In November 2025, a 60-kilogram satellite carrying a single NVIDIA H100 GPU launched into orbit on a SpaceX rideshare mission.
Within days, it did something no machine had ever done before: it trained a large language model in space. Then it ran Google’s Gemma — a version of Gemini — directly from orbit and sent back a response that began: “Greetings, Earthlings!”
That satellite belonged to Starcloud, a startup founded in 2024 by a former McKinsey consultant with degrees from Harvard, Wharton, and Columbia. NVIDIA backed them. Crusoe — one of the fastest-growing AI infrastructure companies in the US — signed on to become the first public cloud provider to run workloads in space by early 2027.
I almost missed this story. Most investors did. Because what happened next moved so fast that if you blinked, you’d think someone was making it up.
On January 30, SpaceX filed with the FCC to launch up to one million orbital data center satellites.
On January 21, Blue Origin announced TeraWave — 5,408 satellites delivering 6 terabits per second.
On February 3, Starcloud filed for 88,000 satellites of its own.
On March 19, Blue Origin filed again — Project Sunrise — for another 51,600 dedicated data center satellites on top of TeraWave.
And somewhere in the middle of all this, Eric Schmidt — the man who built Google into a $2 trillion company — confirmed on X that he acquired rocket manufacturer Relativity Space specifically to launch data centers into orbit. When asked if that was his plan, he replied with one word: “Yes.”
In four months, orbital computing went from “interesting conference topic” to a three-way arms race between the richest people on earth, backed by the most powerful chip company in the world.
Something is happening. And almost nobody in public equities is talking about it.
Why the smartest people in tech are suddenly looking up
To understand why this is happening now — and not five years from now — you need to understand one number.
67 gigawatts.
That’s how much additional power US data centers will need by 2030, according to Eric Schmidt’s testimony before the House Committee on Energy and Commerce. For context, the entire US nuclear fleet generates about 95 GW total. Schmidt told Congress: “These things are industrial at a scale that I have never seen in my life.”
It’s not a future problem. It’s happening right now.
PJM Interconnection — the grid operator covering 13 US states including Virginia, the heart of data center country — has 270 GW of generation and load stuck in its interconnection queue. Average wait time to connect: 5 to 7 years. Some projects won’t get plugged in until after 2032.
In Virginia alone, Dominion Energy received interconnection requests for 60 GW of new data center load. The entire state currently consumes less than that. Their realistic delivery capacity by 2030: 7-10 GW. The remaining 50+ GW? In limbo.
Bloomberg found that electricity costs near data centers have spiked up to 267% compared to five years ago. Communities are blocking new installations. Water for cooling is getting scarce in Arizona and Texas. And every quarter, AI compute demand grows another 50% with no sign of slowing.
The math is simple: demand for compute is growing exponentially. The ability to supply it from the ground is growing linearly. At some point, the curves diverge so far that terrestrial infrastructure alone can’t close the gap.
Now look up.
In sun-synchronous orbit, solar power is near-continuous. There’s no grid to overload. No land to buy. No water to consume for cooling — heat dissipates through radiation into the vacuum of space. No NIMBY opposition. No permitting delays. And it scales without limit — you just launch more satellites.
Starcloud’s CEO Philip Johnston put it simply: “In space, you get almost unlimited, low-cost renewable energy.” He projects energy costs in orbit to be 10x cheaper than terrestrial equivalents, even including launch expenses. Google published a feasibility study in November 2025 concluding that if launch costs to LEO reach $200/kg, orbital data centers become cost-competitive with ground-based facilities.
Current Falcon 9 cost: roughly $2,700/kg. That’s a 13x reduction needed. Musk says 2-3 years with Starship. Deutsche Bank says “well into the 2030s.” Google’s study targets 2035 if Starship scales to 180 launches per year.
The exact timeline is debatable. The direction is not. Every major figure in tech and space is converging on the same conclusion.
The three-way race — mapped in detail
What struck me most when I started pulling this together is that this isn’t one company with a wild idea. It’s a simultaneous, independent convergence of the biggest players in space and tech — all filing within weeks of each other.
SpaceX — 1 million satellites
SpaceX filed its FCC application on January 30, 2026, for what it calls the “SpaceX Orbital Data Center System.” The scale is staggering: up to one million satellites operating at 500-2,000km altitude. For context, there are roughly 14,000 active satellites in orbit today from all operators combined.
The satellites would communicate via optical laser links with SpaceX’s existing Starlink constellation (9,500+ satellites, laser mesh at 200 Gbps, upgrading to 1 Tbps+), which then connects to ground stations. The filing stated: “Orbital data centers are the most efficient way to meet the accelerating demand for AI computing power.”
SpaceX’s advantage is unmatched: it already operates the largest satellite constellation in history, launched 134 Falcon 9 missions in 2024, and has Starship in development — the only vehicle that could realistically deploy data center satellites at the million-unit scale. The Terafab chip factory announced March 21 is designed to produce custom D3 radiation-hardened processors specifically for these orbital satellites — with 80% of Terafab’s production allocated to space applications.
xAI — Musk’s AI company, merged with SpaceX in February at a combined $1.25 trillion valuation — is the built-in anchor customer. The entire vertical stack is controlled by one person: the chips (Terafab), the rockets (Falcon 9/Starship), the network (Starlink), the AI models (Grok/xAI), and now the orbital compute platform. Nobody else on earth has this.
