First AI Training in Space: How Starcloud Put an NVIDIA H100 in Orbit and Changed the Game

Before November 2025, the most powerful computing hardware ever sent to space was measured in single-digit teraflops. Radiation-hardened processors designed for spacecraft — built to withstand the charged particle environment of orbit — typically lag several generations behind their terrestrial counterparts. The RAD750, a workhorse processor used on Mars rovers and military satellites, delivers roughly 266 MIPS of computing power. Compared to the processors powering AI on the ground, space computing existed in a different era entirely.

Starcloud-1 changed that calculus in a single launch. The 60-kilogram satellite, built in just 21 months on $3 million in pre-seed funding, carried a commercial NVIDIA H100 GPU into low Earth orbit — a chip with 80 GB of high-bandwidth memory and approximately 4 petaflops of AI compute power. It was 100 times more powerful than any GPU previously flown in space. And within weeks of reaching orbit, it had demonstrated something no one had done before: training an artificial intelligence model in space.

The Mission: Four Firsts in Orbit

Starcloud-1 achieved four historic milestones that collectively proved the viability of high-performance AI computing in orbit. First, the team trained NanoGPT — a compact language model created by OpenAI founding member Andrej Karpathy — on the H100 using the complete works of Shakespeare. The model learned to generate Shakespeare-like text while the satellite completed orbits at 28,000 kilometers per hour, marking the first time any AI model had been trained on orbital hardware.

Second, the satellite ran Google's Gemma model — an open-source implementation based on the Gemini architecture — demonstrating that large language model inference works on high-performance GPUs in the space environment. Third, the team performed model fine-tuning in orbit: taking a pre-trained neural network and adapting it to new data while the satellite was in space. Fourth, by successfully operating the H100 through thermal cycling, radiation exposure, and the vibration environment of launch, Starcloud demonstrated that commercial AI chips — not radiation-hardened, space-qualified parts — can function reliably in LEO.

The Radiation Question

The single most common objection to commercial GPUs in orbit is radiation. Space is filled with high-energy particles — cosmic rays, solar proton events, and trapped radiation in the Van Allen belts — that can flip bits in memory, corrupt computations, or permanently damage semiconductor junctions. This is why the aerospace industry has traditionally used radiation-hardened processors: chips designed with redundant circuits, error-correcting memory, and hardened transistor structures that can withstand the orbital environment at the cost of vastly reduced performance.

Starcloud's approach bets on a different calculus. In low Earth orbit at typical inclinations, the radiation environment is orders of magnitude less severe than in geostationary orbit or deep space. Single-event upsets (bit flips) occur, but at rates that can be managed through software-level error correction, checkpoint-and-restart strategies, and ECC (error-correcting code) memory — features that the H100 already includes for terrestrial data center reliability. The economics favor this approach: a single radiation-hardened processor might deliver 1/1000th the computing power of an H100 at comparable cost. If software can handle the occasional radiation-induced error, the performance advantage of commercial GPUs overwhelms the reliability advantage of rad-hard chips.

The $3 Million Satellite

Perhaps the most remarkable aspect of Starcloud-1 is its cost. The satellite was designed, built, and launched on $3 million in pre-seed funding — a budget that would barely cover the cost of a single radiation-hardened space processor in traditional aerospace procurement. The team accomplished this by using commercial-off-the-shelf (COTS) components wherever possible, leveraging SpaceX rideshare launches for low-cost orbit access, and applying the rapid iteration methodology that Adi Oltean brought from SpaceX's Starlink manufacturing program.

The 21-month timeline from company founding to satellite launch is equally notable. Traditional satellite programs take 3–7 years from concept to launch. Starcloud compressed this timeline by treating the satellite as a technology demonstration rather than a production system — accepting higher risk in exchange for faster learning. The fact that the H100 worked in orbit, the AI training succeeded, and the thermal management held up validated the core technical thesis and unlocked the $170 million Series A that followed.

What Starcloud-2 Will Change

Starcloud-2, planned for October 2026, represents a 100x increase in computing capacity. The satellite will carry NVIDIA Blackwell B200 GPUs — the successor to the H100 — alongside multiple H100s and an AWS server blade. The power system scales from Starcloud-1's approximately 1-kilowatt solar array to 7 kilowatts, enabling continuous high-performance computing rather than the duty-cycled operation of the demonstration satellite.

The inclusion of an AWS server blade is strategically significant. It suggests Starcloud is already working with Amazon Web Services on orbital computing integration — potentially enabling customers to treat orbital GPU capacity as just another availability zone in their cloud infrastructure. If orbital compute can be accessed through familiar cloud APIs, adoption barriers drop dramatically. The mission will also carry a Bitcoin mining payload, likely as a commercial demonstration to offset mission costs and prove the economic viability of orbital processing for compute-intensive workloads.

Implications for Space Computing

Starcloud-1 fundamentally shifts the conversation about computing in space. Before this mission, space computing meant radiation-hardened processors running at a fraction of terrestrial performance — adequate for satellite housekeeping and basic image processing but incapable of AI workloads. After Starcloud-1, it is demonstrated fact that commercial AI accelerators can operate in LEO, that AI models can be trained in orbit, and that the thermal and radiation challenges are manageable with engineering rather than exotic components.

For the broader space industry, this opens possibilities beyond data centers. Earth observation satellites could run real-time AI analysis on imagery before downlinking, transmitting insights instead of raw data. Communications satellites could perform edge AI processing for latency-sensitive applications. Military satellites could run autonomous decision-making algorithms without depending on ground links. Starcloud-1 proved the hardware works. The question now is how fast the applications will follow.

Frequently Asked Questions