Starcloud Shifts Orbital AI Compute Plan Toward 88,000 Inference Satellites

The emerging orbital data-center model is not simply “cloud computing in space.” Ed Ludlow laid it out as a spacecraft architecture organized around compute, radiation tolerance, power generation, power storage, communications, flight systems, and heat rejection. Philip Johnston endorsed that basic description as accurate.

The headline number is 100 gigawatts: the annual AI compute capacity SpaceX says it ultimately wants to deploy in orbit, according to Ludlow. He framed that target as technically difficult not because the compute concept is mysterious, but because it implies building and maintaining networks of thousands, potentially millions, of specialized spacecraft in a harsh orbital environment.

Bloomberg’s mockup of a space-based data center looked less like a conventional facility than a satellite built around compute. Ludlow described a central body containing radiation-tolerant AI chips, surrounded by networking, power systems, thermal management, and flight control. For communication, the design could use laser links rather than conventional radio frequencies to move large volumes of data back to Earth. For power, it would need very large solar arrays, plus batteries for periods when the spacecraft is out of sunlight.

The hardest constraint Ludlow emphasized was heat. Terrestrial data centers use air or water to cool hardware. In orbit, heat has to be rejected through radiators into deep space. The mockup placed a radiator on the spacecraft and noted that those radiators would need to handle substantially more heat than current designs.

For Johnston, chief executive and co-founder of Starcloud, the premise is no longer purely speculative. Starcloud launched a satellite in November carrying five GPUs, including three from Nvidia and two from Arm. The Nvidia H100 was, in his words, “the most powerful, the most important and interesting one.”

Starcloud’s earlier concept was deliberately large. Philip Johnston described a render showing a four-kilometer-by-four-kilometer solar panel with a five-gigawatt compute cluster in the middle. A simplified diagram labeled the concept as a “5 GW Data Center” with a four-kilometer scale, alongside Starcloud imagery of spacecraft docking with large orbital structures.

Johnston said that large structure reflected assumptions from a couple of years earlier, when AI workloads were still mostly discussed in terms of training. The point of the render, he said, was to show that if a space system could handle the hardest workload, it could handle any workload. That version of the system also included, by his description, a radiator roughly one kilometer by four kilometers running down the back of the array.

But Starcloud is not now focused on docking together a large training facility. Johnston said “99% of all AI workloads will very soon be inference,” and that this changes the architecture. Instead of one enormous structure, Starcloud is pursuing a distributed constellation of much smaller satellites.

Johnston said Starcloud has filed with the FCC for a constellation of 88,000 small inference nodes. He said that constellation would enable roughly 20 gigawatts of compute. The shift matters because it changes the engineering problem from building one orbital megastructure to deploying many smaller compute satellites that can operate as a distributed network.

Starcloud’s supporting visuals showed both the hardware problem and the network model: cleanroom assembly of satellite wiring and components, a user sending an AI prompt, and satellites forming an orbital network connected by laser links. The useful distinction was not the polish of the render, but the workload assumption underneath it. Training led Starcloud to illustrate a huge docked structure; inference led Johnston to describe a distributed constellation.

Ed Ludlow made explicit why Starcloud was relevant to a discussion of SpaceX: SpaceX has pitched a vast network of orbital data centers, while its designs are still expected. Ludlow said Elon Musk had indicated he would show more in coming weeks. Starcloud, by contrast, is already flying compute hardware, though Ludlow characterized its current scale as limited and small.

Philip Johnston did not position SpaceX only as a competitor. He described space compute as potentially “the largest market opportunity ever,” with “trillions of dollars per year of CapEx” deployed in space. In that market, he said, everyone will need a space-compute solution. Some customers will use SpaceX, but others will want independent clouds, which Johnston identified as Starcloud’s target customer base.

Starcloud’s near-term commercial focus is narrower than replacing terrestrial cloud infrastructure. Johnston said the company initially plans to serve workloads for other spacecraft, providing edge and cloud services to them. That makes the first market an orbital one: compute located near other space assets, rather than only compute in space serving users on Earth.

At the same time, Starcloud depends on SpaceX for launch. Johnston said Starcloud is a “very happy customer” of SpaceX’s rideshare program and is potentially interested in dedicated Falcon 9 launches. He did not provide details, but said the company should have “interesting things to say” soon.

The relationship is therefore layered: SpaceX’s own orbital data-center ambitions set the scale of the discussion; SpaceX launch services may help Starcloud deploy; and Starcloud is pitching independent orbital cloud capacity to customers that may not want to rely entirely on SpaceX.

Asked by Ed Ludlow about the ripple effect of a SpaceX IPO, Philip Johnston said Starcloud is “already seeing it,” in the form of enormous interest in the space industry that was not there before. He linked that change to investors understanding space as a large market opportunity, in large part because of the expected SpaceX IPO.

Johnston’s market claim was expansive: he described orbital compute as a potential market for trillions of dollars per year in capital expenditure. Starcloud’s demonstrated scale, as described in the interview, is much smaller: one satellite launched in November carrying five GPUs, including one Nvidia H100. Its next-order claim is not deployed capacity but a filed plan for 88,000 small inference nodes that Johnston said would enable about 20 gigawatts of compute.

He cited Starcloud’s financing as evidence of investor appetite. The company, he said, raised the fastest unicorn round coming out of Y Combinator ever: in 17 months, it went from essentially zero to a $1.1 billion valuation with a $170 million raise. An on-screen graphic listed Starcloud backers including NFX, Nvidia, Y Combinator, Benchmark, and IQT.

Johnston’s claim was not that the technical risks have disappeared. It was that orbital compute is now being treated as financeable at a scale commensurate with its power, launch, cooling, communications, and satellite-manufacturing demands.