Space Data Center $Trillion Bet: Startups to challenge SpaceX, Google, Blue Origin

In the last 90 days, something strange happened to the AI infrastructure conversation: it left the planet.

A dozen-plus organizations filed plans, launched hardware, or committed serious capital to building data centers in orbit. SpaceX and xAI merged into a $1.25 trillion company — and days before the deal closed, filed with the FCC for up to one million data center satellites. Weeks later, Jeff Bezos' Blue Origin filed its own plan for 51,600 satellites, turning the decades-long Bezos-versus-Musk rivalry into an orbital compute race. A Seattle-area startup hit unicorn status faster than almost anyone in Y Combinator history, on a pitch that sounded like science fiction 18 months ago. Google quietly published a paper concluding that machine learning in space is "not precluded by fundamental physics or insurmountable economic barriers."

So which is it — the next great platform shift, or a capital trap dressed up in solar panels?

This edition separates the three things that usually get blurred together: the hype, the physics, and the money. And it gives you the full landscape of who's actually building, what they've raised, who's backing them, and how far they are from delivering.

The case for space-based compute isn't really about space. It's about everything going wrong on the ground.

rrestrial AI buildouts are hitting three walls at once. Grid interconnection queues now stretch past 2030 in many regions. Community and permitting opposition has stalled or killed roughly half of planned U.S. AI data center capacity, by some industry counts. And gigawatt-scale campuses need staggering amounts of water for cooling.

The orbital pitch answers all three in one sentence: continuous solar power, radiative cooling with no water, no land acquisition, and no NIMBY in low Earth orbit. In sun-synchronous "dawn-dusk" orbits, solar panels can generate far more energy per panel than on Earth — no clouds, no atmosphere, no night.

Elon Musk has put a clock on it: within two to three years, he claims, the cheapest way to generate AI compute will be in space. Even Bezos — who just filed his own 51,600-satellite plan — thinks that's "a little ambitious," while agreeing the outcome is "very realistic." The skeptics go further: OpenAI's Sam Altman, Gartner, and short-seller Jim Chanos have variously called space data centers "ridiculous," "AI snake oil," and "peak insanity." Both camps are about to be tested by actual hardware.

Here's the field as of May 2026. I've grouped the players by how far along they actually are, because "announced a data center in space" covers everything from a GPU already flying in orbit to a PowerPoint with a constellation on it.

Raised: ~$200M total, including a $170M Series A at a $1.1B valuation (March 2026) Backers: Benchmark and EQT Ventures led the Series A; Y Combinator; NVIDIA Inception Scale target: 5-gigawatt orbital data center long-term; FCC filing for an 88,000-satellite constellation Distance from delivery: Already flying a demo

The technology: Starcloud's approach is to put data-center-class GPUs directly on satellites powered by large solar arrays, using the vacuum of space for radiative cooling. In November 2025 it launched Starcloud-1, carrying an NVIDIA H100 — a GPU more than 100x more powerful than anything previously operated in space — and became the first entity to train an AI model and run a version of Google's Gemini in orbit. Its roadmap scales aggressively: Starcloud-2 (targeted for late 2026) is meant to have ~100x the power generation of the first satellite, with later iterations using solar and cooling panels roughly 4 kilometers on a side to feed a 5 GW facility. It's the only company here with a paying-grade GPU already proven on orbit.

Raised: Not a fundraise in the normal sense — SpaceX and xAI merged in February 2026 into a ~$1.25 trillion entity, with a reported ~$1.5T IPO targeted Backers: Elon Musk's combined empire; internal capital Scale target: Up to one million data center satellites; a projected ~100 gigawatts of AI compute if one million tonnes of satellites are launched annually Distance from delivery: Filing and vision stage — no dedicated compute satellite flown yet

The technology: The strategy is vertical integration at planetary scale. SpaceX's entire advantage is launch cost — Starship is the enabler that could make hauling massive amounts of compute hardware to orbit economical. xAI brings the models and the demand. The January 2026 FCC filing covers satellites at altitudes between 500 and 2,000 km. Nothing compute-specific is in orbit yet, but no one else controls the launch vehicle, the chips' demand side, and the capital all at once. This is the bet that launch economics, not chip design, decide the race.

