The Sky Becomes a Data Center: The Race to Put AI in Orbit

Over the past nine months, low Earth orbit has become a contested layer of the AI infrastructure war, with SpaceX filing for a million-satellite AI constellation, signing roughly $26 billion in annualized AI compute agreements with Anthropic and Google, and announcing a city-sized factory to build the hardware. Starcloud ran Google's Gemma on an H100 in orbit in November 2025 and deployed Starcloud-1, a 50-kilogram GPU-class data center satellite, in March 2026, while China's Three-Body Computing Constellation completed nine months of in-orbit testing with 10 AI models across satellites. In July 2026 a satellite operated the first multimodal AI combining vision and language understanding entirely in orbit, cutting raw data downlink requirements.

From China's Xingshu Plan satellites to SpaceX's 1-million-satellite Starmind filing and the SpaceX and xAI Gigasat orbital AI factory, every major AI power is racing to move compute into low Earth orbit. The first multimodal AI ran autonomously in space in July 2026, orbital intercept was validated, and SpaceX signed roughly $26 billion in annualized AI compute agreements with Anthropic and Google, turning orbit into a contested layer of the AI infrastructure war. The sky is no longer just a communications medium. Over the past nine months, low Earth orbit has become the newest front in the AI infrastructure war, with American and Chinese actors racing to deploy not just connectivity but raw compute into space. The pace of change is striking: in November 2025, a startup ran the first large language model in orbit on a single GPU; by July 2026, SpaceX had filed for a million-satellite AI constellation, signed roughly $26 billion in annualized AI compute agreements with Anthropic and Google, and announced a factory the size of a small city to build the hardware. The strategic logic is straightforward: orbital compute is unjammable by geography, scalable without land permits, and increasingly capable of acting autonomously without waiting for a ground station to issue instructions. Why does this matter now? Because the transition from experiment to infrastructure is happening faster than most analysts anticipated, and the decisions being locked in today, spectrum filings, manufacturing commitments, bilateral compute contracts, will shape who controls this layer for decades. The starting point is modest but precise. In Starcloud Trains First AI Model in Space https://dev.to/e/737 in November 2025, the Nvidia-backed startup ran Google's Gemma model on a single H100 GPU aboard a satellite. That was a proof of concept, not a product. Four months later, in March 2026, Starcloud followed with Starcloud-1, the First GPU-Class Data Center in Orbit https://dev.to/e/1168 , a roughly 50-kilogram satellite carrying that same H100 into low Earth orbit, the first time a data-center-class compute node had been physically deployed in space. The hardware roadmap accelerated immediately. At GTC 2026 in March, Nvidia unveiled the Vera Rubin Space-1 Module https://dev.to/e/266 , a space-hardened computing platform delivering up to 25 times the AI compute of an H100. That single announcement collapsed the generational gap between what was flying and what could fly. By April, Kepler Communications opened a 40-GPU orbital compute cluster https://dev.to/e/1231 using Nvidia Orin processors across multiple satellites, the largest GPU network in orbit at that point, with Sophia Space as its first commercial customer validating orbital data-center software. The software stack kept pace with the hardware. In April 2026, Loft Orbital's YAM-9 satellite ran Google DeepMind's Gemma 3 https://dev.to/e/1391 via NASA JPL's NAVI-Orbital software, becoming the first Earth-observation satellite to autonomously detect and classify targets without ground-analyst involvement. Then in July 2026, a satellite operated the first multimodal AI combining vision and language understanding https://dev.to/e/5015 entirely in orbit, identifying targets in real time and cutting raw data downlink requirements. The significance is architectural: when inference moves to the satellite, the ground station becomes an endpoint rather than a bottleneck. China was not watching from the sidelines. In January 2026, a Chinese commercial aerospace firm achieved the first deployment of a general-purpose AI model aboard orbiting satellites https://dev.to/e/148 in operational use. By February, China had completed nine months of in-orbit testing for its Three-Body Computing Constellation https://dev.to/e/186 , deploying 10 AI models across satellites with demonstrated inter-satellite networking, the first operational space-based AI edge-computing network anywhere. The scale ambition then became explicit. In July 2026, Shanghai Xingshu Tiansuan Space Technology launched the first satellites of a planned 1,000-satellite constellation https://dev.to/e/5014 designed specifically for space-based AI computing infrastructure. The Xingshu Plan, as it is formally known, represents China's first operational orbital AI data center deployment at constellation scale, with the first satellites launched on July 18, 2026. The timing, within days of SpaceX's own major announcements, underscores that both sides understand the strategic value of establishing orbital compute presence