On October 1, 2026, Google is scheduled to launch the first prototype satellite of Project Suncatcher, a research moonshot testing whether AI accelerators can survive and operate in low Earth orbit. The refrigerator-sized spacecraft, named MVP and built with Earth-imaging firm Planet Labs, carries four Google Tensor Processing Units (TPUs) that will run Gemini workloads in orbit. This is not an orbital data center. It is a hardware survival test, designed to answer three questions Google says can only be answered in space: will the chips survive launch vibration, will radiation corrupt their work, and how do you cool silicon in a vacuum where there is no air to carry heat away. The stakes extend well past Google, because a widening field of companies including SpaceX, Starcloud, Kepler Communications, and India's TakeMe2Space is betting that orbit becomes the next frontier for AI compute.
What Google is actually launching
The mission flies as a rideshare payload on SpaceX's Transporter-18, a Falcon 9 launching from Space Launch Complex 4E at Vandenberg Space Force Base in California. SpaceX lists a 58-minute launch window opening at 11:18 a.m. PT on Thursday, October 1, with a backup opportunity on Friday, October 2 at the same time.
MVP is small by data center standards and deliberately so. It carries four TPUs with roughly the compute of a single Google Cloud TPU v6e-4 slice, supplied by about one kilowatt of solar power, enough to run a microwave. According to Ars Technica, the plan had been to fly two custom satellites in 2027, but Google accelerated the schedule by integrating its AI hardware into a spacecraft Planet Labs had already built for an early test. The satellite will run short Gemini query bursts of roughly 15 minutes, then pause to cool, and operate for only a few months. Google describes the mission as a data-gathering exercise that will "gather in-orbit data and identify potential failure points, rather than demonstrate an operational orbital data center."
That framing is important. Google is not claiming orbital compute works yet. It is instrumenting the failure modes.
Why space is attractive for AI compute
The pitch for orbital data centers rests on three physics advantages that terrestrial sites cannot match. In low Earth orbit, satellites can access near-constant sunlight and generate up to eight times more solar power than comparable systems on Earth, according to Google. There is no atmosphere to absorb, scatter, or reflect incoming energy, and a dawn-dusk sun-synchronous orbit can collect power for most of the year. Space also removes the water problem entirely: waste heat is rejected through thermal radiation into a background near -270 degrees Celsius, consuming no water. And orbit has no permitting queue, no grid interconnection backlog, and no community opposition.
That last point is driving urgency. Hyperscalers plan to spend roughly $400 billion on terrestrial data centers in 2026, and interconnection queues in key markets such as Northern Virginia, California, and Germany stretch 7 to 12 years. U.S. data centers are projected to add about 240 terawatt-hours of demand by 2030, a 130 percent increase over 2024 levels, and the sector tracks toward consuming 8 to 10 percent of global electricity by 2030. For companies that cannot buy power fast enough on the ground, orbit looks less like science fiction and more like a pressure valve.
Google's own research blog frames the end goal bluntly: the Sun emits more power than 100 trillion times humanity's total electricity production, which is why the company argues space may eventually be the best place to scale AI compute.
The three problems Google must solve
Google's September 24 post and the mission reporting lay out the engineering hurdles in unusually candid detail.
Hardware survival. A rocket trip to low Earth orbit lasts about 10 minutes, during which the spacecraft experiences sustained acceleration up to 10 times the force of gravity, while individual components such as the TPU chips can see forces of 50 to 100 g. Google vibrated the satellite on all three axes to mimic launch frequencies. The team said it was "pleasantly surprised that the hardware held up to the force." Radiation is the second threat: solar events and cosmic rays can flip bits and degrade electronics. Google tested its Trillium TPUs in a proton beam facility at UC Davis's Crocker Nuclear Laboratory while running AI workloads. Initial results showed the chips survived a total ionizing dose greater than what a five-year space mission would deliver. Reuters notes that Google still wants orbital data because ground testing cannot fully replicate the real environment.
Cooling in a vacuum. This is the hardest problem. TPUs concentrate large amounts of heat in a small area, and in space there is no airflow, so heat can only be diffused through radiators. Google is testing a combination of heat pipes and radiators and has validated the approach in a thermal vacuum chamber that simulates the space environment. The chips run in roughly 15-minute bursts before shutting down to cool, a direct consequence of the thermal constraint.
Satellite interconnectivity. Future Suncatcher designs call for satellites that each carry dozens of TPUs, flying in clusters. Keeping them coordinated requires high-bandwidth laser links with extraordinary pointing precision, which Google compares to "hitting a coin-size target from miles away while both points are in motion." The company modelled clusters of up to 81 satellites at roughly 650 kilometers altitude and demonstrated 1.6 terabits per second across a single transceiver pair in the lab. The real orbital test of two linked satellites is planned for 2027.
The economics reality check
The physics may permit orbital compute, but the accounting is brutal. The most cited public model comes from aerospace engineer Andrew McCalip at Varda Space Industries, who built a calculator comparing orbital and terrestrial data center costs. His headline finding: orbital compute costs roughly three times more per watt than ground-based alternatives. In an interview with IEEE Spectrum, McCalip said his early rough calculations suggested a 7 to 10 times premium and that the idea "just wasn't practical. Not even close." Revisiting the numbers using public information about Starlink and Tesla technology brought the premium down to about 3x, which at least "nudges the concept out of the instantly dismissible category." His summary: the physics does not immediately kill it, but the economics are savage.
