Why it matters
Starcloud has operated an Nvidia H100 in orbit and is financing the claim that this can become a power and infrastructure business rather than a one-satellite test. CNBC reported in December 2025 that Starcloud-1 was querying Google's open Gemma model on that GPU, which it called the first large language model run on a high-powered Nvidia GPU in space, and that the same satellite trained Andrej Karpathy's nanoGPT, which it called the first model trained in space. Philip Johnston, who co-founded the company in 2024, told CNBC he expected orbital data centers to have about 10 times lower energy costs than terrestrial ones.
That cost claim is a target with a launch condition attached. In March 2026 Johnston told TechCrunch he expected about $0.05 per kilowatt-hour if commercial launch reached about $500 per kilogram, with Starship access in 2028 and 2029, and he said Starcloud would not be competitive on energy cost until Starship flew frequently. By 21 August the nearer constraint was securing rockets. He told TechCrunch that Falcon 9 is scheduled to end in 2028, that the company has asked the FCC for authority to operate 88,000 spacecraft, and that failing to book any SpaceX launch capacity in 2029 would be challenging.
What Starcloud-1 demonstrated
The December demonstration was Gemma, not the Gemini product Google sells. CNBC quoted a downlink beginning "Greetings, Earthlings!" in which the model named itself Gemma, and it quoted Google DeepMind's Tris Warkentin on Gemma running in space. The training run was nanoGPT on the works of Shakespeare. CNBC also reported inference on imagery from Capella Space, the near-term case of processing another operator's data in orbit. In March 2026 Johnston described that same flight to TechCrunch as a version of Gemini. The contemporaneous report, the model's own text, and DeepMind's comment identify Gemma.
A second Nvidia GPU did not survive launch. Johnston told TechCrunch that an A6000 failed during launch. The spacecraft itself is still catalogued. KeepTrack, using a CelesTrak element set with epoch 14 October 2026, lists STARCLOUD-1 as NORAD 66303, an operational payload at 45.40 degrees inclination, with apogee and perigee near 509 and 503 kilometers, launched on 2 November 2025. Altitude had eased only modestly by 13 October. That listing says the satellite is in orbit. It does not say the H100 is still running customer work, and a 45.40-degree low Earth orbit is not the constant-solar regime in the company's energy pitch.
The spacecraft after the demonstrator
In March 2026 Starcloud-2 was described as a later-this-year satellite with Nvidia Blackwell GPUs, an AWS server blade, and a bitcoin-mining computer, plus what Johnston called the largest deployable radiator flown on a private satellite. On 21 August he told TechCrunch the plan was now two 8 kW Starcloud-2 satellites on rideshare flights in 2027, for customers including U.S. government agencies, with a dedicated Falcon 9 under consideration. The same round is meant to advance Starcloud-3, a 200 kilowatt, three-ton spacecraft sized for the dispenser SpaceX designed to deploy Starlink satellites from Starship, and to develop production lines at a 100,000-square-foot site in Woodinville, Washington.
SC-1L is a hosted payload, not another free-flying Starcloud satellite. Firefly said on 30 September 2026 that SC-1L will ride Elytra on Firefly's third lunar mission, targeted no earlier than 2028. Elytra is to deliver Blue Ghost to the Gruithuisen Domes under NASA's CLPS program and then remain in lunar orbit for five years. The payload is to ingest Firefly's Solux vision data, compute on board, and return processed data. Ezra Feilden tied it to the low Earth orbit GPU flights. Firefly stated no price. Nvidia's Vera Rubin Space-1 module, which Johnston hopes to fly in late 2028, had not been built as of the August interview. TechCrunch attributed a $25 million Nvidia check in the extension to a person familiar with the deal, not to a company figure.
Financing and the commercial path
The March Series A was $170 million at a $1.1 billion valuation, led by Benchmark and EQT, and it closed 17 months after Y Combinator demo day. TechCrunch put total capital at that point at about $200 million. The 21 August extension added $250 million at a $2.3 billion valuation, led by Manhattan West Ventures, with Nvidia, Cisco, Benchmark, EQT, Soma, NFX, 776, Cedar Capital, Goanna Capital, and Standard Capital participating. Johnston said the company was 25 employees and growing. These reports do not state recognized revenue.
Johnston told TechCrunch there are two businesses: processing sold to other spacecraft, with the Capella work as the example already claimed, and, if launch costs fall, workloads pulled off terrestrial data centers. Bitcoin-mining hardware on Starcloud-2 was part of the published payload. CNBC reported in December 2025 that a Crusoe cloud module was planned for the satellite then scheduled for October 2026. That month has since moved to 2027 rideshares, and these accounts do not show Crusoe restating the date.
Competition and constraints
Johnston told TechCrunch in March that SpaceX has asked to operate as many as a million satellites for distributed compute, aimed mainly at Grok and Tesla workloads, and that he does not expect SpaceX to sell the energy-and-infrastructure service Starcloud wants to offer. He also named Aetherflux, Google's Project Suncatcher, and Aethero. Starcloud's own scale case is the FCC request for up to 88,000 spacecraft, which the August interview still described as a request. These accounts do not report a grant.
CNBC cited Morgan Stanley analysts on radiation, in-orbit maintenance, debris, and data governance. The H100 flight did not close those issues, and the public catalog on 14 October 2026 still showed one Starcloud payload, NORAD 66303, at 45.40 degrees near 500 kilometers. Operational in that listing means the object is in orbit. It is not evidence of paid GPU hours or of the constant-solar regime in the company's energy pitch.
What to watch
The next proof is whether two 8 kW satellites fly on 2027 rideshares, and whether those orbits test the solar case that Starcloud-1, at 45.40 degrees, has not tested. Woodinville production and a Starcloud-3 Starship slot remain 2028 and 2029 targets gated on launch price and vehicle availability. Johnston said in August that missing SpaceX capacity in 2029 would be a serious problem, and that the company was also looking at a dedicated Falcon 9 and other providers.
SC-1L is a separate item: a hosted lunar-orbit computer on Firefly's no-earlier-than-2028 Elytra mission, with five years of planned time on orbit and no published price. Vera Rubin Space-1 is a late-2028 hope for a chip that had not been built. Also unresolved in these sources: whether the H100 is still running Capella or any other customer job, whether Crusoe or AWS has a current date, whether the FCC request has moved past a filing, and which staffing figure is current. Johnston's August figure of 25 people is the only headcount in the five sources used here.