From One H100 to 200 Kilowatts: Starcloud's Orbital Data Center Roadmap Explained

Starcloud's orbital data center roadmap is easy to summarize and hard to execute: 60 kilograms and one Nvidia H100 became Starcloud-1; 450 kilograms and about eight kilowatts are next with Starcloud-2; three tons and 200 kilowatts are the planned Starcloud-3. But those are not merely larger versions of the same satellite. Each step crosses a threshold where the mission changes from proving compute can survive in space, to delivering useful capacity, to operating something customers can treat as a cloud region.

Step One: Prove the GPU Belongs in Orbit

Starcloud-1, a 60-kilogram spacecraft launched by SpaceX in November, carried the first Nvidia H100 GPU to low Earth orbit. Starcloud says the satellite trained an AI model and ran a version of Google's Gemini in space. That demonstration answered the first-order objection to orbital compute: modern accelerators are terrestrial data-center hardware, so can they function through launch vibration, radiation, vacuum, thermal cycling, and constrained communications? A successful workload is far more valuable than a component test because it validates the complete chain — processor, software, spacecraft, ground link, and operations.

Step Two: Make the Compute Useful

Starcloud-2 is planned as a 450-kilogram spacecraft on a SpaceX Falcon 9 rideshare in January. Its approximately eight kilowatts of power would be about 100 times Starcloud-1's output. The point is not just more GPU-hours. A useful commercial node needs enough capacity to run workloads concurrently, enough communications to accept jobs and return results, and enough operational margin to remain available while the spacecraft manages pointing, power storage, thermal limits, and contact windows. The jump from a successful demo to a product is the jump from can it run to can a customer depend on it.

Step Three: Turn a Spacecraft Into a Data Center

Starcloud-3 is the strategic hinge: a three-ton spacecraft designed to generate about 200 kilowatts for AI inference and training workloads uploaded from Earth. At that scale, the company is no longer demonstrating that a satellite can carry a computer. It is asking whether a spacecraft can perform the functions of a data-center building — generate and distribute power, reject heat, protect electronics, move data, maintain availability, and allow hardware to be refreshed — while moving at orbital velocity and operating beyond physical access.

Inference Is an Easier First Market Than Training

The customer workflow matters as much as the hardware. Training a frontier model requires enormous, tightly synchronized compute capacity and a data pipeline that may still be easier to manage on Earth. Inference — running a trained model against new data — can be more naturally distributed. A satellite constellation can process Earth-observation imagery near the sensor, reduce raw data into alerts, and downlink the result instead of every pixel. A defense customer may value the resilience and custody of orbital processing; an imagery operator may value a faster answer. Starcloud's early commercial case is strongest where the data is already in orbit and the answer is more valuable than the raw file.

• On-orbit Earth-observation analytics that convert imagery into detections before downlink
• Model inference for persistent monitoring where bandwidth and latency limit ground processing
• Distributed workloads that can tolerate intermittent connectivity and schedule around contact windows
• Specialized training or fine-tuning jobs using data collected by a particular mission
• Shared capacity for spacecraft operators that cannot justify a high-end accelerator on every satellite

The Constellation Is an Operations Business

Starcloud has filed plans with U.S. regulators for as many as 88,000 satellites. The number is a ceiling, not a deployment forecast, but it clarifies the ambition: a distributed orbital computing network rather than a handful of premium spacecraft. At that scale, the hard problems become fleet operations. Nodes need standardized software, workload scheduling, inter-satellite or ground networking, radiation-driven hardware replacement, launch campaigns, and a refresh cycle fast enough to keep the accelerator generation competitive. An orbital data center is therefore part satellite manufacturer, part cloud operator, and part launch customer.

The platform has to be designed for obsolescence as well as survival. Nvidia's Space-1 Vera Rubin Module is expected to offer up to 25 times the AI compute of the H100, which is good news for density and bad news for any spacecraft locked to a single generation. Starcloud's advantage will come from modular payload integration and a manufacturing cadence that allows the constellation to adopt better compute without waiting for every old satellite to retire.

The BlacKnight Take

Starcloud's roadmap should be judged as a sequence of commercial gates, not a sequence of impressive numbers. Starcloud-1 established that orbit can host a modern GPU. Starcloud-2 must establish that customers can use a meaningful amount of capacity. Starcloud-3 must establish that the power and thermal architecture can support data-center utilization. Only then does a constellation filing become a business plan rather than an aspiration. The company has earned the right to attempt the climb — the $250 million extension and Nvidia partnership provide real resources — but each step changes the physics and the customer promise. The next decisive metric is not teraflops in orbit. It is paid workload-hours delivered reliably enough that a customer renews.

Frequently Asked Questions