Orbital AI Compute: The Race for Space-Based Data Centers Heats Up

$22 million for Sophia Space. $170 million for Starcloud. $5 million for Orbital. And a filing for up to one million AI compute satellites from SpaceX.

The money flowing into orbital AI compute is still pocket change next to terrestrial data center CAPEX. But it marks a shift: space-based compute is no longer a thought experiment. It's an emerging asset class.

Growing: The startups building orbital compute

The most developed of the new entrants is Sophia Space, a Pasadena-based startup spun out of Mandala Space Ventures and Caltech. On June 23, the company announced a $7 million SAFE financing round, backed by EverGreen (the NVIDIA Alumni Investment Network) and SparkLabs Group, bringing total funding to $22 million.

Its thesis is built around a hardware module called TILE (Thermal Integrated LEO Edge), a one-square-meter panel that combines a solar array on one side with an NVIDIA compute module on the other. The key innovation is passive radiative cooling: instead of pumping heat through a compressor, TILE radiates it directly into the vacuum of space. "If you don't have a physical heat pump, your server is lighter, and most of the power generated goes toward compute," CEO Rob DeMillo told Via Satellite.

It has two in-orbit demonstrations planned. The first, scheduled for late 2026, will validate its orbital operating system on Kepler Communications' in-space network. The second, in 2027, will fly TILE hardware aboard an Apex Nova satellite bus, a standardised platform that lets the startup skip the spacecraft engineering and focus on the compute module. The company has undisclosed defense and Earth observation customers lined up for both missions.

As we wrote last week, Orbital Inc., backed by a16z Speedrun, raised a $5 million oversubscribed pre-seed round in June for its own orbital data center architecture. The Los Angeles startup emerged from stealth in April and plans a 10,000-satellite constellation of fridge-sized compute nodes, each with 100 kilowatts of power. Founder and CEO Euwyn Poon, previously the co-founder of scooter company Spin, has a background in infrastructure logistics and a contrarian view: "There simply isn't enough capacity here on Earth, and the only way is up."

Starcloud (formerly Lumen Orbit) is the furthest along. The Redmond, Washington startup launched its first GPU-equipped satellite in November 2025, ran a large language model in orbit in December, and raised $170 million in March 2026 on a $1.1 billion valuation, making it Y Combinator's fastest graduate to reach unicorn status. Starcloud has filed with the FCC for a constellation of up to 88,000 satellites and is preparing a second satellite, Starcloud-2, for commercial edge workloads later this year.

New: The hyperscalers enter orbit

June 2026 brought two announcements that changed the scale of the conversation.

SpaceX revealed its AI1 satellite on June 8, a dedicated orbital compute node with 150 kilowatts of peak payload power, a 70-meter wingspan, and a deployable liquid radiator spanning 110 square meters. The company has filed regulatory documents referencing a constellation of up to one million AI compute satellites. Elon Musk presented the design during the company's IPO roadshow, framing orbital AI inference as the logical extension of Starlink's global network.

Muon Space debuted the Condor-Ultra satellite platform on June 3, a high-power design delivering 20 to 100 kilowatts for payloads, explicitly targeting orbital data centers. The platform is compatible with Starship's stackable launch configuration and has been developed in consultation with hyperscalers. A pathfinder mission is planned for 2028.

Cowboy Space, formerly Aetherflux, filed in May for a 20,000-satellite constellation dedicated to data center services, taking an architecture approach where each upper stage becomes a one-megawatt data center in orbit. And Star Catcher raised $65 million in June (total $88 million) to build the first space-based solar power grid, an enabler for the entire sector.

Google, meanwhile, is quietly pursuing Project Suncatcher, an internal program to run AI workloads on solar-powered satellites, and reportedly exploring a launch deal with SpaceX.

Comparison: The architecture race

The variety tells you something: the sector has not converged on a winning architecture. Modular tiles compete with dedicated satellites and repurposed upper stages. The bet each company is making reflects different assumptions about how fast launch costs will fall, how quickly customers will materialize, and whether the primary use case is edge inference for existing satellites or wholesale cloud compute in orbit.

"The math said it would work, but the economics didn't quite work out. And then Leon raised his hand and said, 'Well, what if we put a server on this thing?'"— Rob DeMillo, CEO, Sophia Space, on the origin of the TILE architecture from Caltech's space solar research

The defense driver

A common thread across all five startups is the presence of defense and intelligence customers. Sophia Space has undisclosed defense customers lined up for both demo flights. Starcloud's FCC filing notes national security applications. True Anomaly's Golden Dome contract, which we covered at 14.6 hours' notice, signals the scale of government demand for space-based processing.

The logic is straightforward: proliferated LEO constellations generate more data than ground links can downlink. Processing that data in orbit (filtering, inferencing, deciding) reduces latency from minutes to milliseconds and cuts the bandwidth bottleneck. For missile defense, maritime surveillance, and electronic warfare, that difference is mission-critical.

Sophia Space's DeMillo frames it as a dual-use proposition: the same TILE module that processes Earth observation data for climate monitoring can run AI inference for threat detection. The hardware is identical; only the software layer changes.

The insurance market is watching. Reuters reported on June 18 that space startups have approached brokers like Marsh to explore coverage for orbital data center hardware, an early signal that the industry is preparing for commercial operations rather than just experiments.

The terrestrial constraint

The orbital compute thesis rests on one premise that is hard to argue with: Earth's power grid cannot keep up with AI's demand curve. US data center electricity consumption is projected to double by 2030. Permitting timelines for new substations run 4-7 years. Water for evaporative cooling is contested in every drought-prone region. Every new gigawatt of compute capacity on Earth is a political and environmental negotiation.

Orbit removes three of those four constraints at once. Solar energy is continuous and free. Passive radiative cooling consumes zero water. And no one needs a building permit in LEO. The remaining constraint, launch cost, has fallen by an order of magnitude in a decade and is still dropping.

Orbital compute will work technically. The physics is straightforward. The open question is economics: can it compete before terrestrial solutions like more efficient chips, advanced nuclear, or grid-scale storage close the gap?

Key signals to track

Sources