Raised: Internal (Jeff Bezos–funded; Blue Origin is privately held) Backers: Jeff Bezos; the company is self-financed rather than venture-backed Scale target: Up to 51,600 satellites in sun-synchronous low Earth orbit (500–1,800 km) Distance from delivery: FCC filing stage — application submitted March 19, 2026

The technology: This is the other billionaire mega-filing, and it puts Bezos directly head-to-head with Musk. Project Sunrise proposes a constellation of up to 51,600 sun-synchronous satellites to host AI and cloud workloads in orbit, connected via optical links that tie into Blue Origin's separately announced TeraWave satellite communications network (a ~5,400-satellite system for up to 6 Tbps anywhere on Earth). Blue Origin's FCC filing makes the now-familiar pitch: always-on solar energy, no land or displacement costs, and no grid constraints "fundamentally lower the marginal cost of compute" versus the ground. Notably, Bezos himself is the cautious voice here — in May 2026 he called the 2-to-3-year timelines floated by Musk "a little ambitious," while still insisting space-based compute is "very realistic" and "probably going to happen faster than most people think." Like SpaceX, nothing compute-specific is in orbit yet — this is vision and regulatory groundwork backed by a serious launch program.

Raised: Internal (an Alphabet "moonshot" out of X) Backers: Alphabet; hardware partner Planet Labs Scale target: 81-satellite clusters in ~1 km formations; economic viability projected around 2035 Distance from delivery: Two prototype satellites scheduled for early 2027

The technology: Suncatcher envisions fleets of solar-powered satellites carrying Google's own Tensor Processing Units (TPUs), linked by free-space optical (laser) inter-satellite connections to act as a distributed compute cluster. The satellites would fly in tight formation in sun-synchronous orbits for near-continuous power. Google has already done the homework that matters: it ran its Trillium-generation TPUs through a proton beam simulating five years of orbital radiation. The chips survived, though high-bandwidth memory (HBM) showed the most sensitivity — error rates "likely acceptable for inference," with training reliability still an open question. The 2027 demo with Planet will test heat-shedding and formation flying in real conditions.

Raised: $350M Backers: A range of investors including 1789 Capital (linked to Donald Trump Jr.); NASA partner Scale target: Orbital data center nodes attached to its commercial space station Distance from delivery: Early — groundwork phase

The technology: Axiom is best known for building a successor to the International Space Station and flying private astronaut missions (its fifth NASA-partnered mission is in the works). The data center angle is an extension: orbital compute nodes that lay groundwork for secure, space-based cloud computing co-located with its station infrastructure. It's less a pure-play orbital data center and more an existing space-infrastructure company adding compute to its platform.

Raised: $10M (February 2026) Backers: Alpha Funds, KDDI Green Partners Fund, Unlock Venture Partners Scale target: Modular compute tiles added to other companies' satellites Distance from delivery: Early-stage; pre-deployment

The technology: Sophia takes a different shape than the constellation builders. Its "TILE" platform is a solid-state, self-sustaining, radiation-resistant, vendor-agnostic compute module — designed to be bolted onto other operators' satellites rather than launched as a standalone data center. Founder and CTO Dr. Leon Alkalai pitches it as edge compute for the space economy: processing Earth-observation and sensor data on orbit to cut bandwidth and latency for time-critical uses like defense, disaster response, and maritime monitoring. Think "compute next to the sensor" rather than "hyperscaler in the sky."

Raised: Undisclosed (via a16z Speedrun, which typically writes ~$1M checks) Backers: Andreessen Horowitz (a16z Speedrun) Scale target: A LEO constellation of NVIDIA-powered compute satellites Distance from delivery: First test mission targeted April 2027

The technology: Los Angeles–based Orbital is building compute satellites, each housing a cluster of NVIDIA-powered servers with solar panels, for low Earth orbit. CEO Euwyn Poon frames the thesis bluntly: data center economics are dominated by electricity and cooling, and in orbit, solar power is continuous and cooling is fundamentally different. The company is opening an R&D and manufacturing facility ("Factory-1") in LA. Its first mission is explicitly a validation flight — proving sustained GPU operation in a high-radiation environment and handling "bit flips," the single biggest engineering headache in this whole category.

Raised: $233M+ across seven rounds, including a $92M Series C Backers: IA Ventures led the Series C Scale target: An optical data-relay network in orbit Distance from delivery: Operational comms layer, evolving toward compute support

The technology: Kepler isn't building the data centers — it's building the network they'll need. Its optical inter-satellite link infrastructure is the kind of high-bandwidth backbone that moves data between orbiting compute nodes and back to Earth. If orbital data centers become real, someone has to connect them; Kepler is positioning to be that layer.

