Data Centers in Space – Innovation or Not?

Data Centers in Space - Innovation or Not?

by Braden Kelley and Art Inteligencia


Somebody has to say it out loud: the plan to put AI data centers in orbit sounds like something I’d have pitched as a joke in a brainstorm, right before somebody reminded me that radiators are heavy.

And yet Google has TPUs in orbit. SpaceX has asked the FCC for permission to launch up to a million computing satellites. Blue Origin wants 51,600. A startup called Cowboy Space wants to turn the upper stage of its own rocket into a megawatt data center. Starcloud has already flown an Nvidia H100.

Two questions are worth asking, and they’re easy to blur together. Does it make economic sense? And is this really happening because Americans are turning against data centers on the ground? I’ll take the second one seriously, because the politics are real. But I’ll start with the money.

The short answer

Not today, and probably not this decade on cost alone. Orbital compute carries a large price premium over ground-based compute, and the best independent analysis I found expects that premium to shrink but not vanish. The case for space is less “it’s cheaper” than “it’s available.” That’s a different argument, and a more interesting one.

And on the politics: every major orbital data center filing I found was made before the November 3 midterms. The idea didn’t come from the election. The election just made the idea easier to sell.

Who’s building what

Here are the five efforts I’d watch, side by side. Where I’ve given a figure, it comes from the company’s filing, announcement or press coverage, not from a flight record.

Approach Scale they’ve asked for What’s actually flown Next milestone
Google (Project Suncatcher) Small satellites with TPUs, flying in tight formation and linked by lasers, in near-constant sunlight An illustrative cluster of 81 satellites Four TPUs on a Planet-built prototype, launched October 1, 2026 In-orbit tests now underway
SpaceX Massive constellation of AI satellites riding on Starlink’s laser network and Starship launches Up to 1,000,000 satellites (FCC filing) Nothing yet that I could verify Deutsche Bank expects small tests in 2027 or 2028
Cowboy Space Rocket upper stage is the data center: a 1 MW node with five server racks Up to 20,000 satellites (FCC filing, pending) Reason-1, a power-beaming demo launched October 1, 2026 First Big Iron rocket launch targeted for 2028
Blue Origin (Project Sunrise) Satellite constellation tied to its TeraWave network Up to 51,600 satellites (FCC filing) Nothing yet that I could verify No schedule given; seeking waivers of FCC deadlines
Starcloud Modular satellites built around Nvidia chips, launched mostly on SpaceX 88,000 satellites and 20 GW (FCC application) Starcloud-1 with an Nvidia H100, launched November 2025 Two 8 kW satellites, first one January 2027

Google: the careful engineers

Google’s approach reads like a research paper because it is one. The satellites would fly in a dawn–dusk sun-synchronous orbit, where the panels see near-constant sunlight. Google says panels in the right orbit can be up to eight times more productive than on Earth. Its team tested a Trillium TPU in a proton beam and found the memory showed irregularities only after about three times the dose expected over five years of shielded orbit. Google called the chip “surprisingly radiation-hard for space applications.”

Google also did the most useful thing any of these players did: it published the economics. Its analysis suggests launch prices could fall below $200 per kilogram by the mid-2030s, which is where it says space-based data centers become roughly cost-comparable to terrestrial energy costs. Notice how specific that is. It’s a condition, not a promise. And Google itself says “significant engineering challenges remain,” including thermal management and reliability.

The prototype launched on October 1, 2026 on a SpaceX rideshare, and Google reported it was operating as expected. A trade-press roundup from early November says the prototype ran Gemma 4 on an orbital TPU, but only for about 20 minutes of one TPU at a time because of power limits. That’s a proof of concept, not a data center.

SpaceX: the biggest ambition and the loosest details

SpaceX’s FCC filing asks for approval of up to one million satellites at altitudes of 500 to 2,000 kilometers. The filing’s implied math is roughly a million tonnes of hardware per year, adding about 100 gigawatts of AI capacity annually. It gives no satellite mass, no hardware design and no unit cost. One analyst called it “quite rushed and to be very early stage.”

Then there’s the gap between the pitch and the paperwork. Elon Musk has said space would be the cheapest place to run AI within 30 to 36 months, adding “It’s not even close.” But SpaceX’s own IPO filing is far more modest. It says the orbital initiatives depend on “unproven technologies, or technologies that do not exist,” that they “may not achieve commercial viability,” and that the timeline for 100 GW “may be difficult or impossible to determine.”

The launch arithmetic explains the caution. 100 GW is on the order of 10,000 Starship launches. the economics lean on fully reusable Starship boosters, which an early-2026 analysis said SpaceX had not yet demonstrated. And the filing says SpaceX does “not anticipate servicing or repairing processors in space,” so every failed chip is a lost chip.

