Space data centers: SpaceX and Blue Origin race to orbit as scientists question physics

The scale is attractive in its simplicity: AI needs more power than the world’s internet can provide, so it moves data centers into orbit, where the sun never sets and electricity is free. SpaceX, Blue Origin, and a growing constellation of startups are now racing to make that vision a reality. The problem, according to the scientists and engineers who would have to make the physics work, is that the vision violates several chapters of thermodynamics, economics, and orbital mechanics that have not yet been written.

SpaceX filed an application with the Federal Communications Commission on January 30 for permission to launch up to a million satellites in low Earth orbit, each carrying computer hardware that will create what the company described as a “constellation”.unprecedented computing power to power advanced forms of artificial intelligence.” The satellites would operate at ranges between 500 and 2,000 kilometers, in orbits designed to maximize daylight hours, and internet traffic through SpaceX’s existing Starlink network. SpaceX has requested a waiver of the FCC’s standard supply measures, which typically require half of the constellation to be operational within six years.

Seven weeks later, Blue Origin submitted its application. Project Sunrise provides 51,600 satellites in sun-synchronous orbits between 500 and 1,800 kilometers, complemented by the previously announced TeraWave constellation of 5,408 satellites providing ultra-high-speed optical backhaul. Where SpaceX’s filing emphasized raw measurement, Blue Origin emphasized innovation: the system would perform computations in orbit and transmit the results to the ground via the TeraWave network.

The startup ecosystem is moving even faster. Starcloud, formerly Lumen Orbit, raised $170 million at a $1.1 billion valuation in March, it became the fastest unicorn in Y Combinator’s history just 17 months after graduating from the program. The company launched its first satellite equipped with an Nvidia H100 GPU in November 2025 and filed an application with the FCC in March for a constellation of up to 88,000 satellites. Aethero, a security-focused start-up for space-grade computers with Nvidia Orin NX chips wrapped in radiation shielding, has raised $8.4 million and is testing hardware this year.

The business model depends on the real problem. Data center power consumption worldwide it reached about 415 terawatt hours by 2024 and the International Energy Agency projects it could exceed 1,000 TWh by 2026, with fast AI servers driving growth of 30 percent per year. In Virginia alone, data centers use 26 percent of the total electricity supply. Ireland’s share could reach 32 percent by the end of the year. Network bottlenecks are real, enabling delays are real, and political resistance to building more global capacity is real.

What’s also true, scientists argue, is the physics that make orbital computing incredibly difficult by any meaningful measure. The main challenge is the heat. Instead, there is no air to carry heat away from the processors, only radiant cooling, which requires large surface areas. Dissipating just one megawatt of thermal energy while keeping electronics at a constant 20 degrees Celsius requires about 1,200 square meters of radiator, about four tennis courts. A data center of several hundred megawatts, the lower limit of commercial importance, would require radiators thousands of times larger than anything previously used on the International Space Station.

Radiation presents a second structural problem. Low Earth orbit exposes vulnerable chips to cosmic rays and trapped particles that cause flips and permanent orbital damage. Radiation hardening adds 30 to 50 percent to hardware costs and reduces performance by 20 to 30 percent. Another method, triple reduction, means starting three copies of each chip, three times the cooling, three times the power, and three times the weight. Starcloud’s approach to flying commercial GPUs with external protection is an interesting experiment, but no one has shown it to be effective at mass or beyond the lifetime of hardware measured in years rather than months.

Latency is the third obstacle. A million satellites distributed in orbital shells from 500 to 2,000 kilometers cannot achieve the strict connection required for the training of the boundary model, where the inter-node communication parameters must stay in the microsecond range. Low Earth orbit produces a few milliseconds for inter-satellite communications and 60 to 190 milliseconds for round trips, compared to 10 to 50 milliseconds for terrestrial communications networks. That makes orbital architecture possible for heavy workloads, not for training, which is where most of the demand for AI computing currently lies.

Then there is the cost. IEEE Spectrum has estimated that a gigawatt orbital data center would cost upwards of $50 billion, nearly three times the cost of an equivalent space on Earth over five years of operation. Google has said that startup costs must drop below $200 per kilogram before space-based computing starts to make economic sense. SpaceX’s current Starlink economy works out to about $1,000 to $2,000 per kilogram. Some analysts argue that the real competitive price in the world economy is $20 to $30 per kilogram, a figure that has no reliable estimates for the next two decades. The economy looks even better when set against Deep financial technology on the ground, where global infrastructure projects can use established supply chains and proven economies of scale.

Even Sam Altman of OpenAI, who reviewed the multibillion-dollar investment in rocket manufacturer Stoke Space as a competitor to SpaceX for orbital data centers, publicly called the idea “crazy” for the past decade. Altman told reporters that the initial negative cost calculations related to the cost of global energy have not worked, and asked clearly if anyone plans to fix the broken GPU in space.

The star community adds an entirely different resistance. Most of the nearly 1,000 public comments on SpaceX’s FCC filing urged the commission not to proceed. If accepted, the constellation would place more satellites than the visible stars in the sky during the longest night of the year, increasing the military and doing business in the orbital environment already straining under the weight of existing galaxies.

None of this means that orbital data centers will never exist. SpaceX’s Starship, if it meets its cost goals, could change the economics of orbiting large-scale orbiters that make the concept unviable. Starcloud’s incremental approach to flying small payloads and repeating beam operations is the kind of engineering approach that sometimes produces success. And the internet’s barriers to driving interest are endless.

But the gap between filing an FCC application for a million satellites and making orbital computation economically competitive with a warehouse full of GPUs in Iowa is not measured in years. It is measured in physics problems that the current pace of AI infrastructure investment it won’t cut it, no matter how many billionaires are willing to try. The question scientists are asking is not whether space data centers are feasible. That’s why, given the vastness of engineering that has not been solved, anyone treats them as a near-term solution to a problem that requires near-term answers. The sky, it seems, is not the limit. There is a radiator.