News & Events

Data Centers in Space: 16 kW Up There, One Million Satellites on File

October 29, 2026
By Jon Billow

October 29, 2026

By Jon Billow, CRO, BNS

Every AI model you use runs in a building somewhere, and those buildings are running out of places to plug in. Wood Mackenzie has global data center power demand going from about 460 TWh this year to 3,700 TWh by 2040. On the ground, every new campus needs a substation, transmission, generation and a spot in the utility/interconnect queue that can run for years (plus a town council that may or may not want you there). So SpaceX, Google, NVIDIA and a pack of startups have started pitching a way around all of it: put the servers in orbit. A satellite carries the GPUs, a solar array to power them and radiators to dump the heat. Up there the sun shines almost around the clock, there’s no grid to tie into and no permit to pull. Data goes up over a radio or laser link, the work gets done in orbit, and the answer comes back down. A year ago that pitch got eye-rolls in my business. Then the hardware started flying – literally.

In November 2025 a 60 kg satellite called Starcloud-1 went up with a single NVIDIA H100 on board, trained a small GPT-style model and ran Google’s Gemma in orbit. Eleven months later SpaceX has folded xAI into itself, filed with the FCC for up to one million data center satellites (sorry stargazers!), and pitched orbital compute to its IPO investors. Google has Project Suncatcher. Starcloud is valued at $2.3 billion.

Short version: running models on data that’s already up there (radar imagery, defense sensing, other satellites) works today. Replacing a 500 MW campus in Texas takes launch at a few hundred bucks a kilogram, radiators nobody has flown at scale, and a repair model that doesn’t exist yet. I’d put that in the 2030s at the earliest, for a slice of the load. That said? Yeah, it’s coming. My grandkids probably need to learn to be astronauts, so they can actually work as data center techs.

So – How mature is the tech?

What has actually flown as of late October 2026:

  • Starcloud-1 (Nov 2025): one H100 in a 60 kg satellite. First data-center-class GPU in orbit.
  • China’s Three-Body Computing Constellation (ADA Space and Zhejiang Lab): 12 satellites launched May 14, 2025, rated at 744 TOPS each and about 5 POPS combined. The plan is 2,800 satellites, with 100 targeted by 2027.
  • Axiom Space: two orbital data center nodes launched January 11, 2026, riding Kepler’s optical relay network at 2.5 Gbps.

What’s next: Starcloud-2 (Blackwell GPUs plus an AWS server blade) was slated for October 2026, and recent reporting has it sliding into 2027 as two satellites at about 8 kW each. Google and Planet Labs plan two prototype TPU satellites around 2027. SpaceX’s IPO filing targets a 1 MW orbital data center in 2028. NVIDIA announced a space-rated Space-1 Vera Rubin Module at GTC in March.

Add it up and the biggest thing flying or about to fly is measured in kilowatts – nowhere sufficient to actually replace anything terrestrial. It’s a great proof of concept, but it won’t move the needle re: capacity demand. Starcloud-2 as reported – is 16 kW, that’s equivalent to one rack of classic IT equipment, and will not come close to a GPU cabinet of gear. One GB200 NVL72 rack on the ground draws roughly 120 kW, and a single AI hall is 30 to 100 MW. Every gigawatt plan out there also assumes Starship flies like a freight airline. SpaceX talks about one flight an hour carrying 200 tons. Starship made orbit for the first time this fall (Flight 14, 26 V3 Starlinks).

The technical problems

Heat. Space is cold. It’s also a vacuum, so there’s no air or water to carry heat away. The only way out is radiation off a panel. ABI Research ran the numbers in IEEE Spectrum this June: one H100 rack needs about 80 square meters of radiator, roughly a pickleball court. A 100 MW site needs about 2,500 of those courts. The coatings degrade, so you launch about 40% more radiator than day one requires. On the ground a liquid loop dumps heat into a cooling tower. In orbit your cooling tower is a football field of panel pointed at nothing, and it all has to be launched, unfolded and kept edge-on to the sun.

Radiation. No atmosphere means bit flips and slow damage. H100s and Blackwells are commercial parts with no radiation hardening. Inference can shrug off an occasional error. Training runs every chip in lockstep, so a flipped bit can poison the run; the fix is triple redundancy (three times the hardware) or eating the errors. SemiAnalysis assumes orbital operators carry about 20% spare GPUs versus about 5% on the ground.

Power. Sunlight in orbit is 1,361 W per square meter, and a dawn-dusk sun-synchronous orbit sees the sun almost continuously. No night, no clouds, no fuel bill. The catch: solar arrays and radiators together are often 65 to 70% of a satellite’s mass, so every extra kilowatt gets paid for twice in launch mass.

Networking. Training needs a fabric that moves hundreds of gigabits per port in microseconds. Google’s answer is to fly satellites a few hundred meters apart and push free-space optics hard; its bench test hit 1.6 Tbps. The operational link Axiom flies today is 2.5 Gbps. Then everything has to come down through ground stations, and optical downlinks don’t love weather (ask anybody on the Oregon coast about cloud cover).

Support: there’s no truck roll

BNS builds the power and fiber AI sites run on, and what keeps a hall up day to day is the truck roll: a tech with a spare optic, a PDU, a pump seal. In orbit there is no truck roll. Robotic servicing is still in development. A failed GPU, power supply or pump stays failed, so you overbuild at launch and replace whole satellites. Forethought estimates losing around 9% of compute per year with no repair.

