SpaceX, xAI, and the orbital compute bet
Every AI lab has the same three walls: chips, power, and cooling. xAI's answer to the first one was Colossus, a supercluster built at a speed that embarrassed the industry. The answer to the other two, if you believe the trajectory Musk keeps sketching, involves SpaceX. The cheapest place to run a solar-powered computer, once launch is cheap enough, might not be on Earth at all.
I've spent my career being paid to distrust sentences like that one. So let's take it apart properly.
The part that's just physics
In the right orbit, a solar panel works around the clock: no night, no clouds, no atmosphere eating your flux. Power is the binding constraint on AI buildout; I wrote about the five orders of magnitude between today's clusters and where the scaling curves point, and nobody knows where the terrestrial grid fits the next two of those orders. Space solves power elegantly.
It un-solves everything else. Cooling without convection means radiators, and radiators are heavy. Radiation flips bits and ages silicon, so you either shield (heavy) or design tolerant systems that cost you density. Maintenance is a launch, not a technician with a cart. Bandwidth back to Earth exists (Starlink's laser mesh is real engineering, not slideware), but latency and throughput make orbit a place for training runs and batch inference, not for serving your chat app.
Orbital compute is its own kind of machine, not a data center parked in space, and it runs the same weights only for workloads that don't mind being far away.
The part that's economics
The whole case rises or falls on cost per kilogram to orbit, which is exactly the number SpaceX has spent two decades attacking. If Starship hits full reuse at the flight cadence they're aiming for, the launch line in the spreadsheet stops being the dominant term, and the comparison becomes orbital solar versus terrestrial grid interconnect queues, which in 2026 stretch years in every market that matters. That queue, more than any keynote, is why serious people run these numbers at all. The training-versus-serving split I described in the inference economics post maps cleanly: training tolerates distance, serving doesn't.
The xAI angle makes the incentives unusually aligned. One founder controls the model lab that's compute-starved, the launch company with the price curve, and the constellation with the backhaul. Vertical integration like that built Colossus in months; it could plausibly put a training pod in orbit before consensus says it's sensible. Whether Grok 4.5's successors get trained under a solar array in sun-synchronous orbit is a bet on execution, not on physics.
The part that's theater
Some of it is theater, and it's fine to say so. Renders of orbital gigawatt farms arrive on schedule whenever a funding round does. The gap between a demonstration rack and an economically load-bearing cluster is the same gap fusion has lived in for fifty years: real physics, real progress, perpetually eight years out. And every kilowatt-hour claim deserves the discount you'd apply to any number that has never met an auditor.
My honest position: this is the rare moonshot where the skeptic's math and the promoter's math use the same variables and disagree only on dates. Watch launch cadence and interconnect queues (the two curves that decide it) and ignore the renders. As habits go, that one travels well beyond rockets.