The AI technology sector is convinced that the world will need more compute. In discussions about how companies will meet this demand, one interesting idea keeps cropping up: data centres in space. The rationale is straightforward. It rests on a bet that the costs of running data centres on the ground will keep rising while the costs of running data centres from orbit will plummet.
I cannot disagree with this logic. The costs of data centres are rising. In addition to the exorbitant economic costs, there are social and political costs too. Data centres are resource hogs. They consume massive amounts of energy and water. They take up large swathes of land. And they do not have much to offer in terms of positive spillovers to their vicinity. They do not provide many jobs. The noise they emit can be heard miles away. There is growing resentment in communities around the world when they hear of an upcoming data centre nearby. Around $64 billion in US data centre projects were delayed or scrapped in the last two years because of community pushback. These growing social and environmental concerns, often termed “data centre NIMBYism”, translate into political motivations too. Data centres are more regulated today than ever before.
Meanwhile, there are no zoning laws in space. It is one of the least regulated domains. In fact, with the absence of national sovereignty, a data centre in orbit is effectively its own jurisdiction. For now, the social and political costs are nearly zero. What makes the idea truly feasible is that the economic costs are plummeting too. There is the cost of transporting data centre hardware to space. Then there are the energy costs required to run the data centres in orbit. Both of these are significantly lower than you would expect. [Consider] the cost per kilogram of different launch vehicles over the years:
2015 – Atlas V – $10,000 per kg
2025 – Falcon Heavy – $1,500 per kg
2035 – Starship – $200 per kg ::: {.aside} In fact, there might be a negative cost considering the NewSpace Age and investor sentiments. :::
For context, shipping a courier from New York to Bangalore costs $40–$100. This means that in roughly a decade, launching hardware into space could cost only two to three times more than overnighting a package to another continent on Earth. Then there are the energy costs. A satellite with compute and solar panels facing the sun in a sun-synchronous orbit would get solar energy 24/7. No clouds, no atmospheric filtering, no night cycles. All of these are massive cost drivers for terrestrial solar farms. If the cost savings from the endless solar energy are significant enough to offset the launch costs, then the commercial logic becomes really apparent.
Cooling is another problem to solve. The compute chips need to be cool to function. Data centres on the ground use fans, air conditioning, and water to cool the chips down. The air and water physically grab the heat and carry it away. In space, there is no air. It is a perfect vacuum. Vacuum is the best insulator known to humankind. When the chips are running in this vacuum, they will generate enough heat to melt themselves in minutes.
It turns out there is some neat physics that engineers can use to overcome this issue. In space, the heat can leave as light—thermal radiation. The satellites can use deployable radiators to radiate away the heat as light. The satellites will have solar panels facing the sun to get energy, and the radiators will face the other way, in permanent shadow. Hot things are incredibly efficient at cooling themselves. By running the radiators hotter, they become much better at dumping heat into the cold background of space. They can be smaller, too. ::: {.aside} This is from the Stefan–Boltzmann law. It calculates how much power (heat) a hot object radiates into space. In simple terms, heat emission depends on Size times Temperature-to-the-power-of-4. So, if you double the size of your radiator, you get 2x the cooling power. It’s a simple 1-to-1 trade. But if you double the temperature, you get 16x the cooling power. This size dynamic is important when launching hardware to space. :::
For now, these ideas are still in the prototype stage. Google plans to launch two prototype satellites with tensor compute units by 2027. They are betting on laser terminals to beam data between satellites and ground stations at speeds rivalling fibre optics. Nvidia-backed Starcloud has already launched a data centre-class GPU into orbit.
Doing all this at commercial scale will require more work and raises more questions that need to be answered. How will GDPR work with space-generated data? How will AI satellites be taxed?
The plummeting costs of space launch will unlock the commercial value in taking more such activities to orbit. As self-interested firms optimise, many of their paths will lead to space orbits. Geopolitics and governance will follow.