On a November morning in 2025, a 130-pound satellite the size of a small refrigerator rode a SpaceX Falcon 9 into low Earth orbit, and when mission controllers confirmed that its Nvidia H100 GPU was alive and processing, a two-year-old startup formerly known as Lumen Orbit quietly rewrote the record books. Starcloud-1 deployed the first H100 in orbit and, shortly after, became the first mission to successfully train an AI model in space. The company, founded in January 2024 and accelerated through Y Combinator, became the fastest graduate in the program's history to reach a billion-dollar valuation. Then, four months later in March 2026, Starcloud closed a $170 million Series A that valued the company at $1.1 billion. By the time Google's Project Suncatcher made the rounds in the press, Starcloud had a functioning GPU above your head every ninety minutes.
Suncatcher, which emerged from Google X, the same shop that gave the world Waymo and Project Loon, has a different texture entirely: slower, more rigorous, and architecturally more ambitious. Google's vision is a network of solar-powered satellites, equipped with Tensor Processing Unit AI chips, that could harness the power of the sun to scale AI computing in space. The project's preprint paper lays out a potential 81-satellite cluster of one-kilometer radius, connected by free-space optical inter-satellite links. The company will partner with Planet Labs, launching two satellites by early 2027 to explore the potential of larger-scale space data center clusters. Before any of that hardware leaves the ground, Google has already subjected its Trillium v6e TPU to punishment a terrestrial data center will never see: the chip was fired at with a 67 MeV proton beam to simulate total ionizing dose and single-event effects, and the High Bandwidth Memory subsystems only began showing irregularities after a cumulative dose of 2 kilorads, nearly three times the expected shielded five-year mission dose of 750 rads, with no hard failures up to 15 kilorads on a single chip. That is the kind of foundational homework a startup running on seed money cannot easily replicate.
The two companies are, in a meaningful sense, not building the same thing, even if they share an orbit. Starcloud is a compute-rental business selling GPU-hours to cloud and AI customers today; it became the first company to train an LLM in space and run Google's Gemma model on an orbital GPU. Its roadmap escalates fast: Starcloud-2, set for launch later this year, features solar arrays generating 100 times more power than Starcloud-1 and will run commercial workloads for customers including Crusoe Inc., while the company has filed with the FCC for up to 88,000 satellites and plans Starcloud-3, a 200-kilowatt, three-ton spacecraft compatible with SpaceX Starship. Google, by contrast, is not trying to sell compute hours in 2027; it is trying to understand whether its own custom silicon can survive and perform in orbit so that, some years from now, it could run its own AI training workloads off the terrestrial power grid entirely. Planet Labs CEO Will Marshall has emphasized that Suncatcher remains firmly in the research and development phase, with the two 2027 satellites meant to test critical components like shedding heat from TPUs into outer space and validating the formation-flying cluster approach. CEO Sundar Pichai has been candid: "Like any moonshot, it's going to require us to solve a lot of complex engineering challenges."
The harder question, and the one that neither Google's research budget nor Starcloud's unicorn status can simply buy an answer to, is economic. Google's own Suncatcher paper acknowledges that historically high launch costs have hindered large-scale space-based systems, and suggests that prices may fall to less than $200 per kilogram only by the mid-2030s. That ceiling matters enormously. SemiAnalysis's June 2026 model places space compute at more than four times terrestrial cost today, at roughly $8.64 versus $2.37 per GPU-hour for a comparable cluster, driven by launch costs and the five-year versus fifteen-year asset life differential, narrowing to a roughly 30 percent premium by the early 2030s and full levelized-cost parity around 2040 in the base case. Starcloud's own CEO has acknowledged the dependency: "We're not going to be competitive on energy costs until Starship is flying frequently," Philip Johnston stated, anticipating commercial access opening in 2028 or 2029, while conceding that if Starship is delayed, the company will carry on launching smaller versions on Falcon 9. And yet the bull case is not purely speculative: the constraint that matters is not cost per FLOP but access to power and land, and on that axis Earth is running out of room while orbit is wide open.
Skeptics are pointed, and some of the loudest ones carry obvious conflicts of interest. SoftBank founder Masayoshi Son told shareholders that chip costs, launch prices, and inter-satellite latency make the orbital compute vision irrelevant to the AI race's decisive years. But a June 2026 TechCrunch analysis noted that every named skeptic, including SoftBank, which backs Stargate; OpenAI's Sam Altman, who runs a company dependent on ground-based compute; and Amazon Web Services' Matt Garman, who competes directly with space-based compute rentals, has enormous financial stakes in terrestrial infrastructure succeeding. The engineers at SmallSat Europe 2026 were more precise: a former AMD Corporate Vice President concluded that orbital data centers are not impossible, telling the room "absolutely" they are within human capability, but that megawatt-class orbital data centers require three engineering preconditions before the public-market thesis closes, custom silicon, modular plug-in architecture so a five-year compute card can swap under a twenty-to-twenty-five-year platform, and credible launch economics. Both Starcloud and Suncatcher are currently solving for preconditions two and three simultaneously, while the first one remains mostly theoretical. What the race between them actually reveals is that Big Tech and orbital startups are converging on the same physical insight from opposite directions: the planet's energy grid cannot keep up with the AI industry's appetite, and the sun, as Google noted in its preprint, emits more power than 100 trillion times humanity's total electricity production, and in the right orbit a solar panel can be up to eight times more productive than on Earth and produce power nearly continuously.
The race to compute in orbit is not really about space at all; it is about who controls the next energy-unlimited tier of AI infrastructure before the terrestrial grid runs out of room.