In 2026, AI policy is datacenter politics. When the public is dissatisfied about AI, they protest and disrupt the datacenter buildout. When policymakers try to tax or regulate, they don’t need to worry too much about company flight—a search engine can run to Ireland, but the datacenter stays here. When middle powers develop AI strategies, they start with compute: whether it’s hosting American datacenters in exchange for access or building sovereign datacenters for admittedly fuzzier reasons. And when we all have these late-night conversations on how we’d deal with runaway AGI, the answer is also about datacenters: send in the troops, pour some water on the Blackwells, kill the power stations.
But in 2031, Earth might no longer be the premier place to build a datacenter.
Instead, much of the 2030s’ compute buildout could happen in outer space. Datacenter developers are scouring the earth for a place with great unit economics, fast, reliable time to power and cheap electricity. They’re running out of sites that clear the bar. But in space, the sun always shines and there are no anti-datacenter protests. Yes—unfortunately, you need a rocket to go there. But it turns out they’re getting cheaper very fast. And so, with an eye on the launch cost trendlines, firms have started to develop satellites that can carry frontier chips and run them in gigawatt-scale orbital constellations.
If a few remaining technical hurdles are cleared and rocket launch costs keep falling the way the industry expects, our AI models are going to space by the start of the next decade. Once space compute becomes viable, it quickly becomes compelling: datacenters concentrate heavily in the best place to build them, and that might soon be outer space.
Very little has been said about the policy implications of this prospect. Algorithms and AI agents are already hard to grasp through our usual statecraft: we can’t touch, tax, threaten or tariff them. The technology firms that run them are powerful, mobile and elusive. Datacenters, earthbound physical infrastructure, are a rare anchor keeping machine intelligence within reach of public policy and popular protests today. So as the industry is beginning to think seriously about putting compute in space, we should start thinking about what would happen to AI policy after datacenters.
Liftoff
For a front row seat to the literal take-off the industry predicts, you have to travel to the southernmost point of Texas, near a town that used to be called Boca Chica. Pick a cloudless night, camp down on one of the barren fields just near the U.S.-Mexican border, and look to the skies. Right now, most nights would be empty save for an exceptional view of the stars and constellations. But in 2031, on a good night, you might see near-constant traffic: rockets headed for the skies, each deploying dozens of satellites into orbit. With a good telescope, you’d be able to observe these satellites align into their own kind of constellations: each of them carries a rack of AI chips, and they link up into a computing hub that rivals the scale of today’s biggest terrestrial datacenters.
‘Space datacenters’ of this kind are still a nascent industry. But they’re much further along than outside observers usually credit. For space compute to work, you need to do two things: build space datacenters and put them into space. Building the individual satellites is surprisingly easy: a space datacenter is just a satellite full of AI chips that can communicate with other satellites full of AI chips. Already today, we have thousands of Starlink satellites in orbit that can communicate with other satellites. All we need is to make them bigger and put AI chips into them.
Many serious people are pursuing efforts to that effect: Google has a landmark program called Project Suncatcher, SpaceX is devoting serious R&D to its ‘Starmind’ space compute line, Nvidia is developing chips and racks and much of the foundational technology, ex-Google-CEO Eric Schmidt has bought a company to put compute into space. A batch of young hardware startups has embarked on similar pursuits, with Starcloud running an H100 in orbit and raising $250 million on plans for 88,000 satellites and 20 gigawatts of orbital compute. It’s a bit out there, but it’s not just Elon Musk.
They are motivated by a compelling pitch: take reusable, cheap rockets of the Starship design and use them to transport satellites into space, each equipped with a rack of dozens of top tier chips, a few dozen per launch. They assemble into networked constellations in a dawn-dusk sun-synchronous orbit at 500 to 600 kilometres, which means they get sunlight all day, every day, for four times the energy yield of the same panels on the ground. Every 250 launches, you get the equivalent of another gigawatt-scale datacenter online. The satellites are mass-manufactured in efficient factories, no local construction, no in-space assembly, just production and launch. The fastest time to reliable and abundant power you could possibly get.
