Research brief · Frontier infrastructure · October 2026

Data centers
in space

In one week of November 2025, Starcloud flew the first Nvidia H100 into orbit and Google published Project Suncatcher. We read the most detailed research on whether orbital AI compute is actually sustainable, then followed the capital: the billionaires and investors funding the climb.

By Christos Papadimitriou · House of 47 Research
Last updated:

See the build
$200/kgLaunch cost needed for parity, Google
8×Solar power vs on Earth, near 24/7
2.5 to 4×Cost premium in orbit today
2027First prototypes in orbit
A satellite orbiting Earth above the clouds
The build · 2025 → 2050

From slide deck to orbit

The pitch is simple: in the right orbit the sun never sets, so a solar panel is up to eight times more productive and runs almost continuously. No grid, no water, no land. The catch is everything else: launch cost, cooling in a vacuum, and no way to fix a broken server 650 km up.

2025 · Now

First hardware up

H100

Starcloud flew the first Nvidia H100 into orbit on Nov 2, and trained a small model there. Days later Google published Project Suncatcher. The race left the slide deck.

2027 · The test

Prototypes

2 sats

Google launches two TPU satellites with Planet Labs. Starcloud targets first commercial orbital capacity with Crusoe. The real test of cooling and reliability, not just power.

2030s · The hinge

Crossover window

$200/kg

If Starship drives launch toward roughly $200/kg, free solar starts to pay for batch training. If it stalls at today's $1,500/kg, orbit stays a niche for defense and Earth observation.

2050 · The vision

Gigawatt scale

1 GW

Europe's ASCEND study targets 1 GW in orbit before 2050. Bezos predicts gigawatt data centers within 10 to 20 years. The bull case, if five hard things all go right.

The two numbers that decide it

Launch cost to orbit, $/kg (thousands)

Cross-check · Google, ASCEND, IEEE, SemiAnalysis, Aravolta

What everyone agrees on

  • All
    Solar is the real prize

    In the right orbit a panel is up to 8 times more productive and runs nearly non-stop. No night, no weather. This is the one unambiguous advantage.

  • All
    Only batch training fits

    Round-trip latency of 20 to 600 ms and a thin downlink rule out real-time inference and general cloud. Training and in-space data, not your website.

  • All
    Launch cost is the hinge

    Everything pivots on reaching roughly $200/kg with Starship-class reuse proven at cadence. Stuck at $1,500 to $3,000, the economics stay 2.5 to 4 times underwater.

  • All
    Cooling is the hard part, not chips

    A vacuum is a thermos, not a fridge. Heat leaves only by radiation, so rejecting a gigawatt needs radiators measured in square kilometres.

  • Most
    Chips survive better than feared

    Google's TPU showed no hard failures to 15 krad in a proton beam, about 3 times a shielded five-year dose. Radiation is not the blocker people assumed.

  • Most
    Green only with clean launchers

    Europe's ASCEND finds orbit cuts digital CO2 only if a launcher roughly 10 times less emissive exists. Otherwise launch emissions erase the grid savings.

Where they split

Is it viable this decade?

From "ridiculous" to "cost parity within years." Placed by stance, not a single metric.

Snake oilParity this decade

Cost premium in orbit, today

How much dearer than a terrestrial data center, per unit of compute, right now.

1× parity4× dearer

When gigawatt scale?

A gigawatt-class data center operating in orbit.

NeverWithin ~15 years

Google Project Suncatcher 2025 ($200/kg crossover, 8x solar, TPU radiation test) · Thales ASCEND 2024 (1 GW by 2050 if launchers 10x cleaner) · IEEE Spectrum 2026 (cooling physics) · Aravolta 2026 (radiators 3-10 km2 per GW, 3-5 yr life) · SemiAnalysis / Varda / BCG (2.5 to 4x cost premium) · Bezos, Musk, Altman (public statements 2025-26).

Probability of outcome · House of 47 synthesis

The odds

Synthesis estimates across the sources above, not market forecasts. Google's own framing: the concept is "not precluded by fundamental physics or insurmountable economic barriers," but viability rides on an unproven launch-cost collapse plus unsolved thermal and reliability engineering.

Viability scorecard · the five things that must hold

What has to be true

Orbital compute is sustainable only if every line below turns green at once. Today, two are advantages, one is manageable, and three are hard.

ConstraintVerdictThe number

Figures from Google Suncatcher, Aravolta, IEEE Spectrum and launch-cost data (CSIS, SemiAnalysis). The binding constraints are launch cost, cooling and in-orbit reliability.

Follow the capital · who is actually building it

Who builds it

A frontier bet this early is really a map of who can afford to be wrong for a decade. Mapping exactly who stands behind bets like this, their stakes, and how to reach them, is the work of 47HNWI.

Jeff BezosBlue Origin · net worth ~$225B

Predicts gigawatt data centers in space within 10 to 20 years, on 24/7 solar at cost parity. New Glenn is a launch enabler. Believer

Elon MuskSpaceX · Starlink · xAI · ~$600B+

The launch and laser-mesh backbone, and the most aggressive timeline. Starship plus Starlink V3 optical links are the proposed substrate, framed as scaling toward 100 GW/yr of compute. Key enabler

GooglePage, Brin, Pichai · Alphabet

Project Suncatcher: TPUs in orbit, two prototype satellites with Planet in 2027. Corporate R&D, the most credible big-tech entrant. Believer

Jensen HuangNvidia · ~$150B

Nvidia invested in Starcloud and put the first H100 in orbit. The chips are the compute layer for the leading private play. Enabler

Eric SchmidtRelativity Space · ~$25B

Took a controlling stake in launch startup Relativity and became CEO, explicitly to launch data centers into orbit as AI power demand explodes. Believer

Starcloudthe pure-play · $2.3B valuation

Ex-Lumen Orbit, founded 2024. About $450M raised, backed by Benchmark, EQT Ventures, Nvidia and Cisco. Fastest YC startup ever to unicorn. Operator

Sam AltmanOpenAI

The loudest sceptic. Calls near-term orbital data centers "ridiculous" and is betting the other way, on the terrestrial Stargate buildout. Sceptic

Investable proxies

Nvidia (NVDA, chips plus a Starcloud stake), Rocket Lab (RKLB, launch), Alphabet (GOOGL, Suncatcher optionality), and since its 2026 IPO, SpaceX itself. The pure-plays, Starcloud and Relativity, stay private.

The capital pattern

A classic frontier syndicate around one pure-play: YC and scout funds at seed, Benchmark and EQT at growth, strategic corporates anchoring the stack. No disclosed sovereign-wealth or family-office checks yet. The sovereign vector is institutional, through Europe's ASCEND.

Free sunlight overhead, expensive everything else. The physics may yet allow it. The open question is who keeps funding the climb.

Sources · figures verified against primary reports

Sources

Changelog · what changed in this brief, newest first

Changelog