Blue Origin — TeraWave + Project Sunrise (57,000+ satellites)
Jeff Bezos is building a two-layer architecture.
Layer 1: TeraWave, announced January 21, 2026. A communications backbone of 5,408 satellites — 5,280 in LEO and 128 in MEO — delivering up to 6 Tbps globally via optical inter-satellite links. This is the network layer. Deployment starts Q4 2027 using Blue Origin’s New Glenn heavy-lift vehicle.
Layer 2: Project Sunrise, filed with the FCC on March 19, 2026. Up to 51,600 dedicated orbital data center satellites — the actual compute layer. These satellites route data through TeraWave’s optical mesh backbone. Together, the two constellations form a complete compute + connectivity stack.
Blue Origin explicitly positioned TeraWave as “enterprise-grade” — targeting up to 100,000 enterprise, data center, and government customers. This is not consumer broadband. This is infrastructure for hyperscale AI.
Blue Origin is the only other player besides SpaceX that builds both the rockets and the satellites. The vertical integration parallel is exact. And with the appointment of former ULA CEO Tory Bruno to lead Blue Origin’s National Security Unit, they’re also going after the defense market — a direct challenge to SpaceX’s Starshield.
Starcloud — the startup that already proved it works
Starcloud is the most advanced in terms of actual demonstration, which is remarkable for a company founded in 2024 by a team of three.
What they’ve already done:
Launched Starcloud-1 in November 2025 with an NVIDIA H100 — 100x more powerful GPU compute than had ever been in space before
Trained the first LLM in orbit
Ran Google’s Gemma model from space, producing inference responses beamed back to Earth
Partnered with Crusoe to become the first public cloud provider to operate workloads in space by early 2027
Filed for 88,000 satellites with the FCC in February 2026
What’s coming next: Starcloud-2 launches October 2026 with NVIDIA Blackwell platform integration (10x the performance of H100), 100x the power generation of the first satellite, and — according to the company — it will generate more cash than it costs to build and launch. That’s the unit economics inflection point.
Long-term vision: a 5-gigawatt orbital data center with solar and cooling panels approximately 4 kilometers across. Johnston predicts that within 10 years, most new data centers will be built in space.
NVIDIA is deeply embedded. At GTC 2026, Jensen Huang announced the NVIDIA Space-1 Vera Rubin Module — delivering 25x more AI compute for space-based inference versus the H100. Named partners: Starcloud, Aetherflux, Axiom Space, Kepler Communications, Planet Labs, and Sophia Space.
The supporting cast
Eric Schmidt / Relativity Space: Schmidt acquired the rocket company in March 2025 for ~$800M, becoming CEO with a controlling stake. The Terran R rocket (33.5 tons to LEO, first flight expected 2026) gives him independent launch capability that doesn’t depend on Musk or Bezos. He confirmed on X that orbital data centers are the explicit reason for the acquisition. His congressional testimony: data centers could go from consuming 3% of US electricity to 99% without space-based alternatives.
China — CASC “space+” initiative: China Aerospace Science and Technology Corporation announced plans for “giga-watt level space infrastructure” for data processing. China has proposed a 200,000-satellite constellation focused on data sovereignty and in-orbit processing.
Axiom Space: Developing an orbital data center capability (ODC T1) for its commercial space station, using NVIDIA accelerated computing platforms. The first module is planned to dock with the ISS as early as 2026.
Aetherflux: Pioneering space-based solar power beaming and orbital compute. Using NVIDIA platforms. Their CEO stated at GTC: “We’re pioneering a new paradigm for power and compute in space.”
The market size
The orbital data center market is projected to grow from approximately $1.77 billion in 2029 to $39.1 billion by 2035, at a compound annual growth rate of 67.4%.
For comparison, the global cloud infrastructure market was worth about $2.5 billion in 2008. By 2025, it exceeded $300 billion. That’s 120x in 17 years.
Nobody knows if orbital compute will follow that trajectory. But the parallels are uncomfortable to ignore: a new infrastructure paradigm, skepticism from incumbents, a forcing function (energy constraints vs bandwidth constraints), and a handful of companies racing to build it before everyone else realizes it’s real.
The question every public equity investor should be asking right now is: where is the supply chain?
The supply chain that nobody’s mapped
Most of the orbital data center operators are private. SpaceX ($1.25T, IPO this summer). Blue Origin (private, Bezos-owned). Starcloud (YC-backed startup). Relativity Space (Schmidt-owned).
But the companies that supply them — the rockets, the components, the laser links, the chips, the satellite infrastructure — include several public equities that the market hasn’t connected to this thesis yet.
I’ve identified six positions across the risk spectrum. Some are picks-and-shovels plays that win regardless of which operator dominates. Others are direct bets on the orbital compute market materializing at scale.
The full breakdown — with the investment thesis for each, the specific link to orbital compute, why each one has a floor even if the space thesis is delayed by 5 years, my risk framework, entry points, and the one date that determines whether this thesis accelerates or stalls — is in the premium section below.
What’s inside:
— 6 publicly traded companies positioned on the orbital compute supply chain, ranked from most defensive to most speculative
— For each: the thesis, the direct orbital compute link, the terrestrial floor (why it works even if space is delayed), and current valuation context
— My risk framework: the 5 things that could kill this thesis — technical, economic, regulatory, execution, and competition — and how each one affects the positions differently
— Sizing and entry logic: how I’m building these positions, at what prices, and the one catalyst date I’m watching that will determine whether I scale up or step back