Google, SpaceX, and Now OpenAI Are Eyeing Orbital Data Centers. Here's What the Hype Gets Right — and What It Gets Very Wrong

The physics of space computing is real. The business case is far murkier than the headlines suggest.

The Story Everyone Is Running With

The Wall Street Journal reported that OpenAI CEO Sam Altman has explored discussions to invest in or acquire Stoke Space, a rocket manufacturer, with the goal of developing space-based data centers. The framing is irresistible: the AI industry's most prominent figure, reportedly trying to launch his own rocket company to compete with Elon Musk in aerospace. It sounds like the plot of a near-future thriller. That's precisely why it deserves a slower read. The source coverage here is thin — the Digitimes report citing WSJ is paywalled beyond the headline — so what follows is an honest separation of what the engineering actually supports versus what the coverage implies. Because orbital compute is a concept where the physics genuinely helps in narrow, specific scenarios, and is close to fantasy in the broad ones most headlines are imagining.

Where the Physics Actually Helps

There are real engineering arguments for putting compute in orbit, and they are worth taking seriously before dismissing the whole idea. First, cooling. On Earth, data centers burn enormous energy fighting heat. In the vacuum of space, radiative cooling is free — you point a radiator at the cold of deep space and shed heat without chillers, compressors, or water. For compute-dense workloads, that is a genuine thermodynamic advantage. Second, solar power. In low Earth orbit, a satellite in sunlight receives solar energy with no atmospheric filtering and no night cycle interruption for significant portions of its orbit. Power availability per square meter of solar panel is meaningfully higher than ground-based solar. Third, latency to satellite constellations. If your compute needs to talk constantly to a Starlink-type constellation, putting the processing node in orbit eliminates the ground round-trip entirely. For specific latency-critical applications, that matters. These are narrow, real advantages — not a general argument for moving the internet to space.

The Problems the Coverage Glosses Over

Every honest engineering conversation about orbital data centers runs into the same four walls almost immediately. Radiation hardening. Consumer-grade chips fail in the radiation environment of low Earth orbit within months. Radiation-hardened chips exist — they power satellites and military systems — but they cost orders of magnitude more per unit of compute than ground-based silicon, and they lag commercial performance by years. The AI compute revolution is built on cutting-edge consumer-derived GPUs. Those do not survive in orbit without massive redesign. Repair impossibility. When a server rack fails in a terrestrial data center, a technician walks over. In orbit, you write off the hardware. Redundancy requirements for orbital compute would be extreme, adding cost on top of already brutal launch economics. Launch cost per kilowatt of compute. Even with SpaceX's reusable rockets dramatically cutting launch prices, getting a meaningful amount of compute into orbit and keeping it there remains extraordinarily expensive per unit of processing power compared to just building another data center in a cold climate. Regulatory jurisdiction. Data processed in orbit exists in a legal grey zone. Which country's data sovereignty laws apply? This is unsolved, and for enterprise and government customers, it is not a minor footnote.

Why This Story Is Appearing Now

The timing of these reports is not accidental, and that context matters as much as the technology itself. Sam Altman exploring Stoke Space — a rocket manufacturer — places OpenAI in a direct competitive posture with Elon Musk's SpaceX at a moment when the relationship between Altman and Musk is openly adversarial. Whether the Stoke Space discussions represent a genuine infrastructure roadmap or a negotiating position designed to signal independence from SpaceX's launch monopoly is a question the available sourcing cannot answer. What the pattern does suggest: when AI's biggest players start talking about owning rockets, the story is rarely purely about the rockets. It is about controlling the full stack — compute, power, connectivity, and now potentially launch capability. Orbital data centers may be the stated destination. The actual near-term goal may simply be leverage.

What to Watch Next

The Stoke Space discussions, if they progress, will be the signal to watch — not because orbital data centers are imminent, but because who funds whom in the launch industry will shape which AI companies control their own infrastructure destiny. Watch whether any concrete technical specifications emerge from Google, SpaceX, or OpenAI about what orbital compute would actually look like — payload weight, orbit altitude, compute density. Vague announcements stay vague for a reason. Specific engineering commitments are where real infrastructure roadmaps begin. And watch the radiation hardening problem. Until someone credibly solves how to run modern AI accelerators in the space radiation environment at reasonable cost, every orbital data center announcement is a concept, not a product.