Google and SpaceX Are in Talks to Put Data Centers in Orbit
Google and SpaceX are reportedly in early talks to build orbital data centers, positioning space as a long-term home for AI compute workloads — despite current costs that dwarf terrestrial alternatives by orders of magnitude.
Original sourceAccording to a new report from TechCrunch, Google and SpaceX are in active discussions to explore placing data center infrastructure in low Earth orbit. The pitch centers on space as an eventual home for AI compute, leveraging SpaceX's Starship launch cadence and reusability improvements to make the economics of orbital deployment less catastrophic over time. Neither company has confirmed terms, timelines, or whether a formal agreement is in place.
The underlying logic, at least in theory, is compelling for a narrow set of use cases: orbital data centers could operate closer to satellite-based sensors, reduce certain latency profiles for space-native applications, and sidestep terrestrial power and land constraints that are increasingly throttling AI infrastructure buildout on the ground. Google has been aggressively expanding its data center footprint globally to support Gemini and other AI workloads, and land, water, and power availability are real bottlenecks.
The economics, however, are not close to viable at scale today. Launching compute hardware into orbit costs orders of magnitude more per unit than building equivalent on-the-ground capacity, and the challenges of cooling, radiation hardening, and maintenance in a vacuum are unsolved at data center scale. SpaceX's improving launch economics are real, but the gap between 'cheaper than before' and 'cheaper than a data center in Iowa' remains enormous.
This is best understood as an early-stage strategic alignment — two companies with overlapping infrastructure ambitions testing whether a long-term bet on orbital compute makes sense to develop together. Whether it produces actual hardware in orbit or quietly fades into an NDA is an open question. The signal here is less about imminent deployment and more about where Google believes the outer boundary of its infrastructure roadmap might eventually extend.
Panel Takes
The Futurist
Big Picture
“The thesis here is falsifiable and specific: orbital compute becomes cost-competitive with terrestrial infrastructure within 10-15 years if Starship achieves its per-kg launch cost targets and AI workloads continue demanding more power than terrestrial grids can supply. Both dependencies are real and have clear failure modes — Starship economics could plateau, or grid buildout could outpace demand. The second-order effect nobody is discussing is that whoever controls orbital compute infrastructure controls a layer of AI capacity that is jurisdictionally ambiguous in ways that no nation-state has fully worked out yet, and that matters enormously for which regulatory regimes apply to frontier model inference at scale.”
The Skeptic
Reality Check
“The thing that kills this in 12 months isn't SpaceX's launch costs or Google's ambition — it's the cooling problem, which is not a footnote but the entire engineering challenge, and neither company has publicly demonstrated a credible solution for dissipating megawatts of heat in a vacuum at data center density. 'In talks' from two companies with aligned PR interests is not a product, a prototype, or a timeline; it's a press release with plausible deniability. I'll revisit this when there's a mass budget, a radiation-hardened hardware spec, and an actual maintenance plan for hardware that's traveling at 17,000 mph.”
The Founder
Business & Market
“The buyer here is Google itself, which means this is a vertical integration play, not a product — and the economics only pencil out if you believe Starship gets launch costs below $100/kg and stays there while Google's terrestrial power costs continue rising. The moat, if it materializes, is genuinely structural: an orbital compute layer that competitors cannot replicate without their own heavy-lift rockets creates a durable infrastructure advantage that no amount of VC funding can shortcut. The risk isn't the vision, it's the 10-year capex runway before any of this generates a return, and the assumption that AI compute demand doesn't shift to a form factor that makes orbital placement irrelevant.”
The Builder
Developer Perspective
“Speaking purely from an infrastructure engineering standpoint: the unsolved primitive here isn't the rocket, it's the ops model — how do you patch firmware on a server rack moving at orbital velocity with a 400ms round-trip latency to ground control, and who gets paged at 3am when a cosmic ray flips a bit in your memory controller? Until there's a credible answer to 'how does a human intervene on a hardware fault,' this is a research project with a very expensive delivery mechanism, not deployable infrastructure. I'd be far more interested in this conversation once someone publishes actual radiation tolerance benchmarks on commodity GPU hardware, because right now we don't even have the vocabulary for the failure modes.”