Nvidia Just Bet $25M on AI in Orbit — Starcloud Raises $250M to Build the Orbital Inference Layer

Starcloud closed a $250 million extension round on August 21, 2026, pushing its valuation to $2.3 billion — and the most notable line item in the cap table is Nvidia's $25 million check, the chip giant's first direct investment in a company running its GPUs in low Earth orbit. The round was led by Manhattan West Ventures, with Cisco, Benchmark, EQT, Soma, NFX, and 776 also participating.

Background You Need

Before November 2025, putting serious compute in orbit meant accepting hardware from a different era. The RAD750 — a radiation-hardened processor that has guided Mars rovers and military satellites — delivers roughly 266 MIPS of processing power. That's a number that made sense when the goal was surviving a decade in space, not running a language model. The gap between what space-qualified hardware could do and what AI workloads actually require was enormous.

Starcloud, a 25-person startup based in Woodinville, Washington, bet that the gap was bridgeable with commercial off-the-shelf silicon and fast iteration. BlacKnight Space Labs, which reviewed the mission in technical detail, reports that the company built its first satellite — Starcloud-1, a 60-kilogram spacecraft — in just 21 months on $3 million in pre-seed funding. It launched aboard a SpaceX rocket on November 2, 2025, carrying an Nvidia H100 GPU with approximately 4 petaflops of AI compute power in a package roughly the size of a dorm-room refrigerator. BlacKnight puts that at about 100 times more powerful than any GPU previously operated in orbit.

The rationale for moving compute to space isn't primarily about novelty. Data Center Frontier's analysis of the launch points to a concrete terrestrial problem: AI data centers are running into hard limits around land, power, and water cooling. Orbital infrastructure offers solar energy and radiative cooling as essentially free inputs, with no land-use permitting required. One source puts the potential energy cost advantage at up to ten times lower than comparable ground facilities — though that figure is vendor-adjacent and has not been independently verified at commercial scale.

What's New

Starcloud-1 didn't just survive launch. Multiple sources confirm it achieved four distinct milestones in the months following deployment. The satellite trained NanoGPT — a compact language model created by OpenAI co-founder Andrej Karpathy — on the complete works of Shakespeare, the first time any AI model was trained on orbital hardware. It then ran inference using Google's Gemma model, an open-source implementation of the Gemini architecture. It performed fine-tuning of a pre-trained network in orbit. And in doing all of this, it validated that commercial GPUs — not radiation-hardened, space-qualified chips — can operate through thermal cycling, radiation exposure, and launch vibration. Karpathy publicly acknowledged the accomplishment, and Demis Hassabis of Google DeepMind congratulated the Starcloud team for running Gemma in space, per the company's own site.

That operational record is what moved Nvidia to write the $25 million check. According to TechCrunch's reporting on the round, Nvidia described the investment as requiring far more technical due diligence than a typical venture investment — a signal that the company treated this as hardware validation, not a financial bet. This is significant because Nvidia doesn't routinely take minority stakes in customers; it builds tools and sells chips. A direct equity investment suggests the company sees Starcloud's orbital operational data as useful for its own product roadmap.

That roadmap has a name: the Vera Rubin Space-1. It will be the first GPU purpose-built for the space environment — with radiation shielding, thermal management for orbital extremes, and structural tolerance for launch G-forces built into the design from the start. It doesn't exist yet; Nvidia is targeting orbit for late 2028. Starcloud is actively feeding operational data from Starcloud-1 to Nvidia's design team, making the startup a de facto test partner for the chip's requirements.

The $250 million raise is also a response to a supply-chain risk that has nothing to do with chips. Falcon 9, currently the most reliable and cost-effective heavy-lift launch vehicle available, is expected to retire around 2028. SpaceX's Starship is the intended successor, but it remains unproven for commercial satellite rideshare at scale. Blue Origin's New Glenn and ULA's Vulcan are not flying with regularity; Rocket Lab's Neutron is not yet on the pad. Starcloud has requested FCC permission for 88,000 spacecraft and is planning its Starcloud-2 constellation — 8 kW compute satellites, with rideshare launches targeted for 2027 — but locking in launch contracts before Falcon 9 winds down requires capital now. Part of this raise is launch capacity insurance.

The Pushback

The commercial case for orbital compute isn't settled. Data Center Frontier's December 2025 analysis — the most measured of the corroborating sources — notes explicitly that while Starcloud-1 proved technical feasibility, the economics of launch, deployment, and long-term operation remain open questions. A single H100 in orbit at experimental scale is not a data center; it is a proof of concept. The cost per GPU-hour in orbit has not been published, and the energy efficiency claims don't account for the capital cost of placing and maintaining hardware in space in the first place. Data Center Frontier called these "critical" unanswered questions for the transition from experimental to commercial viability — a notably more cautious framing than the other sources reviewed here.

Our read is that the real validation story is narrower than the headline figures suggest. Starcloud's current customers are US government agencies with use cases — real-time satellite telemetry processing, surveillance, disaster detection — where the data originates in space and transmitting it to the ground creates latency and bandwidth problems that orbital inference directly solves. That's a specific, defensible niche. Scaling to general-purpose commercial AI inference is a separate claim, one that depends on both Starship becoming a reliable rideshare option and the Vera Rubin Space-1 GPU delivering on specifications after it reaches orbit sometime after 2028. The launch market fragility also cuts both ways: if Starship matures and costs drop, Starcloud's window to lock in favorable contracts closes; if it doesn't, the 88,000-spacecraft constellation faces a hard bottleneck regardless of how good the hardware gets.

Sources