I’d be doing you a disservice if I said this was definitely going to happen. The sticky parts, roughly in order, are that launch costs aren’t actually that low just yet; cooling these datacenters still seems very difficult, networking them at high-enough bandwidth is difficult, and maintaining them is near-impossible. Yet the industry deems these problems solvable: maintenance through redundancy, cooling through high-temperature radiators, and communications through lasers and the Starlink precedent. Launch costs are plummeting by the familiar magic of straight lines on a log graph: rockets are getting cheaper, and once they’re reusable at scale, costs could fall far below the break-even point for terrestrial versus orbital compute.1
No one’s backyard
There are clever answers to these claims as well, and just-as-clever responses to these answers. For the purposes of this piece, I’ll assume the tech and economics will work out—enough to make space-based inference clusters extremely attractive and to make space-based training fringe but doable in the 2030s. That’s both because I think it’s likely enough that this happens for us to begin considering the what-if, and because I’m compelled by how strong the incentives seem.
The first incentive is from datacenter economics: industry trends point toward pricing reliably fast time-to-power over all else, even at the price of more wasteful processes or more difficult cooling. Space datacenters, done right, are fast in a predictable way. Terrestrial datacenters are locally concentrated, project-based, lump-risk investments: they can go wrong or right, move fast or slow, and you never quite know until the GPUs start running. Space datacenters, in comparison, are easy to predict: they’re assembled at scale in centralised factories and deployed on a steady schedule of launches. That’s a much more attractive modus operandi for a terawatt-scale financialised industry than scattered megaprojects.
The other driving force is datacenter politics. Even in 2026, building compute anywhere near people is out: the anti-datacenter mood is growing, chasing the buildout across state borders, and there’s no equitable resolution in sight. Local communities in particular have proven resistant to bribery and pressure and increasingly are a serious risk to datacenter builders. It’s entirely fair for these communities to leverage their attractive sites, of course. But as the buildout continues, communities like these are becoming more scarce and more extractive. If there aren’t enough of them to produce a race to the bottom, eventually they’ll start asking for concessions that hurt.
Leaving the datacenters on earth also saddles hyperscalers and AI developers with an unfortunate vulnerability to policy action: as long as a municipality, a state, even a nation controls your datacenter’s access to power, it can pass AI policy and force you to comply if you don’t want to lose your chips. Getting away from most of that sounds very attractive to an industry with an ever-deteriorating approval rating. But there’s no place on earth without politics: move to allied countries and the datacenter politics might follow you; try to move to autocracies, and you’re vulnerable to export controls and regional unrest.
If there’s any way at all to escape this planet and its local politics, a trillion-dollar industry faced with this much pressure to leave will find it.
The domestic policy of going to space
For a first glance at the policy implications, we need to look away from the constellations in the sky for a moment and instead look toward the compound from which the rockets keep launching. The village of Boca Chica is of course no longer known by that name—it is now called Starbase, Texas instead. Its mayor is a SpaceX executive, and it’s owned by the firm and its ruler, Elon Musk. Starbase is an extraordinary facility: an incorporated company town of 1.5 square miles, home to some 500 residents and over 4,000 SpaceX workers, set to double within this year alone. It’s within reach for democratic governance in an abstract sense, but even today, it’s shaping up to be as sovereign a facility as you can build within the borders of the U.S. Who knows how it will look by 2035.
By then, facilities like Starbase, along with a few transmission-receiving facilities, might be all that’s left of the AI industry’s physical nexus on earth.2 That won’t be quite true as soon as space compute becomes viable. The datacenters we’re currently building on Earth will stay relevant for another few years. But eventually, chips degrade and datacenters deteriorate. If building in space is cheapest, you should expect constellations to take the place of more and more datacenters year by year.
By that time, the other relevant physical anchor—scarce and capable employees—will also have weakened. The lion’s share of AI research and development will instead be carried out by AI agents, themselves running on space datacenters in short order. The transmission sites that receive the satellites’ communications don’t strike me as lasting leverage in extreme scenarios: losing them is really bad for their owners and reduces downlink traffic for a while. But having to relocate the earth-based receiver is neither all that catastrophic for the agents hosted in the space datacenters nor for anyone else who has a backup site.
When the chips go to space, two interesting things happen to domestic AI policy. Government loses much of its ability to affect already-deployed compute and the AI systems running on it. It’s relegated to governing firms over infrastructure, and future launches over present deployments. And at the same time, political control over AI infrastructure moves from being distributed across countries, states, and municipalities toward being concentrated with the federal regulators of spaceflight and a handful of launch sites.
