The race to build ever larger AI models has created an unexpected bottleneck. It is no longer chips that limit progress, but power. Modern data centers already consume vast amounts of electricity, and demand is rising faster than infrastructure can keep up.
Orbital believes the solution lies far above the grid.
Article continues below
The company is developing AI data centers designed to operate in low Earth orbit, powered entirely by solar energy and cooled by radiating heat directly into space. Without weather, night cycles, or grid dependency, solar arrays in orbit can generate continuous power, while the vacuum of space provides a natural way to dissipate heat — two constraints that dominate the economics of terrestrial data centers.
Orbital-1
Backed by funding from a16z Speedrun, Orbital is preparing its first test mission, Orbital-1, scheduled to launch on a SpaceX Falcon 9 in April 2027.
The satellite will host Nvidia-powered compute hardware and is intended to validate sustained GPU operation in orbit, test radiation resilience, and begin running AI inference workloads once initial validation is complete.
The company chose inference over training because inference tasks can run independently across a satellite network. Unlike training clusters, which require tightly coupled GPUs operating in near-perfect synchronization, inference tasks can be distributed across many independent nodes — making them better suited to satellite constellations.
The idea of running data centers in space may sound radical, but the pressures driving it are becoming increasingly real.
I wanted to know more so I spoke to Orbital founder Euwyn Poon about the economics, engineering challenges, and practical realities of running AI infrastructure in orbit.
- Why is everyone suddenly looking at space as the new frontier for AI infrastructure? What has changed? (cratering cost of putting goods in orbit, the OPEX appeal, exploding AI demand, the lack of red tape in space?)
A few things converged at once. Launch costs are going to collapse with Starship, going from $7,000/kg on the Falcon 9 today towards the target of $10/kg.
Meanwhile, AI demand is skyrocketing beyond the capacity of the power grid, with US data centers using about 25 gigawatts today and growing 3 to 4 times by 2030. Building gigawatt data centers in the US is also becoming increasingly difficult.
Communities across the US are pushing back against data center construction. Getting a gigawatt facility permitted and connected to the grid now involves a ton of uncertainty and risk.
- More specifically on the energy infrastructure, can you give us a sense of how much a MwH of power would cost in space as opposed to Earth-based (or perhaps how easily would it be to get solar power up there).
On Earth, data center electricity runs $60 to $100 per megawatt-hour before cooling. Add the 40% thermal overhead that every facility carries and you land at $85 to $140 per MWh in effective energy cost.
In orbit, the marginal cost of energy is zero. The sun delivers 1,361 watts per square meter in LEO, constantly, with no fuel cost, no grid fees, no utility contract.
What you actually pay is the amortized capital cost of the solar array and launch costs. Our ultimate goal is to push amortized space power below $10 per MWh.
- What could prevent rogue nations from disabling/harming/taking down these orbital data center networks in the sky à la Moonraker? What sort of resilience features are being implemented to prevent that from happening?
The short answer is that distributed constellations are inherently harder to attack than buildings on the ground. We are not building one big space station. We are deploying thousands of small, independent satellites spread across multiple orbital planes. Taking out a single satellite removes a fraction of a percent of total capacity.
To meaningfully degrade the network, you would need to disable hundreds or thousands of targets simultaneously. That is an enormously expensive and conspicuous military operation.
On the legal side, the Outer Space Treaty of 1967 establishes that no nation can claim sovereignty over space. Satellites remain under the jurisdiction of the launching country, but no nation can seize or shut down another country’s satellite without committing an act of war.
Every terrestrial data center can be raided with a warrant. Orbital infrastructure cannot.
- How does one deal with the extreme heat differences in space where I assume one side of the satellite will remain immensely hot and the part in the shadows, a few Kelvins above absolute zero (dissipation of compute heat but also thermal management).
The thermal environment in space is extreme. On the sun-facing side, solar irradiance hits at 1,361 watts per square meter. On the shadow side, space is roughly 3 Kelvin. Inside the satellite, you have GPU waste heat with no air for convection.
On Earth, you blow air over heat sinks or run water through cooling towers. In vacuum, none of that works. Managing this thermal environment is one of our core engineering efforts that we are not yet discussing publicly in detail.
- Given your focus on inference, how do you plan to deal with signal latency for real-time AI applications? Would latency be inversely proportional to the number of satellites deployed?
At 550 kilometers altitude, the speed-of-light round trip is under 4 milliseconds. A typical API call to a terrestrial cloud provider already takes 20 to 100 milliseconds. For inference workloads like chatbots, code generation, and agentic AI, users already wait hundreds of milliseconds for a response. The orbital penalty is imperceptible.
More satellites helps, but it is not inversely proportional. For inference, these are independent, parallel requests, not a single job being split across multiple nodes.
- Your satellites are going to be launched using Space X’s platform. Space X has already disclosed its ambitious to be a major player in YOUR field. How do you plan to tackle this David vs Goliath conundrum?
This is a big space, and we are in different lanes. Elon merged SpaceX with xAI, and that validates the entire thesis. The biggest startup risk is that the category does not exist. SpaceX just eliminated that risk for us.
Elon has announced he will focus on its own custom silicon and its own models. That creates an opening for other operators like us. Just as there are many hyperscalers on earth, there is room for many winners here.
- What happens if the satellites reach EOL? Will they be decommissioned or sent to burn in the Earth atmosphere or actually never decommissioned? What’s the rough lifespan you expect to eek out from them?
Every satellite deorbits and burns up in the atmosphere. Each satellite carries ion thrusters with propellant reserved specifically for the deorbit maneuver. The structure is made from materials that burn up cleanly on re-entry.
Design life is roughly matched to the useful life of the GPU architecture on board, around 5 to 7 years. By that point the chips are multiple generations behind and the economics favor launching fresh satellites with current hardware rather than maintaining older ones.
- What else is on Orbital’s roadmap? Autonomous robots? Strategic partnership with hyperscalers? Different class/size of orbital data centers — e.g. SSO for training/LEO for inference? Self repairing ones?
Right now we are focused on two things: our test mission and our factory. Orbital-1 launches in early 2027 In parallel, we are standing up Factory-1, a robotic satellite assembly facility in Los Angeles.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
https://cdn.mos.cms.futurecdn.net/Sj3zGjiiDaJzfGKaFSAzf5-1600-80.png
Source link
desire.athow@futurenet.com (Desire Athow)
