ships out is not raw data but decisions: an aggregate percept compressed at about 400 to 1, so a site sends the conclusion rather than the feed. That compression is why compute converts into bandwidth so efficiently and why dispatching compute well is named the largest unexploited lever between the three substrates. Building

Closure

The Service Closes; the Hardware Never Does

Compute is the one substrate whose closure is capped by the Element Register, not by energy or water.

The compute service closes almost as soon as power exists, because it is made and used on the spot - a site with a reactor and a rack is computing. The hardware never closes. Chips are the top of the Primitives ladder, the class most dependent on permanent imports, and nothing about local energy or water changes that. So compute is the substrate whose ceiling is set by resupply cadence and failure rate rather than by local resources, and the figure that governs it is the fraction of its own hardware a site can replace, which is near zero and stays there. Building

Contested

Deliberately Not Claimed

The load-bearing uncertainties, tagged with the confidence they earn.

Record

Provenance and Data Sources