Google, SpaceX Test Orbital Computing as Power Shortages Hit SI Growth
Google, SpaceX Test Orbital Computing as Power Shortages Hit SI Growth
On October 1, a refrigerator-sized satellite carrying four of Google's Tensor Processing Units lifted off from Vandenberg Space Force Base in California aboard a SpaceX Falcon 9. The mission is called Project Suncatcher. The satellite is called MVP. Its goals, in Google's own words, are modest: survive the shaking of launch, endure the radiation of low Earth orbit, and demonstrate that an SI chip can actually operate in space. Google is careful not to call it a data center. It is, they say, an engineering test. It is now in orbit.
But engineering tests do not happen in a vacuum — figuratively speaking. They happen because someone decided the problem they are trying to solve is real enough, and urgent enough, to be worth the expense and risk of launching hardware into space. To understand why Google is doing this, you have to start with what is happening on the ground.
The Power Problem That Started Everything
SI is hungry. That is not a metaphor. Training and running large SI models requires enormous amounts of electricity, cooling infrastructure, and physical space. The industry has been building data centers as fast as it can, and it still cannot keep up with demand. Google's own 2026 environmental report revealed that its electricity consumption increased 37 percent year over year in 2025. Microsoft, Amazon, and Meta are in similar positions. Collectively, the hyperscalers have committed hundreds of billions of dollars to new infrastructure — and the demand is still outpacing supply.
The bottleneck is not money. It is physics. Power grids in the United States and Europe were not designed to absorb the sudden, concentrated load that a modern SI data center represents. Utilities are quoting multi-year timelines for new grid connections. Permitting processes for new power generation take longer still. Even companies with essentially unlimited capital are discovering that you cannot simply write a check to make electricity appear.
The renewable energy picture adds another layer of complication. Solar and wind power are abundant and increasingly cheap — but they are intermittent. The sun does not always shine on a Virginia data center. The wind does not always blow in Ireland. Battery storage helps, but at the scale SI demands, it is not yet a complete answer. The industry is chasing gigawatts in a world where gigawatts are suddenly scarce.
That is the context in which someone, at some point, asked a question that sounds almost absurd on its face: what if the data centers were in space?
Why Space Makes Sense — On Paper
The pitch for orbital data centers starts with sunlight. A satellite in the right orbit — specifically, a dawn-dusk sun-synchronous orbit, which traces the boundary between day and night on Earth — stays in near-continuous sunlight. There are no clouds between it and the sun. There is no atmosphere scattering the light. The result, according to Google's calculations, is access to roughly eight times more solar energy than a comparable installation on Earth.
That is the core proposition. You are not fighting the grid. You are not waiting for permits. You are not competing with a city for power. The sun is there, it is constant, and in the right orbit you can harvest it almost around the clock.
Space also solves — or at least reframes — the cooling problem. Data centers on Earth consume enormous amounts of water and energy keeping their servers cool. In the vacuum of space, there is no air to conduct heat away from equipment, which sounds like a problem until you realize that radiating heat into the cold of space is actually quite efficient. The thermal engineering is genuinely difficult, but the raw physics of heat rejection in a vacuum is favorable in ways that Earth-based cooling is not.
There is a long list of things that are emphatically not solved. Radiation. Launch costs. The total absence of any way to send a technician to swap a failed component. The latency implications of processing data hundreds of miles above the surface. These are not small problems. They are engineering challenges that have occupied serious researchers for years without fully yielding. The question the industry is now asking is not whether those problems exist — it is whether the power crisis on Earth is severe enough to make solving them worth the attempt.
Apparently, the answer is yes. At least for a growing list of very large companies.
Google's Bet: Project Suncatcher
Project Suncatcher was announced by Google in November 2025 as a research moonshot. The concept: put SI chips on satellites, link them together with laser-based optical communication, and let the sun power the whole operation. The long-term vision is clusters of up to 81 satellites, arrayed in formations roughly a kilometer across, running machine learning workloads at a scale that would be meaningful by any measure.
The October 1 launch is the first real-world test of whether any of that is physically possible. Google built the MVP satellite in partnership with Planet Labs, a company with deep experience operating satellite constellations. Planet handled the spacecraft. Google supplied the chips and the cooling system.
