Google Is Putting AI Chips in Space to Break the Datacenter Bottleneck

The satellite that could change everything about AI infrastructure

Google is loading four of its newest AI chips onto a satellite and firing it into orbit next month. The hardware fits inside a refrigerator-sized box. The ambition behind it is anything but small.

The MVP satellite — short for minimum viable product — launches on September 1 aboard a SpaceX Falcon 9 as part of the Transporter-18 rideshare mission. Google built it with Planet, the Earth-imaging company, and it carries four TPU 8i chips, Google’s reasoning-specialized eighth-generation tensor processing unit. It will spend roughly a year in low Earth orbit runningGemini and collecting data on how AI silicon behaves when exposed to the space environment.

This is the first orbital test of Project Suncatcher, Google’s long-range bet that the future of AI compute doesn’t live on land.

What Google is actually testing

The MVP mission is not about running production AI workloads from space. It is about answering three blunt engineering questions: can TPUs survive launch, radiation, and vacuum long enough to matter?

The answers so far come from ground tests before the satellite even left the pad. Google subjected the hardware to vibration profiles mimicking the multi-G forces of ascent — up to 10G on the bus, with individual components seeing 50 to 100 times normal gravity. It fired a 600 MeV proton beam at a sixth-generation Trillium TPU at UC Davis’s Crocker Nuclear Laboratory to simulate years of orbital radiation exposure.

The results are mixed but directional. High-bandwidth memory began showing anomalies at a cumulative dose of 2 kilorads. A shielded five-year mission would see roughly 750 rads, so the margin is tighter than hoped. But the TPU chip itself survived up to 15 kilorads of total ionizing dose with no hard failure. That is useful news for hardware designers — the processors are tough. The memory is the fragile link.

Cooling is the harder problem. Space is a vacuum, so there is no convection. Google is using heat pipes and radiators to dump TPU waste heat into the environment, but the system has limits. On MVP, each TPU will run for about 15 minutes before throttling back to cool down. Cyclic operation under those constraints is not a path to the kind of sustained throughput datacenters require — but it is a path to data.

Why Google is doing this

The case for orbital compute rests on one simple arithmetic problem: AI is eating electricity faster than power grids can expand.

Google’s pitch for Suncatcher is that solar panels in low Earth orbit can harvest sunlight nearly continuously, achieving up to eight times the generating efficiency of equivalent ground-mounted arrays because there is no night cycle, no weather shadow, and no atmospheric attenuation at those angles. Pair that with TPUs and you get a compute node that does not compete with a city’s grid for megawatts.

The economics are speculative but framed concretely by Google. If launch costs continue declining toward the $200 per kilogram analysts expect by the mid-2030s, the cost of building and operating a space datacenter could approach the power costs of an equivalent ground facility. That is a threshold claim, not a budget. But if it holds, the argument shifts from engineering curiosity to competitive strategy.

Who wins and who loses

The semiconductor industry is the first obvious winner, at least in narrative terms. Google’s TPU program has always been designed to create internal demand outside the Nvidia cycle. Space is a new arena where that internal demand becomes the only demand — Google cannot buy its way out of radiation hardening or thermal design, it has to build its way out. That pulls more design cycles, more tape-outs, and more supply-chain attention toward custom AI silicon.

Launch providers win too. Transporter rideshare missions like the one carrying MVP are already a revenue model for SpaceX. Adding heavier, more complex payloads like MVP on top of the smallsat traffic expands the business line without requiring new rockets.

The losers are harder to name precisely because the project is early. But the ground datacenter model faces a real constraint: land, water, and power. Every major cloud builder is negotiating for gigawatts of electricity and thousands of acres. An orbital alternative, even a partial one, changes the negotiation dynamics over time.

The infrastructure gap: connecting satellites is the real challenge

Project Suncatcher’s end state requires dozens or hundreds of satellites talking to each other at datacenter speeds. Google is developing laser-based free-space optical communications to make this happen. Ground tests showed a single transceiver pair achieving 800 gigabits per second in each direction — 1.6 terabits total. That is competitive with some terrestrial fiber links and far above what most inter-satellite microwave systems deliver.

But ground tests and orbital tests are different. Atmospheric turbulence, point-and-track accuracy over hundreds of kilometers, thermal distortion of optical benches, and the sheer difficulty of maintaining a laser lock between two objects moving at seven kilometers per second are unsolved at scale. Google plans to launch two more satellites in 2027 to test exactly this — laser links between orbiting nodes alongside continued TPU validation. Those flights will tell us whether the concept survives contact with reality.

What happens next

The MVP data will feed into designs for the 2027第二批 satellites. If the 15-minute duty cycle under thermal limits proves manageable with better radiator area or more efficient chips, the path toward longer continuous operation opens. If HBM radiation sensitivity remains a hard limit, Google will need to decide whether to shield heavier, redesign memory architectures, or accept lower density per node and compensate with more nodes.

The wider implication is timing. Google is not racing to ship a space datacenter in 2027. It is racing to prove that one is physically possible and economically plausible by the mid-2030s — the same window in which ground-based AI buildouts will face their steepest power and permitting constraints. If the orbital model works, the company that gets there first reshapes where and how AI compute grows for the next decade.

The satellite carrying four TPUs toward orbit next month is small. The bet behind it is not.