Starcloud

Every frontier model release shifts the bottleneck further upstream. The constraint is no longer GPU supply — it is megawatts. Hyperscalers are sitting on chips they cannot plug in because the grid, the substations, and the water rights are still being negotiated.^[13]

The terrestrial path forward asks AI to compete with cities, factories, and farms for the same kilowatt-hours and the same gallons of cooling water — under multi-year permitting timelines that did not anticipate AI.^[4]

Power. In a dawn-dusk sun-synchronous orbit, a satellite stays in continuous sunlight year-round — a >95% capacity factor versus the 24% median for US terrestrial solar farms, plus ~40% higher peak irradiance with no atmosphere to absorb the spectrum. Johnston puts the net effect at roughly 8× more energy per square meter of array — before you account for the permitting, transmission, and storage required on Earth.^{[5] [18]}

Heat. Cooling is harder in vacuum, not easier — there is no convection or conduction, only radiation. This is the genuinely difficult part of the design, and the whitepaper budgets 70% of engineering hours on thermals. The math works: a 20°C radiator panel emits ~633 W/m² net to the 2.7 K cosmic background per Stefan-Boltzmann, with no chillers and no cooling water. And the architecture is already operational at scale — NASA's International Space Station has rejected ~70 kW of waste heat continuously since 1998, using two ammonia coolant loops feeding deployable radiator panels that emit infrared into deep space. Starcloud's design is bigger; the building blocks are flight-proven.^{[5] [18] [19]}

Launch cost. Cost-to-LEO fell from ~$20,000/kg in 2010 to ~$1,500/kg today with Falcon Heavy. McKinsey's analysis sets ~$500/kg as the threshold at which orbital compute becomes cost-competitive with a terrestrial cluster of the same capacity; below that, it is cheaper outright. SpaceX's stated Starship target at flight rate is $100–200/kg. This is the curve that triggers everything else.^[4]

Total infrastructure cost is roughly $5M per megawatt in orbit versus $12–15M on Earth, with no batteries, chillers, cooling towers, or AC/DC conversion to pay for.^[4] Starcloud is signing near-term LOIs at 3¢/kWh, headed toward sub-half-a-cent in the medium term and roughly $0.002/kWh over a 10-year amortization.^{[5] [18]}

The trigger is launch cost. Below roughly $500/kg to LEO, orbital compute is cost-competitive with terrestrial; further down the curve it's outright cheaper. Starship's flight cadence puts that threshold inside this decade.^[4]

Marginal cost dynamics also invert. On Earth, every additional gigawatt is harder to build — substations and transmission rights bid up. In orbit, every additional satellite is cheaper, because manufacturing rate and launch cadence both improve with volume.

Johnston's working number is ~$1 trillion of annual capex moving into orbital compute within ten years, with most new compute capacity deploying to space inside the same window. The full constellation Starcloud has filed for — 88,000 satellites — would deliver roughly 20 gigawatts of compute capacity, primarily for inference workloads.^{[4] [5]}

The near-term market is narrower and hotter. Defense and Earth-observation customers already pay roughly a thousand times more per GPU-hour than terrestrial inference — and they discard 90%+ of the data they collect because there's nowhere on the satellite to process it. Edge inference at orbit is a real product today, not a 2030 thesis.^[5]