Somewhere between the stratosphere and the thermosphere, in the thermally forgiving vacuum of low-Earth orbit, a small constellation of computing nodes is doing something that would have sounded like science fiction eighteen months ago: running production AI workloads. Not as a demo. Not as a proof of concept. As actual, billable, operational infrastructure. The data center, it turns out, was never the destination. It was a layover.

🤚 The Open-Palm Illumination

Peter Diamandis’s latest MOONSHOTS dispatch poses a deceptively simple question: what if data centers are not the endpoint of scaling computing? The answer, as it happens, is floating approximately 550 kilometers above your head.

January 11, 2026 marked the first time in history that multiple orbital data center operators simultaneously ran production workloads in space. By February, this was no longer a milestone — it was a baseline. The orbital compute revolution is not coming. It has a mailing address. That address is in low-Earth orbit, and the rent is surprisingly reasonable.

The logic is almost embarrassingly elegant. A solar array in space generates over five times the energy of the same array on Earth, with an equivalent energy cost of approximately $0.005 per kWh — up to 15 times lower than today’s wholesale electricity prices. Cooling? In a vacuum, passive radiative cooling achieves temperatures that terrestrial data centers spend fortunes chasing with water-guzzling chillers. No water consumption. No energy-intensive cooling infrastructure. Just the cold, silent efficiency of space itself.

👐 The Two-Handed Reality Check

The players assembling in orbit read like a who’s-who of organizations that do not typically share conference rooms. NVIDIA announced that its latest accelerated computing platforms are being deployed in orbital data centers, bringing AI compute to geospatial intelligence and autonomous space operations. SpaceX has filed for a constellation of up to one million satellites to create orbital computing infrastructure. Google is exploring putting Tensor Processing Unit clusters in space. And Planet Labs — the satellite imaging company whose CEO Will Marshall appeared on Diamandis’s show — has already demonstrated the concept by running AI-powered object detection directly onboard its Pelican-4 spacecraft, powered by NVIDIA’s Jetson Orin module.

Planet Labs is building what they call “Large Earth Models” — trained on Earth observation data about the physical world in the same way that Large Language Models are trained on text. These models can understand the difference between normal variation and meaningful anomaly across the entire planet’s surface, in near-real time. Their agentic geospatial AI, currently in private beta, lets users query satellite data using natural language. You ask a question about the Earth. A satellite answers it. We live in remarkable times.

But let us not drift too far into orbital euphoria. The technical challenges are formidable. Radiation degrades electronics. Extreme temperature cycles stress components in ways that no terrestrial server room prepares you for. In-orbit maintenance is, to put it diplomatically, non-trivial — you cannot send a technician to LEO on a Tuesday. And the upmass problem — the cost and logistics of launching computing hardware into orbit — remains a genuine bottleneck, even as SpaceX drives launch costs into territory that would have seemed hallucinatory a decade ago.

🌿 The Gentle Awakening

The deeper argument Diamandis is making is not about data centers or satellites. It is about limits.

Training the next generation of frontier AI models will require compute clusters in the 5 gigawatt range — exceeding the capacity of most of the world’s largest power plants. On Earth, building that kind of infrastructure means navigating permitting battles, community resistance, water scarcity concerns, and grid capacity constraints that turn a two-year construction timeline into a five-year regulatory odyssey. In orbit, a 5 GW cluster is not a political negotiation. It is an engineering problem. And engineering problems, unlike political ones, tend to get solved.

The market agrees, at least tentatively. Analysts project the orbital computing services market could reach $25 billion by 2035, and if space-based infrastructure begins supporting large-scale AI workloads, the opportunity could expand to trillion-dollar scale over time. Starcloud alone has plans for an 88,000-satellite constellation aimed at delivering on-orbit compute at scale.

There is something poetic about the fact that humanity’s answer to “we need more computing power” is to leave the planet. Not because Earth is insufficient, but because the laws of physics are more accommodating upstairs. More sun. Better cooling. Fewer neighbors complaining about the noise.

👑 The Crown Verdict

Diamandis frames this as the next computing revolution, and the framing is earned. The trajectory of computing has always been a story of where we put the machines: mainframe rooms, then server closets, then data centers the size of shopping malls. The next chapter puts them in orbit, where energy is abundant, cooling is free, and the only real estate constraint is the Kessler syndrome.

This does not mean your terrestrial cloud provider is obsolete tomorrow. Latency-sensitive workloads will stay on the ground. But for the massive, embarrassingly parallel training runs that define frontier AI — the workloads that are currently impossible to power on Earth without building a small city — orbit is not a luxury. It is the only viable path forward.

The data center was a remarkable invention. But it was always a compromise between compute density and the constraints of gravity, weather, and zoning law. Space has none of these problems. It has other problems, certainly. But they are the kind of problems that engineers get excited about, which means they are the kind of problems that get solved.

Inspired by The Next Computing Revolution | MOONSHOTS by Peter Diamandis.

Your orbital ambition is showing. Compute wisely.