early. The Chinese approach differs from the American one in one important structural way: it is state-adjacent commercial development, with Shanghai municipal backing visible in the Xingshu announcements, rather than purely private capital. That distinction affects how quickly regulatory and spectrum coordination happens domestically, and how aggressively the constellation can expand without shareholder pressure. The single most consequential structural move of this period was the SpaceX acquisition of xAI in a $1.25 trillion deal https://dev.to/e/1108 announced in February 2026. The stated rationale was explicit: merge Starlink's orbital infrastructure, Grok's AI models, and xAI's compute into a unified space-based AI data center business. That is not a product announcement. It is a vertical integration play that eliminates the boundary between launch, connectivity, compute, and AI model serving. The regulatory and commercial follow-through came fast. In June 2026, SpaceX filed with the FCC for Starmind https://dev.to/e/560 , a constellation of up to one million AI compute satellites targeting 100 gigawatts of orbital AI compute power. That filing is the largest orbital compute regulatory claim in history by a wide margin. In July, SpaceX revealed the AI1 satellite design https://dev.to/e/3626 : a 70-meter wingspan, 150-kilowatt solar array, 120 to 150 kilowatts of AI compute payload, with laser inter-satellite links to beam compute to Earth's surface. Launches are targeting late 2027. The commercial validation arrived on July 19, 2026, when SpaceX signed roughly $26 billion in annualized AI compute agreements with Anthropic and Google https://dev.to/e/4601 , more than doubling the company's total revenue built over two decades. SpaceX also announced plans to manufacture AI chips with Terafab for orbital data centers. Three days later, SpaceX and xAI announced the Gigasat facility in Bastrop, Texas https://dev.to/e/6056 : 11 million square feet of satellite manufacturing space targeting 1 gigawatt of orbital AI compute capacity by end of 2027. The Anthropic and Google contracts deserve particular attention. Both companies are signing for orbital compute capacity that does not yet exist at scale. They are, in effect, pre-purchasing a layer of infrastructure whose physics, latency, and reliability characteristics are still being validated. That is a significant bet, and it tells you something about how seriously hyperscalers are taking the terrestrial power and land constraints that are currently throttling their ground-based data center expansion. One event sits slightly apart from the infrastructure narrative but connects directly to it. On July 3, 2026, the U.S. Space Force confirmed that True Anomaly's autonomous Jackal spacecraft intercepted and characterized Rocket Lab's Puma satellite https://dev.to/e/585 in 61 hours, the first autonomous commercial orbital intercept ever validated. The mechanism matters: Jackal used onboard AI to navigate, approach, and characterize a target without real-time ground control. This is the military application of the same stack being built for commercial orbital compute. When satellites can act autonomously, identify targets, and maneuver without waiting for ground instructions, the distinction between an orbital data center and an orbital sensor-shooter collapses. The Gigasat factory, the Starmind filing, and the True Anomaly intercept are all expressions of the same underlying capability curve. Starmind spectrum coordination : The FCC filing for one million satellites will require international spectrum and orbital slot coordination through the ITU. Watch for Chinese or European objections that could constrain the constellation's operational parameters before a single AI1 satellite launches. Vera Rubin Space-1 first flight : Nvidia's purpose-built orbital AI module is the hardware that makes the economics of Starmind and Xingshu viable at scale. The first operational deployment will set the benchmark for compute-per-watt in orbit and determine whether the 2027 launch targets are credible. Anthropic and Google compute delivery timelines : The roughly $26 billion in annualized contracts are contingent on SpaceX actually delivering orbital compute capacity. If AI1 launches slip past late 2027, watch for contract renegotiation or parallel terrestrial fallback investments from both customers. Xingshu Plan constellation cadence : China's first Xingshu satellites are in orbit, but the path from initial deployment to 1,000 satellites is the real test. Launch rate and inter-satellite networking performance will determine whether China's orbital AI network reaches operational parity with Starmind before 2030. Autonomous on-orbit inference regulation : The July 2026 multimodal AI milestone means satellites can now identify and classify targets without human review. No international framework currently governs autonomous orbital inference for dual-use applications. The first regulatory or treaty proposal in this space will be a leading indicator of how governments intend to manage the military-commercial boundary in orbit. This piece was originally published on Present of AI https://presentofai.com/insights/orbital-ai-race , where we cover what AI is actually doing in the world, no hype. Read more https://presentofai.com or get it in your inbox https://presentofai.com .