Raised: $60M total ($50M Series A in 2025) Backers: Index Ventures and Interlagos (co-leads); Breakthrough Energy Ventures, Andreessen Horowitz, NEA; angels including Robinhood's Vlad Tenev and actor Jared Leto; plus DoD funding Scale target: Space-based solar power beamed to Earth — with an "orbital data center" extension targeted for Q1 2027 Distance from delivery: First solar-power demo planned, with the compute angle a later add-on

The technology: Founded by Robinhood co-founder Baiju Bhatt, Aetherflux is primarily a space solar power company — its satellites harvest solar energy and beam it to ground stations via infrared laser, roughly one kilowatt per beam in early form. The data center connection comes through its announced "Galactic Brain" project, which aims to run orbital compute powered by that same solar infrastructure. I'm including it because it's the clearest example of the power-first version of this thesis — but it's worth being honest that the data center piece is still downstream of getting space solar working at all.

A note on the numbers: The SpaceX/xAI, Blue Origin, and Google figures aren't "raised for orbital data centers" — they're corporate valuations or internally funded programs. Comparing them directly to a $10M seed round would be apples-to-orbital-oranges. And every figure here moves fast; treat this as a May 2026 snapshot.

This is where credibility is won or lost, so let's not hand-wave.

Radiation and bit flips. Charged particles strike chips and flip bits, corrupting computation. Google's own testing found HBM memory most vulnerable, with error rates probably fine for inference but unresolved for training. Orbital's entire first mission exists to validate radiation-hardening. This is the single hardest near-term problem.

Thermal management. On Earth you dump heat into air or water. In a vacuum, the only way to shed heat is to radiate it — which requires enormous panels. Starcloud's 5 GW vision needs cooling surfaces roughly 4 km on a side. Whether that scales is genuinely unproven.

Orbital congestion. Starcloud's 88,000-satellite plan is over 12x the size of today's operational Starlink fleet (~7,500 satellites). SpaceX's filing imagines a million. The collision-cascade risk in low Earth orbit is not theoretical.

Station-keeping. Tightly clustered formations — like Google's 1 km satellite clusters — need constant adjustment, which burns propellant and adds complexity.

None of these are obviously fatal. But none are solved, either.

One number decides everything: launch cost per kilogram.

The bull case cites solar power in orbit at as little as $0.005–0.05 per kWh and zero cooling water. Google's paper says the whole thing becomes viable only if launch costs drop below ~$200/kg by the mid-2030s — seven or eight times cheaper than today.

The bear case has a name and a calculation: Varda Space Industries estimates orbital compute currently costs roughly 3x more per watt than terrestrial equivalents. At today's prices, the math doesn't close.

For market sizing: orbital data centers are pegged at roughly $1.77B in 2029, projected toward ~$39B by 2035 at a ~67% compound annual growth rate. Big growth, small base.

So the real question isn't "will orbital data centers happen?" It's "which wave of deployment reaches cost parity with the ground first?" — and that answer rides almost entirely on Starship-class launch economics.

Here's why this belongs in a Physical AI newsletter and not just a space one.

Orbital data centers may be the most extreme Physical AI deployment environment that exists. You can't send a technician. Everything — power management, thermal regulation, fault recovery, station-keeping — has to be a closed-loop autonomous system reacting to a hostile physical environment in real time. On-orbit assembly and servicing will increasingly be robotic. And the most compelling near-term use case isn't hyperscale training at all — it's edge inference next to the sensor, like Sophia's pitch to process Earth-observation and SAR data in orbit instead of shipping raw bandwidth back down.

The companies that win here won't just be good at chips or rockets. They'll be good at autonomous physical systems operating where no human can intervene. That's Physical AI in its purest, most unforgiving form.

Strip away the trillion-dollar headlines and the field sorts into four honest tiers:

The hype is real but front-loaded. The physics is hard but not obviously impossible. The money is enormous but mostly betting on a launch-cost curve that hasn't happened yet.

My read: orbital compute is not a question of if but of when cost parity arrives — and that's a launch-economics question more than a chip or cooling one. Watch three signals over the next 18 months: the Starcloud-2 launch, the 2027 prototype missions from Google and Orbital, and every incremental drop in Starship's cost per kilogram. Those will tell you far more than any FCC filing.

Bull or bear on orbital compute reaching cost parity by 2030? Drop your take in the comments. 👇

If you found this useful, subscribe for weekly coverage of where Physical AI, space tech, and advanced manufacturing collide.