Starcloud Put an NVIDIA H100 in Orbit — 17 Months Later, It’s Worth $1.1B

A startup that launched its first satellite five months ago just raised $170 million at a $1.1 billion valuation. Starcloud, a Y Combinator company building data centers in low Earth orbit, closed its Series A today led by Benchmark and EQT Ventures. That makes it the fastest company in YC history to hit unicorn status — 17 months from demo day to a billion-dollar valuation.

The round is also more than double the size of the next largest YC Series A ever. Total capital raised now sits at $200 million. Benchmark’s Chetan Puttagunta — a six-time Midas lister whose deal sheet includes GitHub, Confluent, and Elastic — is joining the board. When that guy decides to go big on putting servers in space, you pay attention.

But here’s what separates Starcloud from every other ambitious space startup: this isn’t a PowerPoint company. They already put an NVIDIA H100 GPU into orbit, trained an AI model in space, and ran Google’s Gemini up there. The question isn’t whether orbital compute works. It does. The question is whether it can scale fast enough to matter before terrestrial infrastructure catches up.

The Energy Crisis That Makes “Servers in Space” Sound Rational

The IEA projects global data center energy demand will blow past 1,000 terawatt-hours in 2026. That’s roughly equal to Japan’s entire electricity consumption. And it’s accelerating — AI data center power demand in the US alone could jump from 4 gigawatts today to 123 GW by 2035, according to Deloitte.

Meanwhile, 11 gigawatts of planned data center capacity worldwide is currently stalled because power grids simply cannot keep up. Utilities in the US and Europe, choked by permitting backlogs and infrastructure bottlenecks, are building too slowly for the AI boom. Every major lab is fighting over the same limited pool of electricity, land, and water.

Space solves all three constraints simultaneously. A solar array in orbit generates over five times the energy of the same panel on Earth — no clouds, no night cycles in sun-synchronous orbits, no weather. Radiative cooling in the vacuum of space eliminates the massive chiller systems and water supplies that terrestrial data centers burn through. And there’s no NIMBY opposition to a satellite 700 kilometers above your town.

Starcloud estimates orbital energy costs around $0.005 per kilowatt-hour. That’s up to 15 times cheaper than wholesale electricity prices on the ground. The company’s pitch isn’t about doing something cool in space. It’s about doing something cheaper in space. And when you frame it that way, $170 million from Benchmark starts making a lot more sense.

One Satellite, Three World Firsts

Starcloud-1 launched in November 2025 aboard a SpaceX rideshare mission. It carried a single NVIDIA H100 GPU — the first time that chip has ever operated in orbit, delivering roughly 100 times more powerful GPU compute than anything previously sent to space.

With that one satellite, the company stacked three firsts that nobody else has touched. They trained an LLM in space — specifically nanoGPT, the model architecture created by Andrej Karpathy. They ran inference on a version of Google DeepMind’s Gemma model in orbit. And they proved that a high-performance GPU can survive and function in the harsh radiation and thermal cycling of low Earth orbit with enough reliability to do real work.

The team behind this reads like someone assembled it specifically to build this exact company. CEO Philip Johnston came from McKinsey, where he worked on satellite projects for national space agencies. He holds degrees from Harvard, Wharton, and Columbia, plus a CFA. Chief Engineer Adi Oltean spent years at SpaceX as the responsible engineer for Starlink’s tracking beams — the system that enables Starlink connectivity for other spacecraft, including Starship. Before SpaceX, he was a principal engineer at Microsoft Azure for 20 years and holds over 25 patents. CTO Ezra Feilden came from Airbus Defence and Space with a decade of satellite design experience and a PhD from Imperial College London.

SpaceX for launch expertise. Microsoft for cloud infrastructure. Airbus for satellite hardware. All three domains converge in one company. The fact that they went from founding to first orbital hardware in under two years — and from founding to unicorn in 26 months — starts to make sense when you look at the pedigree.

The company was originally called Lumen Orbit when it was founded in January 2024 in El Segundo, California. A trademark dispute with Lumen Technologies forced a rebrand to Starcloud in March 2025. They relocated to Redmond, Washington — right in Microsoft and SpaceX territory.

