The Orbital Bid: Google, SpaceX, and Schmidt's Relativity Gamble for AI Compute in Space

The AI industry's most pressing constraint is not intelligence. It is electricity. Every frontier model, every agentic workflow, every multimodal generation consumes power at a scale that terrestrial grids were never designed to deliver. Data center operators are buying nuclear plants. Hyperscalers are partnering with utilities on gigawatt-scale commitments. And now, the biggest names in technology are looking up — not for inspiration, but for outlet strips in the sky.

In July 2026, Bloomberg reported that Google is in active discussions with SpaceX to provide launch services for Project Suncatcher, its orbital data center initiative. The talks confirm what has been visible since Google published its pre-print paper on the concept last year: the company intends to build a scalable compute system for machine learning in space, using fleets of satellites equipped with solar arrays, inter-satellite free-space optical links, and Google's own Tensor Processing Units. Planet Labs is manufacturing the prototype satellites. Two are slated for launch by early 2027. The full constellation, as described in Google's technical paper, would span 81 satellites across a 1-kilometer radius — an in-orbit compute cluster.

Why Space?

The logic is not as unhinged as it sounds. In a sun-synchronous orbit, a solar panel can be up to eight times more productive than on Earth. It produces power nearly continuously, eliminating the need for batteries and the thermal cycling that degrades ground-based systems. For AI workloads that demand constant, reliable compute capacity, the sun is an essentially inexhaustible power source that requires no grid interconnection, no permitting battle, and no community opposition hearing.

Google's own analysis, published in a blog accompanying the Suncatcher paper, argues that if launch costs continue their sustained decline through the mid-2030s, the cost of launching and operating a space-based data center could become roughly comparable to the reported energy costs of an equivalent terrestrial facility on a per-kilowatt-per-year basis. That is a speculative claim, built on an extrapolation of SpaceX's Starship cost curve. But it is not a frivolous one.

The Players

Google is not alone in this bet. The competitive landscape reads like a who's-who of billionaires and their rockets.

SpaceX has filed for approval to launch a constellation of up to one million satellites delivering 100 kilowatts of compute power per tonne, also in sun-synchronous orbit. Elon Musk's company has the launch infrastructure, the satellite manufacturing capacity, and the vertical integration that every analyst agrees is the prerequisite for making orbital data centers economically viable. SpaceX filed its application in January 2026, and while no target launch date has been disclosed, the company's own Starlink constellation already proves it can deploy and maintain orbital assets at scale.

Eric Schmidt, former CEO of Google, acquired controlling stake in Relativity Space in March 2025 for approximately $800 million, reportedly to put data centers in orbit. Relativity Space was on the brink of bankruptcy when Schmidt stepped in. The company's 3D-printed rocket architecture had failed to deliver on its promise, but its launch capability and manufacturing infrastructure were exactly what Schmidt needed. The move was telling: a man who spent a decade running the world's most sophisticated data center operator decided that the next frontier of compute infrastructure was not on Earth.

Starcloud, formerly Lumen Orbit, raised $170 million in Series A funding at a $1.1 billion valuation in March 2026, led by Benchmark and EQT Ventures. In November 2025, the company launched Starcloud-1, a satellite equipped with an NVIDIA H100 GPU, into low Earth orbit. Within a month, it was running Google's Gemini model on the chip and training Andrej Karpathy's nano-GPT LLM. Starcloud claims its orbital data centers will operate with 10x lower energy costs than terrestrial facilities. Starcloud-2, featuring a complete GPU cluster with persistent storage and thermal management, is targeted for 2027.

Axiom Space launched its first data center prototype aboard the International Space Station in August 2025, running cloud computing, AI, and cybersecurity workloads on Red Hat Device Edge. It was awarded up to $5.5 million from the Texas Space Commission and deployed its first two orbital data center nodes in January 2026. Lonestar Data Holdings is pursuing lunar data centers for disaster recovery. Jeff Bezos has said there will be gigawatt data centers in space within a decade through Blue Origin.

The Economics: Still in the Crater

For all the momentum, the economic case remains unproven. Andrew McCalip, an aerospace engineer and head of R&D at Varda Space Industries, modeled the full cost stack and found that 1 gigawatt of orbital solar compute would cost $51.1 billion versus $15.9 billion for the same capacity on Earth. That gap is not a rounding error. It is a factor of three.

McCalip's most important finding was not about launch costs or solar panel efficiency. It was about margins. If you have to buy launch, buy satellite buses, buy power hardware, buy deployment services, and pay margin at every interface, the margin stack and the mass tax eat you alive. Vertical integration, he concluded, is not a nice-to-have. It is the whole ballgame.

This is why the players are who they are. SpaceX has its own rockets, its own satellites, and its own constellation management. Google has its own chips, its own AI workloads, and a 6.1 percent stake in SpaceX. Schmidt bought a rocket company. Starcloud is the outlier — a startup trying to compete by assembling a supply chain rather than owning one — and its $1.1 billion valuation is a fraction of what vertical integration costs.

The Real Question

The orbital data center race is not really about space. It is about the limits of terrestrial power. The AI industry's compute demand is growing faster than grid capacity can expand. Nuclear deals, geothermal pilots, and behind-the-meter gas plants are all attempts to bridge the gap. Orbital compute is the most extreme version of the same logic: if you cannot get enough power on Earth, go where the power is free.

Whether the economics close by the mid-2030s, as Google projects, or remain a factor of three apart, as McCalip's model suggests, may be the single most consequential infrastructure question of the decade. If orbital compute becomes competitive, the entity that controls the cheapest launch wins everything — not just the space economy, but the AI economy, because it will have broken the power constraint that every terrestrial competitor still faces.

If it does not, the orbital data center will join the space elevator and the O'Neill cylinder in the archive of beautiful ideas that the physics and economics did not quite support. Either way, the bet is being placed now, by people with the capital to place it, and the rest of the industry should be watching. Because the constraint that drives AI compute into orbit is the same constraint that will drive every other creative and desperate solution to the power problem. The question is not whether the frontier of compute is moving. It is whether it is moving up.