Happy Saturday.
Today's Deep Dive is the most interesting venture bet of 2026 — and most of the takes I've read are missing the punchline.
The company is Starcloud. Y Combinator alum, originally Lumen Orbit, founded January 2024 in El Segundo by Philip Johnston (ex-McKinsey), Adi Oltean (ex-SpaceX, ex-Microsoft Azure), and Ezra Feilden (ex-Airbus). This week they closed a 170 million dollar Series A at a 1.1 billion dollar valuation, led by Benchmark and EQT Ventures. Total raised: 200 million. Time from YC graduation to unicorn: 17 months. Fastest in YC history.
Their thesis is simple enough to fit on a napkin: put AI compute in orbit.
And here is what almost every article I've read missed.
This is not a moonshot deck. The proof already flew.
On November 2, 2025, Starcloud-1 launched on a SpaceX Falcon 9. A satellite the size of a refrigerator, roughly 60 kilograms, with an Nvidia H100 inside — the first data-center-class GPU ever placed in orbit. In December 2025, they trained their first AI model in space. This is not theater. It already happened.
Why the math suddenly works
Three things had to become true at the same time, and they finally did.
First, solar power in low earth orbit is 24/7. No clouds. No grid. No interconnect queue. The sun never sets on a satellite in the right geometry.
Second, space is the largest heat sink in the universe. Radiative cooling is essentially free. No chillers. No water. No fan walls. No utility bill.
Third, SpaceX has driven launch cost to roughly two thousand dollars per kilogram — down about 80% in a decade. A full rack of H100 / Blackwell-class compute weighs around 200 kilograms.
Power and cooling are about 70% of data center operating expense on Earth. Both drop to near zero in orbit. Three years ago this did not pencil. Today it barely does.
"Barely" is enough when the alternative is a US power grid that can't keep up with AI demand.
The part that stunned me
Starcloud-2 launches later this year. Nvidia Blackwell chips. 100 times the compute power of Starcloud-1. The largest commercial deployable radiator ever sent to space. And confirmed customers.
Crusoe is committed. Partnerships are in place with Amazon Web Services, Google Cloud, and Nvidia. Starcloud-2 will also be the first satellite to mine Bitcoin from orbit — because why not.
Read that customer list one more time.
Crusoe. AWS. Google Cloud. Nvidia.
The hyperscalers are not chasing Starcloud. They're already plugging in.
The skeptic's case is real
I want to be fair to the short side of this, because it is not small.
▪ Orbital debris is accelerating. Every new satellite raises the collision risk for every other satellite. LEO is getting crowded fast.
▪ Cosmic radiation degrades silicon. The mean-time-to-failure problem in space is brutal and not fully solved at this form factor.
▪ There is no service call. If a rack dies in orbit, it is dead until the next launch. On-orbit repair is years away.
▪ There is no regulatory framework for orbital data sovereignty. Who owns a bit that crosses the Kármán line?
Every watt that comes from space is a bet on reliability engineering that has never been tested at commercial compute scale.
My take
This is the most interesting venture bet of 2026 — not because it is likely to work, but because if it does, the data center wars move off-planet permanently.
Two paths. One, Starcloud becomes a trillion-dollar infrastructure company in five years and every hyperscaler either builds their own orbital fleet or pays Starcloud to rent one. Two, it becomes the most expensive flameout in venture history.
There is almost no middle outcome.
The capital allocators I respect are paying close attention. Benchmark does not lead nine-figure rounds in "interesting experiments." When Peter Fenton's firm writes a check like this, it's because their internal model says the downside is survivable and the upside is generational.
Starcloud-2 launches this year. That's the data point that decides it.
Watch that launch.
Until tomorrow — because money never sleeps.
Steve Kaplan
stevekaplan.ai
