There are two opposite answers to the same bottleneck, and both are being sold as space. One puts the GPU in orbit and sends the answer down. The other leaves the GPU in a hall in Oregon or Texas and sends the energy in. Overview Energy, 27 April 2026: a deal to beam up to 1 gigawatt to Meta data centres when terrestrial solar is dark, collected in GEO, delivered as near-infrared into existing solar farms. In-space demo 2028, commercial 2030. Micah Walter-Range of Caelus, to SpaceNews: seems easier than putting the data centres in space. Starcloud and SpaceX are making the other bet.
Overview–Meta offtake shape, 27 Apr 2026, SpaceNews. In-space demo 2028. Not a joule on a farm yet. Walter-Range: “Seems easier than putting the data centers in space.”
You cannot hold both as the same trade.
Two answers to the same bottleneck
What Overview actually sold
The Meta sentence is an offtake shape, not a watt in the air. GEO collection, a laser into farms Meta already owns or contracts, night and weather as the product. That is a capacity-factor story. It does not require a radiation-hardened Vera Rubin. It does require beam safety, atmospheric loss, and a farm that will accept a laser as if it were the sun. The 2028 demo is the first object. The 2030 commercial date is a slide. Until a joule lands on a named Meta array and is metered, this is a power-purchase conversation with a space adjective.
The useful argument is the anti-orbital-hall one. If the scarce thing is electrons at night, you do not have to lift the chips. You lift the collector. The chips stay next to the people who repair them, next to the water, next to the last-mile fibre. GPU refresh stays on a truck. That is why Walter-Range's sentence belongs on this page. It is a logistics claim.
What the orbital-compute side is buying
Johnston's 21 August raise and Musk's 14 August post are the other logistics claim: permitting and public hostility will make new terrestrial halls illegal or late, so the only scale path is orbit. The PJM IRAS filing is the US version of that hostility, written as a tariff. Aetherflux's May pivot, from beaming at Earth toward Galactic Brain compute plus a rocket, is what happens when a company discovers that electrons on the ground are harder to book than a compute story Nvidia will join. Cowboy Space Corporation is still targeting a 2026 space-based solar demo sat. The equity story moved.
Google's Suncatcher is a third shape: TPU satellites in dawn-dusk sun-synchronous orbit, optical inter-satellite links, two prototypes with Planet by early 2027. It is two birds and a research blog. Treat it as a prototype, which is what the blog says.
If Overview's 2028 demo puts a measured beam on a Meta-affiliated array, the beam-down side has an object. If Starcloud-2 or a SpaceX Starmind node is running billed inference for a named customer in 2027, the fly-the-GPU side has an object. If neither happens and the FCC pile grows, we have a filings market. Both can work. They are not one theme. Beam-down is a night-time capacity-factor hedge that leaves Nvidia's customers on the ground. Fly-the-GPU is a permitting hedge that makes Nvidia's next package a spacecraft. Meta is on the first. Nvidia is on the second. SpaceX is trying to be the vehicle for both and the compute vendor for one.
X collapsed them into orbital AI. Do not. The laser that hits a farm and the laser that hits a downlink dish are different photons with different customers, different regulators, and different failure modes. Astronomy and aviation will show up for the beam. Debris and thermal area will show up for the hall.
Two photons are not one trade.
Satellite bus on the cleanroom floor.
Wafer and a radiator plate.
Launch-slot tags on an aluminum bus.