There’s no National Guard in space
Losing the option to physically go into a datacenter makes AI policy harder. Satellites are not out of the legal reach of American authority. Launched from U.S. sites, pursuant to the Outer Space Treaty, space datacenters would still be subject to rulings, legislation, and executive action. And as long as the labs running them remain American firms, the legal entities, too, remain within American reach. Someone is still in charge.
But legal options don’t provide practical intervention capacity. Take some of the more extreme scenarios: if a frontier lab believes itself on the final stretch to superintelligence goes rogue and stops responding to government intervention, an unprepared U.S. government couldn’t do much to affect the chips already in space. If the machine entity that inhabits the chips themselves goes rogue and stops responding to earthly communications, intervention is similarly limited.
Even in today’s Congress, policymakers are introducing an ‘AI Kill Switch Act’ that is supposed to ensure unilateral ability to stop this kind of AI deployment gone wrong. But a post-hoc legislative act won’t do it if you can’t reach deployed agents: no national guard to go in and cut the copper wires, no electricity company to turn off the substation. The spectrum of available action collapses to the two unsatisfying extremes of ‘try to shoot down your own industry’s satellites’ and ‘hope the legal entity still cares about injunctions’.
To my institutionalist mind, the consequence of that is fairly clear: let’s not put ourselves into that position. There’s just no way you can let these guys put the chips in space without an ironclad remote oversight and kill switch system, anchored in reasonable governance and placed in the hands of controlling governments. If you want to be maximally paranoid, you’d even want this to be a dead man’s switch so that autonomous constellations can’t cut off communications and recursively self-improve on their own.
This is something you have to do before you launch enough inference capacity into space to pose a serious danger, meaning we should really develop and perfect the hardware-enabled mechanisms everyone’s talking about for purposes of executing US-China treaties anyways. That also requires figuring out the precise governance ahead of time: fine-tuned remote oversight is also tremendously power-concentrating, and you wouldn’t want to put it together ad hoc and place it into the hands of an unconstrained executive.
To retain meaningful control over autonomous AI systems in space, we would need remote oversight. If we don’t, we might under-react to a dangerous development—say, a self-exfiltrating AI or recursive self-improvement loop running on a completely autonomous constellation of satellites that cannot be reached by plane, ship, or tank. Or we might overreact to a mildly concerning instance of rogue AIs that can only be shut down by blowing up our hard-won space-based infrastructure. This is a solvable problem, but one that requires policymakers to not dismiss the constellations going up before they reach concerning scale.
AI policy as launchpad governance
Today, political control over AI is widely distributed. California can regulate AI because it’s built and developed there. New York and the European Union can regulate AI because they control important markets that they can deny to non-compliant developers. And any county or state with a datacenter has some leverage over the decisions of its owner, too. So much leverage, in fact, that datacenters have become a salient anchor for the AI politics conversation in recent weeks: they’re construction projects with all the vulnerabilities to local opposition and state pushback that construction projects ordinarily have. They yield local taxes, and they can be disrupted by local protests.
AI policy is losing real-world political anchors one by one. Consumer markets become less and less important as labs rely on fewer and fewer customers to pay for increasingly scarce frontier tokens; and the talent nexus will matter less as more researchers are automated away. The datacenters will be the main policy anchor: whoever controls a lab’s datacenter has hard leverage without a feasible opt-out for the lab. With terrestrial datacenters, that control is spread throughout jurisdictions. Not perfectly, but good enough that a range of governments and agencies can try their hand at different regulatory approaches: we can see how AI policy in California and in New York goes, what local AGs or federal agencies think. And when scepticism of big tech and political money boils over, it can be channelled through local resistance against datacenter projects.
If the only measure available to anti-datacenter protests were to convince the U.S. Senate to pass an anti-space-datacenter law, I think the industry would be much less concerned: much more than local officials, national bodies are in a position to ignore public unrest in favour of national security and broader economic arguments. If I was an anti-AI activist, I’d be very worried about any plan that removes infrastructure from the immediate reach of grassroots movements.
