But in a surprisingly quiet but important front in the longstanding feud between Jeff Bezos and Elon Musk, the latest battlefield in the struggle between the two billionaires may now shift not from rocket technology but from the design of artificial intelligence systems itself this time, not on planet Earth but in orbit. Both Blue Origin and SpaceX are reportedly pressing ahead with plans to use artificial intelligence to create data centers in space. This is in a bid to take advantage of the benefits that space offers to help supplement the increasing need for computation on planet Earth.
The appeal is simple yet profound. The earth-based data centers already account for huge power consumption, about 4.4% of the total power consumption in the US as of 2023, and is predicted by the US Energy Department to double or rather triple to between 6.7% and 12% by the year 2028. The AI computing workloads, owing to the high GPU and TPU computing density, put additional pressure. In space, solar panels are able to generate power from almost constant sunlight, unlike earth-based solar power systems. The sun-synchronous orbit provides anywhere from two to eight times more power per solar panel as compared to earth-based systems, making the need for battery or fossil fuel-based power backup negligible.
Space also provides an inherent heat sink. Rather than relying on cooling towers that annually devour huge volumes of water as evaporative cooling towers do space platforms can exhaust their waste heat directly into the vacuum of space at a temperature of 3K by means of radiators. This is not an insignificant factor; according to NASA analysis, radiators can take up more than 40% of the power-related spacecraft mass during high loads; nevertheless, this eliminates what has been the most resource-intensive factor for Earth stations. Companies such as Starcloud are already taking advantage of this for their Starcloud-1 satellite, which weighs only 60 kg and comes with an NVIDIA H100 GPU capable of delivering performance 100 times greater than any previous space-based computation.
Radiation protection is another area for engineers to explore. Google’s Project Suncatcher exposed its Trillium-generation TPUs to a 67 MeV proton beam to model five years in Low Earth Orbit. The chips withstood the ordeal unscathed, although High Bandwidth Memory subsystems were found to be vulnerable to Total Ionizing Dose radiations. Radiation shielding techniques should therefore take into account mass considerations and reliability, since extended AI training sessions might experience Single Event Effects.
Communication between satellites is also necessary. To provide low latency in ground-based data centers, terrestrial facilities interconnect with multi-terabit-speed networks; DWDM optical interconnects are needed for the similar requirements in space. Google identified that 10-Tbps total bi-directional bandwidth can be supported for commercial transceivers, but satellites will have to orbit in close formation to mitigate beam divergence and intensity; this requires station-keeping modeled through Hill-Clohessy-Wiltshire equations and further correction for effects of the earth’s odd gravitational distribution.
The key to low-cost LEO access is to focus on the last few kilometers of the orbital insertion maneuver, according to Jeff Bezos. The primary issue is non-launch economics. Today the cost is between 1,500 and 2,900 USD per kilogram. This is too high to have a large constellation of satellites. There is a chance that by the mid-2030s, SpaceX’s Starship can reach 200 per kilogram, and when that happens, the cost of access to space could be competed away by terrestrial energy providers. Bezos has asserted, “there will be gigawatt data centers in space in 10+ years,”
These are just some examples in terms of compute. But on-orbit AI might allow real-time analysis of earth observation imagery, enabling analysis in minutes instead of hours or days, such as real-time fire detection, crop analysis via hyperspectral imaging, or disaster response. It may provide an isolated environment within which to run so-called high-security AI models. Finally, on-orbit compute nodes located in Lagrange points may preprocess data from telescopes before they are downloaded, optimizing science return in the presence of bandwidth limitations.
Bezos and Musk are not alone in this race. Google, Eric Schmidt’s Relativity Space, and an increasing number of startups are zeroing in on the same physics-based opportunity: abundant solar power, effective cooling, and planetary coverage independent of earth’s limitations. While federally and privately supported rocketry demonstrations are impressive, and their underlying technology is certainly promising, so too are their formidable engineering hurdles. Yet their potential, too, is staggering. What began as a rocketry contest is now shaping up into a contest for what will constitute the computational infrastructure of tomorrow and one that orbits well beyond today’s power grids and data centers on earth.