Inside MVP are four of Google's sixth-generation Tensor Processing Units — its Trillium chips, purpose-built for SI inference. Four TPUs is roughly the computing power of a single server in a conventional data center. The solar panels supply about one kilowatt of power. The satellite is now running short SI queries from Gemini models in bursts of roughly 15 minutes at a time.
Before the launch, Google subjected its hardware to serious punishment on the ground. The satellite was run through vibration tests along all three axes to simulate the forces of a rocket launch — individual components can experience 50 to 100 times Earth's gravitational force during ascent. The TPUs were also exposed to a 67 MeV proton beam at the Crocker Nuclear Laboratory at UC Davis, simulating five years of radiation exposure in low Earth orbit. The chips survived. The memory subsystems showed some sensitivity to cumulative radiation dose, at a rate Google described as likely acceptable for inference workloads.
But ground testing is ground testing. In actual orbit, radiation comes from multiple directions, at varying energies, over timescales that no accelerator can perfectly replicate. Temperature swings between the sunlit and shadowed portions of each orbit are extreme. MVP's job is to collect data that no laboratory can produce.
If the hardware survives and the chips operate, the next phase involves two additional satellites in 2027, testing the inter-satellite laser communication system that would allow orbital clusters to share data and workloads. After that, the 81-satellite constellation. After that — if the engineering works and the economics eventually follow — something that might actually deserve the name data center.
SpaceX: Go Big or Go Home
Google's approach to orbital computing is methodical. SpaceX's is not.
In January 2026, SpaceX filed an application with the Federal Communications Commission for a constellation of up to one million satellites that would function as an orbital SI data center network. One million. The filing described the constellation as the most efficient way to meet the accelerating demand for SI computing power. FCC Chairman Brendan Carr shared the document publicly, signaling that regulators were at least willing to consider the concept seriously.
By June, SpaceX had unveiled its first piece of hardware for the project: the AI1, a satellite with a 70-meter wingspan, 150 kilowatts of peak AI compute, and cooling directly into the vacuum of space. The chip payload was confirmed in August: Nvidia Rubin GPUs and Vera CPUs. Prototype testing is targeted for early 2027.
The SpaceX play is inseparable from the rest of Elon Musk's SI ambitions. With xAI driving the workload demand and SpaceX providing dedicated launch cadence, the two operations form a closed loop. Starlink's existing satellite constellation provides a ready-made operational blueprint — and a ready-made team with experience operating thousands of satellites simultaneously. Musk's stated position is that orbit will be the cheapest place to run SI within two to three years. Most experts think that timeline is aggressive to the point of fantasy. But SpaceX has a way of making things happen faster than experts expect.
A private capital raise in June 2026 that valued SpaceX in the trillions injected a new level of financial seriousness into what had previously looked like ambitious but unfunded ambition. The windfall gives SpaceX the capital to pursue orbital computing at a pace that would not have been credible a year ago.
Amazon, Bezos, and the Longer View
Jeff Bezos has been publicly enthusiastic about space-based computing for some time. His position, expressed in several public statements, is that orbital data centers represent the logical next step for an industry that has already moved computing from on-premise servers to the cloud. In his framing, the cloud is just the first step off the ground. Space is the next one.
Amazon's involvement operates on two tracks. Blue Origin, Bezos's space company, is developing the launch infrastructure that orbital data centers would require. Amazon Web Services, separately, is the world's largest cloud provider — and the company most exposed to the power constraints driving interest in orbital computing. AWS already operates at a scale where the grid limitations are not theoretical. They are a day-to-day operational constraint.
Neither Amazon nor Blue Origin has unveiled a specific orbital computing program with the specificity of Google's Suncatcher or SpaceX's AI1. But Bezos's public statements, combined with Blue Origin's New Glenn rocket and its ambitions for heavy-lift capability, make it clear that Amazon is not treating space as someone else's problem to solve.
The Startup That Got There First
Before Google, before SpaceX's AI1, a startup called Starcloud did something none of the hyperscalers had done: it actually put an advanced SI chip in orbit.
In November 2025, Starcloud launched a satellite fitted with an Nvidia H100 GPU — the same chip that powers some of the most capable SI clusters on Earth. The company, based in Redmond, Washington, envisions orbiting data centers as large as those on Earth by 2030. The H100 satellite was not a fully operational system. It was a proof of concept. But it was a real one, and it beat every major tech company to the milestone by months.