Concretely, IEEE Spectrum's illustration of a one-gigawatt orbital network assumes about 4,300 satellites, each with a 1,024 square meter solar array generating 250 kilowatts and carrying at least 175 GPUs. The network would cost around $51 billion including launch and five years of operations, versus about $16 billion for a comparable terrestrial system.
Consulting firm BCG reaches a similar conclusion from a different direction. Its 2026 analysis puts the current orbital cost premium at 2.5x to 3x, with a 20-year total cost of ownership of roughly $660 million to $750 million per megawatt, against $230 million to $300 million per megawatt for terrestrial data centers. BCG projects the premium narrows to about 1.5x over the next decade but does not disappear, concluding that space-based data centers could become technically feasible at scale within five to ten years but are unlikely to compete with terrestrial infrastructure on cost.
The wildcard is launch cost. McCalip's model and related analyses note that everything hinges on getting mass to orbit cheaply. At Starship's target of $100 per kilogram, space-based compute stops being 10 times more expensive; at the longer-term aspirational $10 to $20 per kilogram, the calculus shifts further. That dependency is why so much of the debate reduces to a single variable.
Who else is racing
Google is a late mover in a crowded field. SpaceX merged with xAI on February 2, 2026, creating a combined entity valued at roughly $1.25 trillion, and filed an FCC application on January 30 for up to one million orbital data center satellites at altitudes between 500 and 2,000 kilometers. The filing projects that launching one million tonnes of satellites annually would generate 100 gigawatts of AI compute capacity.
Starcloud, formerly Lumen Orbit, put the first Nvidia H100 GPU in space in November 2025 and trained a language model in orbit in December 2025. It filed separately for an 88,000-satellite constellation, with Starcloud-2 planned for October 2026 carrying Nvidia Blackwell hardware. Kepler Communications launched 10 optical relay satellites in January 2026, and Axiom Space deployed two orbital data center nodes on the same flight. Aetherflux, founded by Robinhood co-founder Baiju Bhatt, raised a $50 million Series A for a plan that combines orbital compute with power beaming to Earth.
The same rocket carrying Google's MVP illustrates how broad the field has become. Hyderabad-based TakeMe2Space is flying MOI-1A on Transporter-18, a sub-50 kilogram satellite carrying Nvidia Orin NX edge processors with about 150 watts to spend. The company describes it as India's first orbital computing satellite. Instead of beaming down raw Earth-observation data, MOI-1A lets customers upload containerized AI models that process imagery in orbit and transmit only the results. TakeMe2Space says 23 customers have booked space on the mission, spanning mapping firms, schools, and businesses in farming, mining, supply chains, and insurance. The satellite is a replacement for MOI-1, lost on January 12 when the third stage of ISRO's PSLV-C62 failed. Founder and CEO Ronak Samantray drew a clear line between his 150-watt edge computer and the high-power Nvidia chips that Starcloud and SpaceX plan to fly, a distinction that captures the difference between orbital edge inference and true orbital data centers.
The governance gap
The technical and economic debates have a third leg: policy. A Brookings commentary by COMPASS Fellow Lauren Tokos argues that the feasibility gap itself is a governance risk. While researchers question whether large orbital data centers can be cooled at scale, some scientists calculate that expelling the heat from a single orbital data center through thermal radiation alone would require about 2.15 million square feet of radiators. Ultraviolet exposure can degrade that radiator infrastructure, limiting performance and lifespan, and orbital debris risk rises with every added satellite.
On the regulatory side, Brookings points out that authority is split across the FCC for spectrum and licensing, the FAA for launch activity, the NTIA for interagency liaison, and a supporting cast including the FTC, the White House Office of Science and Technology Policy, and NASA. Tokos argues these silos stagnate policymaking and "enable unfettered technological expansion without proper oversight," and warns that orbital data centers could become an extension of existing terrestrial monopolies rather than a redistribution of compute power. Rest of World has made a parallel point: without shared governance, orbital compute could leave developing countries as data suppliers and AI consumers while control and value stay concentrated in the U.S. and China.
What to watch next
The near-term signal is narrow and concrete. If Transporter-18 flies on schedule, Google gets its first real radiation and thermal data from orbit, feeding directly into the two-satellite laser-link test it has planned for 2027. If the Falcon 9 slips to the October 2 backup window, the same hardware waits another day on the pad, a reminder that even the most advanced compute roadmap still depends on a launch schedule.
The larger question is whether orbital compute crosses from engineering experiment to commercial reality before the terrestrial power crunch forces the issue. The physics is credible. The economics carry a 2.5x to 3x premium today that BCG expects to narrow but not vanish. SpaceX's one-million-satellite filing and Starcloud's 88,000-satellite proposal describe scale that would reshape the global compute map, yet neither has demonstrated a working gigawatt-class orbital data center. Google's MVP satellite will not settle that debate. It will, however, produce the first hard data from a major tech company on whether its AI silicon works in space at all, and that is the threshold every orbital compute claim downstream depends on.
Sources: Google Research, Reuters, Ars Technica, SpaceX Transporter-18 mission page, IEEE Spectrum, BCG, Brookings, Introl, The Next Web, Gadgets Now.