Credit where due: SpaceX has a record of doing the improbable. The Colossus cluster, from xAI, scaled to 100,000 GPUs in 122 days against a 24-month estimate. That’s a reason not to dismiss it. It isn’t evidence that orbit works.

Cowboy Space: the most original architecture

Cowboy is the one I find most interesting, because it attacks a different bottleneck. Most orbital plans launch a satellite on someone else’s rocket and discard the upper stage. Cowboy’s idea is to keep the upper stage, make it the data center, and skip the redundant structure so more of the launched mass goes to compute, power and cooling.

The company was formerly Aetherflux, a space-based solar power startup, and was founded in 2024 by Baiju Bhatt, a Robinhood co-founder. It raised a $275 million Series B led by Index Ventures at a reported $2 billion valuation. Its Stampede constellation is filed at up to 20,000 satellites, each a 1 MW node with five server racks, using Nvidia’s Space-1 Vera Rubin modules. The rocket, Big Iron, is an 86-meter stainless-steel vehicle with nine reusable methalox engines. The first full vehicle is targeted for the end of 2027 and a maiden launch for 2028.

The pitch is blunt: “Earth’s energy grid can’t run at the pace of AI. We can.” The company also says average grid connection lead times for new data centers in major U.S. markets “can run five to seven years, or more.” I haven’t verified that figure independently, so treat it as the company’s claim.

The trade-off is concentration risk. One analysis of the sector put it plainly: smaller nodes lose little when one fails, while Cowboy’s larger nodes lose far more, and the company has to build a new rocket and win FAA authorization first. Its own site is candid that its capabilities and timelines are “design objectives and goals, not demonstrated or guaranteed outcomes.” And the FCC sent a deficiency letter in August asking about satellite dimensions and debris mitigation.

The rest of the field

Blue Origin has asked the FCC for up to 51,600 satellites under Project Sunrise, and Jeff Bezos has said that “We will be able to beat the cost of terrestrial data centers in space in the next couple of decades.” Note the timeframe: couple of decades. That’s a more candid number than the ones coming out of SpaceX. Starcloud is the most flown, with an H100 in orbit since November 2025, and has raised about $450 million at a $2.3 billion valuation. Its CEO, Philip Johnston, concedes that if he can’t book SpaceX launches in 2029, “that will be challenging for us.” Which tells you who the whole industry is renting from. Orbital Compute, a Los Angeles startup with a $5 million pre-seed, is proposing up to 100,000 modular 100 kW satellites.

Is it cost effective?

The best independent number I found comes from Boston Consulting Group. Its conclusion: orbital facilities carry a 2.5x to 3x cost premium today. Over 20 years, BCG puts the cost at roughly $660 million to $750 million per megawatt in orbit versus $230 million to $300 million on the ground. Its closing line is one I’d put on a slide: “Space-based data centers won’t win on cost—but they don’t have to.”

BCG’s improvement scenarios are worth reading closely:

  • A realistic 5-to-10-year path (launch costs down about two-thirds, satellites half as heavy) leaves a premium of roughly 1.5x to 1.8x, depending on how many satellites fail.
  • An aggressive path (full Starship-level economics) could get the premium down to 1.1x to 1.2x, but only if failure rates stay very low.
  • Parity this decade is judged unlikely.

At roughly $1,500 per kilogram, BCG says putting 1 GW of compute in orbit would cost around $30 billion in launch alone, which it calls economically impractical. A separate open calculator built by Andrew McCalip, which Tom’s Hardware covered, found a baseline of about $51 billion for a space installation against $16 billion on the ground. Its author’s verdict on the whole idea was “FOMO and aesthetic futurism.” That model covers only launch logistics, and he says so.

Three things keep the economics stubborn:

  1. Heat. In a vacuum you can only shed heat by radiation. SpaceX’s own filing says its compute satellites would need “substantially larger radiators” than Starlink’s. The International Space Station sheds tens of kilowatts; an AI cluster has to shed megawatts.
  2. Failures. Chips fail, especially under full load, and nobody is going up to swap a card.
  3. Obsolescence. GPUs age out in a few years. A satellite is a poor place to hold an asset that depreciates that fast.

My own arithmetic on scale makes the point vivid. A single 1 GW campus is a thousand Cowboy-sized 1 MW nodes. Starcloud’s first two flight satellites total 16 kW. That’s the distance between a demo and a business, and it’s why I’d treat any “cheaper than Earth within three years” claim with the same skepticism SpaceX’s own lawyers apparently do.

Is this about the backlash against data centers?

Now to the question you didn’t ask but I know you’re thinking about. Is space only attractive because building data centers on the ground has become politically harder?