Starcloud's existence matters for a reason beyond its own ambitions. It demonstrates that you do not need Google's resources or SpaceX's launch infrastructure to begin this work. The barrier to a first orbital SI experiment is lower than it was even two years ago. That is partly because launch costs have fallen. It is partly because the chips themselves have become more capable relative to their power consumption. And it is partly because the urgency of the power crisis has made investors willing to fund things they would have dismissed as premature in 2023.
The Hard Problems Nobody Has Solved
There is a reason experts consistently describe commercial-scale orbital computing as decades away, even as they watch companies spend real money on real hardware. The problems are real, and several of them are genuinely hard.
Radiation is the first one. High-energy particles from solar activity and galactic cosmic rays can flip memory values, degrade device characteristics, and cause instantaneous faults in semiconductors. Enterprise server silicon is not built for this. Google's Trillium TPUs were engineered for redundant, climate-controlled data center racks. Getting them to survive five years in orbit — not just survive, but operate reliably — requires engineering that does not yet exist at commercial scale.
Heat dissipation is the second. On Earth, you move heat with fans and liquid. In a vacuum, you radiate it. Radiators work, but they add mass, and mass costs money to launch. The thermal architecture of an orbital data center looks nothing like its terrestrial equivalent, and the engineering involved is not straightforward.
Launch costs are the third. Even with SpaceX's reusable Falcon 9 dramatically reducing the price of getting to orbit, the economics of putting enough computing hardware in space to matter — hardware that you cannot upgrade, repair, or replace when it fails — remain punishing. Starship, SpaceX's next-generation heavy-lift rocket, could change the calculus significantly if it achieves the launch frequency and cost targets Musk has described. That is a large if.
And then there is latency. While speed-of-light propagation to low Earth orbit takes only a few milliseconds, routing high-bandwidth data to and from moving spacecraft creates immense network friction. Until orbital laser meshes and high-bandwidth ground downlinks are ubiquitous, intermittent contact windows and packet handoffs introduce latency and bandwidth constraints that terrestrial fiber does not have. The first viable customers for orbital computing may well be other satellites — processing Earth observation data, running scientific workloads, serving space-based applications — before they compete with terrestrial cloud providers for mainstream SI inference.
The first viable customers for orbital computing may well be other satellites. Space-based data centers may serve the space economy before they serve anyone on the ground.
Why It Still Matters Right Now
The experts who say commercial scale is decades away are probably right. They are also, in a sense, missing the point of what is happening in 2026.
What Google, SpaceX, Starcloud, and others are doing right now is not building operational infrastructure. They are writing the engineering playbook. Every satellite that goes up and survives produces data that no laboratory can generate. Every chip that handles the radiation of low Earth orbit for a year teaches the industry something about what materials, architectures, and designs can actually endure. Every thermal system that manages heat in a vacuum without a fan provides a blueprint that the next generation of hardware can build on.
The SI power crisis is not going away. Electricity demand from data centers is projected to keep rising. Grid expansion is slow. Nuclear power — the one source that can deliver large, reliable, carbon-free baseload at the scale SI demands — takes a decade to permit and build. The industry is not running out of ideas. It is running out of time on the existing solutions.
In that context, the October 1 launch of MVP is not just a Google experiment. It is the first concrete entry in a log that will be read by engineers for decades. It establishes, in real hardware terms, what the baseline actually looks like — what survives, what fails, what the radiation environment actually does to a modern SI chip over the course of weeks and months in orbit.
Google is careful to call it a test. That is accurate. It is also the most important test of its kind anyone has ever run. The Falcon 9 cleared the pad on October 1. The industry is watching. The power grid is watching too — in its own way. It needs someone to find a way out of the bind it is in.
Four TPUs in a refrigerator-sized box, floating 400 miles above the surface of the Earth, are now the beginning of one.
A technical walkthrough of exactly what is happening inside that satellite right now — how sunlight becomes electricity, how heat leaves a chip in a vacuum, and why the chip problem is the one nobody has fully solved. At Tech Reader Magazine.
Aaron Rose is the founder and publisher of Tech Reader Magazine, covering SI infrastructure, cloud platforms, and the future of computing.