The most expensive line item in any AI data center is not the silicon. It's the power to run it and the water to cool it. Starcloud, a startup that launched its first satellite with an NVIDIA H100 GPU in November 2025, is building its entire business case on moving those costs off-planet [TechCrunch, March 2026]. The company's pitch to its first customers, NASA and the U.S. Department of Defense, is a 10x reduction in energy costs for inference workloads, with zero freshwater consumption and a fraction of the carbon footprint of terrestrial operations [The Innovator]. It is a bet of staggering physical scale, backed by a $170 million Series A from Benchmark and EQT Ventures that valued the company at $1.1 billion just 17 months after its Y Combinator demo day [TechCrunch, March 2026].
The Wedge: Government Inference in Orbit
Starcloud is not chasing the hyperscalers for model training contracts, at least not yet. Its initial ideal customer profile is a U.S. government agency with earth observation satellites that need to run AI models on the data they collect. The value proposition is latency, security, and cost. By placing the compute next to the sensor, in orbit, Starcloud argues it can deliver faster, more secure inference at a lower operational cost than building ground stations [The Innovator]. If Starcloud can land these initial contracts, it plans to expand to enterprise customers later in 2026.
A Team Built for Hardware at Scale
The founders assembled a team with the specific credentials to get hardware into space and make it work. CTO Ezra Feilden brought a decade of satellite design experience from Airbus Defense and Oxford Space Systems [TechCrunch, December 2024]. Chief Engineer Adi Oltean was a Principal Software Engineer at SpaceX, where he worked on the tracking beam technology that lets Starlink communicate with other spacecraft [Y Combinator Hard Tech YouTube]. CEO Philip Johnston, a former McKinsey consultant who worked with national space agencies in the Middle East, provides the commercial and strategic lens [TechCrunch, December 2024].
| Founder | Role | Key Background |
|---|---|---|
| Philip Johnston | CEO | Ex-McKinsey, space agency strategy |
| Ezra Feilden | CTO | Satellite design, Airbus Defense, Oxford Space Systems |
| Adi Oltean | Chief Engineer | Principal Software Engineer, SpaceX |
The Scaling Hurdle: Cooling a Constellation
Traction so far is a proof of concept, not production. The company successfully trained lightweight AI models like NanoGPT and Google's Gemma on its first satellite, Starcloud-1 [News9live, 2025]. But the real test is scaling the infrastructure. Cooling consumes roughly 70% of the engineering team's attention, and the proof-of-concept system on Starcloud-1 cannot run continuously [Observer, March 2026]. The next satellite, Starcloud-2, slated for an October 2026 launch, is designed with a production-grade liquid-loop radiator system [Observer, March 2026]. The ultimate vision is a constellation of up to 88,000 satellites forming a distributed, 5-gigawatt data center [Data Centre Magazine, CNBC, December 2025].
The Realistic Competitive Set
Starcloud's competition is not other space-based data center startups, but the status quo: building more ground-based data centers and paying for the bandwidth and latency to shuttle data to and from orbit. Its advantages are its focused government ICP, its SpaceX pedigree on the engineering team, and its deep partnership with NVIDIA, which provides a clear roadmap to future GPU platforms like Blackwell [NVIDIA Blog].
What to Watch in the Next Twelve Months
The milestones are hardware-shaped and customer-focused. The successful launch and operation of Starcloud-2 with its production cooling system in October 2026 is the first critical gate. Following that, the company needs to announce its first contracted customer, almost certainly from its stated target list of NASA or the Department of Defense. A contract would validate not just the technology but the sales motion and pricing.