On top of that: collision-avoidance maneuvers, a deorbit plan for every unit (the FCC’s five-year rule for low orbit), a ground segment with landing rights in every country you serve, and software updates that can brick hardware you’ll never touch again. Somebody gets to write the MOP for a firmware push to a box doing 17,000 mph.

Serious or PR?

Both.

Serious: Starcloud has raised about $450M and has hardware in orbit. Google published a real engineering paper and hired Planet to build the satellites. NVIDIA is building a space-rated module. China has 12 computing satellites up and ITU-approved spectrum for the rest.

PR: Musk’s not wrong that sunlight in orbit is free and nearly constant. But SpaceX filed for one million data center satellites on January 30, announced the xAI merger days later, and made orbital compute a centerpiece of its IPO story. In July, Kalshi traders gave a 1 MW orbital data center about a 14% chance of going live before January 2029. Meanwhile Anthropic agreed in May to pay SpaceX $1.25 billion a month through May 2029 for capacity at its ground-based Colossus data centers. So the loudest orbital pitch in the market is collecting roughly $15 billion a year renting out compute on the GROUND.

How the economics would work

The bull case is simple. No utility, no queue, no permit, no fuel bill, no town council. You pay once to launch and the sun is free. The bear case is everything you pay for to get there and keep it running.

The technicals (napkin math, using the bulls’ own numbers):

  • SpaceX claims about 100 kW of compute per metric ton of satellite. So 1 MW in orbit is about 10 tonnes.
  • 1 MW on the ground for a year is 8,760 MWh. At $0.08/kWh that’s about $700K a year, or $3.5M over a five-year GPU life. At $0.20/kWh it’s $8.76M.
  • Launching 10 tonnes costs about $30M at $3,000/kg (Falcon 9 territory today), $5M at $500/kg (Starcloud’s stated parity point) and $2M at $200/kg (Google’s mid-2030s target).
  • Break-even on electricity alone lands around $350/kg with $0.08 power and around $875/kg with $0.20 power.

And that’s before the satellite bus, arrays, radiators, the radiation tax, the 20% spare pool, ground stations and insurance. To be fair, the ground site carries costs orbit skips: substation, switchgear, generation, cooling plant, land. Those don’t close a 10x gap.

The published full-cost models are all over the place. Wood Mackenzie puts a 1 GW orbital data center at about $170 billion, over three times a terrestrial build. Andrew McCalip at Varda gets $42.4B versus $15.9B on the ground. ABI says at least 10x the cost per GPU-year even at an optimistic $44/kg launch. Forethought, the most bullish serious analysis I’ve seen, says orbit could be cost-competitive within three to five years if Starship stays on track, and still doesn’t expect a meaningful share of compute before 2030. That’s a 10x spread between people who all did real math.

Where it pays first: processing data that’s already in orbit. A radar satellite generates far more data than it can cheaply downlink, so running the model in space and sending down the answer cuts hours to minutes. Defense sensing is the same story. The third case is the buyer whose ground option is “no power until 2031.” Some of them will pay a premium for capacity that skips the queue, if it ever ships at scale.

How would sovereignty work?

Orbit isn’t a data haven. The rules are 60 years old and clearer than the hype suggests:

  • Outer Space Treaty, Article II: no nation can claim orbit.
  • Article VIII: the country that registers a satellite keeps jurisdiction and control over it. A US-registered satellite is US-jurisdiction hardware even while it crosses Germany roughly 16 times a day.
  • Article VI: that country must authorize and supervise its private operators, and it’s on the hook internationally for what they do.
  • The CLOUD Act already reaches data in a US provider’s control regardless of where it sits. Orbit doesn’t change that.
  • GDPR still follows EU residents’ data. The EU’s draft Space Act (Article 27) would require space data offered in the EU to come from objects on an EU registry.

There’s a live proposal for a “digital flag state” rule: treat the satellite as its registry’s soil and apply national data laws at the ground stations where data enters and leaves. Every bit comes up and down through a gateway licensed by some government, so landing rights are the control point.

Two consequences. First, a country that wants sovereign AI compute won’t get it by renting a US-registered constellation; it’ll want its own registry and its own birds, which favors state-backed programs in China and Europe. Second, on government & defense work the first questions from any contracting officer will be who holds the command uplink keys and where the ground stations sit. GPUs come up later.

Other considerations

  • Debris. About 11,000 Starlinks are up today. A million more satellites, each with big radiators and arrays, raises collision odds and the Kessler cascade risk for everyone in low orbit.
  • Astronomy. Acres of reflective panel do the night sky no favors, and astronomers are already fighting the current constellations.
  • Carbon. The EU-funded ASCEND study (Thales Alenia Space) found orbital data centers only beat the ground on emissions if the launcher puts out less than 370 kg of CO2 per kg delivered.
  • Single-vendor risk. Nearly all of this depends on SpaceX launch capacity, and SpaceX is also a competitor.
  • Export controls. Advanced GPUs in orbit are dual-use by definition, which matters more for China’s constellation than anyone’s press release admits.
  • The ground build doesn’t go away. Ground stations, fiber backhaul and the terrestrial clusters that train the models all still need a substation, a permit and a crew.

Bottom Line:

Nothing in orbit takes load off a US grid before 2030. Every serious forecast says so, including the bullish ones. The orbital players are honest about it when you read past the headline: their near-term customers are other satellites and the defense community.

So the megawatts that land in 2027 & 2028 still come from a substation, a gen yard and a fiber route somewhere in Texas, Virginia or Ohio. If your capacity plan has a line item for orbit, I’d love to see the radiator drawings….

Sources