The race to build data centers in orbit stopped being a thought experiment and became a capital allocation. On August 21, Starcloud closed a $250 million extension to its Series A at a $2.3 billion post-money valuation. Nvidia and Cisco Investments joined the round. The Redmond, Washington company has now raised $450 million since it was founded in 2024.
That is a large cheque for a company whose first satellite carried a single graphics processor. The number is not the point. The point is what the money buys: launch capacity, radiation-tolerant compute, and a position in a market that has no enterprise customer at scale yet.
Starcloud is the best-funded pure play in the category, and its economics rest on a rocket that has not yet flown a commercial payload at cadence.
The risks that are not yet priced are launch availability and insurance, not the engineering.
Starcloud Series A extension, August 2026
Led by Manhattan West, with Nvidia and Cisco Investments joining existing backers Benchmark and EQT. · Starcloud, 2026
Why compute is moving off the planet
Three constraints pushed this idea from science fiction into a funding round. Terrestrial data centers are queuing for grid connections that take years to deliver. Electricity prices for large loads are climbing in the same markets that host the densest clusters. Water-based cooling is running into local limits.
Orbit offers the opposite profile. A satellite in low Earth orbit (LEO) sees the sun almost continuously, so it needs no grid interconnection queue. Waste heat radiates into deep space instead of into a cooling tower. The trade is that every unit of compute has to be launched, and nothing can be repaired on site.
The ambition is now written into regulatory filings. SpaceX asked the Federal Communications Commission (FCC) in January for up to one million orbital data center satellites, projecting 100 gigawatts of AI compute — roughly a fifth of current U.S. generating capacity. Starcloud's own application, accepted in March, requests 88,000 satellites and 20 gigawatts. These are applications, not approvals, and the gap between a filing and a deployed watt is the entire investment question.
As we wrote in September, Loft Orbital and Marlan Space committed $1 billion to run AI workloads in orbit. It is attacking the same thesis from the compute side rather than the platform side.
From one GPU to a three-ton spacecraft
Starcloud-1 launched in November 2025 carrying an Nvidia H100, the first data-center-grade GPU to reach orbit. The satellite trained a small language model in space, then ran Google's Gemma. Starcloud-2, due this year, is built around Nvidia's Blackwell B200. Starcloud-3 is a three-ton spacecraft designed to fly on SpaceX's Starship. The company is fitting out a 100,000-square-foot factory in Washington to build the line.
“Last November we put the first NVIDIA H100 in orbit. Today this fresh capital empowers us to build the infrastructure to launch many more of NVIDIA's most advanced GPUs into space.”— Philip Johnston, co-founder and CEO, Starcloud
Nvidia is not a passive backer. The two companies are co-developing the Space-1 Vera Rubin Module, a compute package engineered for orbital conditions, targeted for 2028. Radiation exposure and heat rejection, handled by large radiators, set the design limits long before raw throughput does.
The capital stack behind the bet
Its March round raised $170 million at a $1.1 billion valuation, led by Benchmark and EQT. That made it the fastest company in Y Combinator's history to reach unicorn status, 17 months after finishing the program. The August extension, led by Manhattan West, doubled the mark to $2.3 billion. Total capital raised now stands at $450 million.
The investor list reads like a supply chain. Nvidia supplies the accelerators. Cisco Investments brings networking and optical expertise. Benchmark and EQT provide the venture track record. Each backer is buying a piece of the same wager: that compute in orbit becomes infrastructure rather than a demonstration.
| Parameter | Orbital data center | Terrestrial data center |
|---|---|---|
| Power source | ✔ Continuous solar, no grid queue | ✗ Interconnection queue, rising prices |
| Cooling | ✔ Radiative rejection to space | ◐ Water and air cooling, constrained |
| Latency to users | ✗ Orbital plus ground-segment hops | ✔ Single-digit milliseconds in metro |
| Scaling constraint | ✗ Every unit needs a launch slot | ◐ Land, power and permits |
| Repairability | ✗ No on-site servicing at scale | ✔ Standard field replacement |
Comparison of orbital and terrestrial compute economics · Nexithon analysis, 2026
The bottleneck nobody has priced
Launch is the constraint that decides whether any of this scales. SpaceX is retiring Falcon 9 in favour of Starship, a vehicle that has not yet demonstrated commercial cadence. The company is weighing a dedicated Falcon 9 purchase and contracts with other providers to bridge the gap. A constellation of 88,000 satellites cannot be deployed on a manifest that does not exist yet.
Insurance is the second gap. In August, Marsh's U.S. aviation and space practice described orbital data centers as a new frontier for underwriting, while conceding that pricing models for unrepairable compute hardware do not exist. A satellite that cannot be serviced in orbit is an asset with a hard, unhedged failure curve.
Launch capacity: the deployment plan assumes Starship cadence that has not been proven.
Underwriting: insurers have no actuarial base for compute that cannot be repaired.
Demand: no enterprise has yet signed an anchor inference contract at scale.
What has to be true
Probability: 45% — the hardware is moving faster than the demand side, and the first anchor contract is what converts a constellation plan into a business.
✅ Arguments for
Terrestrial power constraints are worsening, which improves the relative case for orbit every quarter.
Confirmation criteria: a named hyperscaler or government agency discloses an orbital inference contract.
❌ Arguments against
Enterprise buyers hesitate to move production inference to hardware they cannot physically reach.
Disconfirmation criteria: Starship slips past 2027, or insurers decline to cover orbital compute.
Starship's first commercial payload at cadence
A named enterprise or agency inference contract
Starcloud-2 launch and B200 in-orbit benchmark results
An insurer publishing an orbital compute risk model
Development scenarios
🟢 Optimistic scenario (30%)
Implications: orbital compute becomes a fundable asset class, and the $2.3 billion mark looks conservative.
🟡 Base-case scenario (50%)
Implications: capital remains available, but revenue arrives later than the 2028 models assume.
🔴 Pessimistic scenario (20%)
Implications: the pure plays lose their premium, and orbital compute folds back into launch providers that own their own rockets.
Starcloud has bought itself the right to find out. The $2.3 billion valuation is a claim about deployment, not discovery, and the next two years will settle it.