Bitcoin, Blackwell, and the 88,000-Satellite Filing

The next milestone is Starcloud-2, scheduled to launch in October 2026. It’s a significant step up from the proof-of-concept first satellite. The payload includes multiple GPUs with an NVIDIA Blackwell chip — the latest generation — plus an AWS Outposts server blade. That AWS partnership, announced in February, means Starcloud will essentially extend Amazon’s cloud infrastructure into orbit. Not a simulation. Actual AWS hardware running above the atmosphere.

Then there’s the bitcoin angle. Starcloud-2 will also carry bitcoin mining ASICs, making it the first attempt at cryptocurrency mining in space. CEO Johnston’s reasoning is characteristically pragmatic: mining ASICs are roughly 30 times cheaper per kilowatt than GPUs. If your biggest competitive advantage is cheap solar power with zero electricity bills, mining is an obvious way to monetize excess capacity. Whether you think orbital bitcoin mining is brilliant or absurd probably says more about your priors than about the economics.

The real scale play, though, is Starcloud-3 — a 200-kilowatt, three-ton spacecraft designed specifically for SpaceX’s Starship. Each Starship flight could carry about 50 Starcloud-3 satellites using the same “pez dispenser” deployment system SpaceX built for Starlink. That works out to roughly 10 megawatts of computing capacity per single launch. The company expects to start flying on Starship in the mid-to-late 2028 timeframe.

And in February 2026, Starcloud filed an FCC application to deploy a constellation of up to 88,000 satellites in low Earth orbit, positioned between 600 and 850 kilometers altitude in sun-synchronous orbits. For context, SpaceX’s Starlink constellation currently has around 7,000 satellites in orbit, with a full plan topping out at roughly 42,000. Starcloud is proposing more than double that number.

Whether they get anywhere close to 88,000 depends on launch costs, manufacturing throughput, and regulatory approvals that don’t exist yet. But the filing signals what the long-term vision actually looks like — not a handful of demo satellites, but computing infrastructure at planetary scale.

The Space Compute Race Is Getting Crowded

Starcloud isn’t alone in chasing this idea, though it’s furthest ahead in terms of actual hardware in orbit. Sophia Space, a Pasadena startup founded by a retired NASA JPL fellow, raised $10 million in a seed round in February. Their approach is different — modular TILE compute units using NVIDIA Jetson Orin chips rather than full H100s — with a more conservative timeline aiming for a first orbital demo in late 2027 and full operations in the 2030s.

ThinkOrbital and TakeMe2Space are also developing concepts, though at much earlier stages. And then there’s the elephant floating in zero gravity: SpaceX itself. When a partnership between Starcloud and robotic assembly company Rendezvous Robotics was announced for self-assembling data centers in orbit, Elon Musk responded publicly that “SpaceX will be doing this.” Whether that means SpaceX sees Starcloud as a customer, a partner, or a future competitor is anyone’s guess — but when the company that controls launch costs signals interest in your market, the dynamics shift fast.

The skeptics have real ammunition too. Some lifecycle analyses estimate orbital compute carries an effective carbon intensity of 800 to 1,500 grams of CO2 equivalent per kilowatt-hour when you factor in launch emissions, hardware manufacturing, and eventual reentry. That’s worse than any country’s power grid. Industry figures that only count the operational energy dimension land much lower, around 134 to 165 grams, but the full-lifecycle numbers are uncomfortable for anyone selling space compute as a green solution.

There’s also the latency question. Low Earth orbit adds round-trip signal delays that matter for real-time applications. Starcloud’s pitch smartly focuses on AI training workloads where latency is irrelevant compared to raw throughput and cost — you’re not running a stock trading algorithm from space, you’re crunching gradient updates with cheap solar power. But that narrows the addressable market compared to a generic data center.

The real bet isn’t about whether space compute works. Starcloud-1 settled that. The bet is about timing. SpaceX’s Starship, with its projected 10x reduction in cost per kilogram to orbit, is the critical enabler. If Starship delivers on its economics, orbital compute could genuinely undercut terrestrial alternatives on price. If launch costs stay high, 88,000 satellites stays a number in an FCC filing and a slide deck.

Seventeen months from a YC demo stage to a billion-dollar valuation. One GPU orbiting Earth, 200 million dollars in the bank, and a filing for a satellite constellation that would dwarf Starlink. The space race for AI infrastructure is no longer hypothetical — and Starcloud is the only company in it that’s already running models above the atmosphere.

Top AI Product

Leave a comment Cancel reply