The backlash is real, and it’s big. On November 3, voters in 10 states decided 41 local data center measures, and nearly every one limited development. Ohio alone had 21. In Ashville, Ohio, more than 80% of voters rejected a data center already under construction. In Texas, Gina Hinojosa won the governor’s race after calling for a moratorium on new data center approvals. Analysts cited by one trade outlet say up to 60% of planned 2028 U.S. data center projects could face delays, and Goldman Sachs puts 45% of planned capacity at risk. Early this year, at least 75 projects worth about $130 billion were reportedly stalled or blocked.

So yes, Democratic candidates ran on this issue and many won. But if the story is “one party’s policies caused the backlash,” the evidence doesn’t back it up:

  • The opposition crosses party lines. Gallup found about 70% of Americans oppose local AI data center construction, including 75% of Democrats and 63% of Republicans. In Box Elder County, Utah, where Trump took nearly 80% of the vote, a 40,000-acre project drew intense backlash from rural conservative voters over water and tax breaks.
  • Republican governors acted first. Greg Abbott ordered a statewide audit of data centers seeking grid connections, saying “Texans must come first.” Ron DeSantis signed new requirements on hyperscale data centers. Nevada’s Joe Lombardo cut sales-tax benefits and told companies to cover their own power and water costs. Nebraska’s Jim Pillen said “we cannot foot the bill for them.”
  • The concerns are about bills, water and tax breaks. One Federal Reserve Bank of Dallas paper estimated data centers raised wholesale electricity prices 6.9%. Corpus Christi’s reservoir fell to 8% of capacity. Those aren’t partisan facts.

There’s a fair argument on the other side, and I’d make it too. Blanket bans and moratoria are blunt instruments. They can push the same demand to another state or country without fixing the bill problem, and they can slow the very grid and generation upgrades communities need. People who think that way aren’t wrong to worry. The administration has taken that line, with President Trump calling Texas’s opposition a “mistake.”

But here’s my main point. Space doesn’t escape the politics, it changes the venue. Cowboy’s constellation filing is pending at the FCC with a deficiency letter outstanding. Blue Origin is asking for waivers on deployment deadlines. SpaceX wants the standard six-year milestones waived. And the Center for Space Environmentalism has already raised concerns about collision risk, re-entry hazards and interference with astronomy. A million satellites will have their own opposition, and I’d bet it shows up faster than anyone is planning for.

What would have to be true

For orbital data centers to be a real answer, five things need to line up:

  1. Launch costs fall by an order of magnitude, roughly from about $1,500 per kilogram toward the $100 to $200 range.
  2. Heat rejection works at megawatt scale, in hardware that actually flew.
  3. Failure rates stay low enough that losing a satellite isn’t losing a company.
  4. Regulators approve constellations of a size no one has licensed before.
  5. Ground alternatives stay stuck. If efficiency, flexible demand and local small models ease the pressure, the premium stops being worth paying.

That last point is the one I’d emphasize. OpenAI’s Grid Protectors program, which offers up to 1 GW of flexible demand at a planned Georgia campus that utilities could cut in emergencies, is a competing answer to the same constraint. So is running smaller models locally, which I wrote about in the hybrid intelligence piece. Space competes with every one of those, and they’re all available now.

My verdict

Innovation or Not? Innovation, with an asterisk the size of a Starship.

The engineering effort is genuine, the capital is real, and Google’s disclosure of its economics is the kind of honesty this field needs more of. But the evidence says orbital compute is a premium product for buyers who can’t get power on the ground, not a cheaper way to run AI. It won’t relieve the pressure on communities any time soon, because the first real megawatt nodes are a 2028-or-later story and the numbers needed to match a single large campus are enormous.

And the politics? The backlash gave the idea a better story, but it didn’t create it. The filings came first. The more useful lesson for innovators is the one the voters taught: people don’t mind new infrastructure, they mind paying for it, drinking less water because of it, and not being asked. A technology that solves a real constraint and respects the people near it wins. Whether the data center sits in Ohio or in a dawn–dusk orbit, that rule travels with it.

I’d watch two signals: whether the Google prototype’s in-orbit tests show the chips and links holding up, and whether SpaceX’s real launch cadence starts to look like the plan. Those will tell us more than any filing.

Sources

Image Credits: Gemini

Content Authenticity Statement: The topic area, key elements to focus on, etc. were decisions made by Braden Kelley, with a little help from Claude to clean up the article.

Subscribe to Human-Centered Change & Innovation WeeklySign up here to get Human-Centered Change & Innovation Weekly delivered to your inbox every week.

This entry was posted in Technology, Innovation on by .

About Art Inteligencia

Art Inteligencia is the lead futurist at Inteligencia Ltd. He is passionate about content creation and thinks about it as more science than art. Art travels the world at the speed of light, over mountains and under oceans. His favorite numbers are one and zero. Content Authenticity Statement: If it wasn't clear, any articles under Art's byline have been written by OpenAI Playground or Gemini using Braden Kelley and public content as inspiration.

Leave a Reply

Your email address will not be published. Required fields are marked *