Once datacenters go to space, the nexus of regulatory leverage moves, too. The most important nodes become the launch sites needed to put new satellites into space as old ones degrade and the buildout continues; and perhaps the transmission sites that reintegrate space inference into the terrestrial economy. The practical authorities that would govern this version of AI are highly centralised. They’d be concentrated in the incidental agencies overseeing (a) commercial spaceflight (like the FAA) or the communications involved (like the FCC) and (b) the regional authorities of way fewer sites than in the datacenter paradigm. In 2031, the FCC chairman and the mayor of Starbase might well make the TIME 100 list of most influential people in AI.
No matter the governance regime then, these authorities matter. As we’ve seen in both the Fable episode and the Anthropic-Pentagon conflict, considerable power still rests with anyone who has executive authority over what AI labs are doing. Most recently, that was the Pentagon and the Supply Chain Risk designation, or Commerce and export control authority; in the future, the same kind of power would by default be afforded to whoever incidentally got the Starbase or the satellite portfolio when no one thought it would matter.
The thing about space—as well as about most of Elon Musk’s ventures—is that things tend to move pretty fast once they start working at scale. That frequently catches people off-guard: you keep thinking this is surely not going to work, up until it really works and it’s too late to make the institutional changes necessary. In other words: if you don’t think the FCC chairman should be the AI czar, if you think Texas shouldn’t call the shots, or if you think the notion of space datacenters is broadly objectionable, then you should be thinking about the institutional arrangement of space compute before the rockets launch.
Outer space for middle powers
Most keen-eyed readers will have noticed that Starbase is located in the United States of America. That means that, for the rest of the world, concentration of government authority onto launch sites is a massive problem.
I’ve long advocated for a ‘compute-for-access’ thesis on here: middle powers should build datacenters for compute-hungry Americans and get them to guarantee enduring access in return. The short-term benefit of that strategy is unchanged by space compute; for the next few years, datacenter deals are the best leverage play we have. But the long-term prospect of compute as national policy is seriously imperilled if we go to space.
If you build a datacenter today, you get two things: a datacenter and a track record. The former is good for immediate access and leverage. But the latter is what carries the endgame of compute-centric strategies: once you have been successful at delivering on one or two ambitious compute projects, you become a prime location for future deals. You can argue that you should continue receiving chips because you’re the best place to get them online fast. To risk-averse, compute-hungry labs in a few years, that pitch will be very hard to ignore, and so you should expect to be able to turn short-term compute success into long-term leverage: an AI petrostate is not just a place that has datacenters, but a place that can make the best pitch for future datacenter projects.
But that entire premise falls apart if you think that space compute will eventually happen: at some point, your track record doesn’t only compete against everyone else’s track record, but against the prospect of much more convenient orbital buildouts—and then, your petrostate play is over. That means space compute turns datacenter building from an endgame in itself to a midgame play that needs to bootstrap into something else.
Most immediately, that means middle powers need to hurry up on the compute-for-access; time is running out. On the national level, that simply means accelerating dealmaking. Globally, that means we might not have time for an iterated strategy where we first make a deal work in one country and then export the blueprint elsewhere; instead, we need many countries to take the plunge at the same time. If Canada waits to see if the play works out in Australia in two years, it might already be too late. Even if there are earthbound datacenters in the space compute era, the compute-for-access play loses its bite: the space option removes the hard constraint of limited terrestrial sites, which means you can only negotiate so hard before the lab decides to go for satellites instead.
Space compute would also mean it’s time to think about what comes after datacenters even sooner. In brief: the way to think about what matters when compute goes to space is to think about what the trucks going into Starbase will carry. It’ll still be chips and the equipment needed to manufacture them. It’ll also, at much greater scale, be solar panels and the equipment needed to manufacture them—a currently highly concentrated supply chain more or less monopolised by China. Look at how hectic even today’s scramble for gas turbines to power the terrestrial buildout is; now imagine a world where solar panels are the key bottleneck to otherwise near-abundant compute deployment capacity.
And of course, launch capacity itself will be a highly coveted asset. A domestic space program able to deploy satellites at scale would be worth the most. But even launch sites without a space program would matter: there are only so many of them, they take ages to build, even longer to permit. There also aren’t that many democratic nations in a good place to launch to sun-synchronous orbit specifically—you’d ideally want a clear north-south corridor over the ocean, so Norway, Sweden, Scotland, New Zealand, Canada and Australia come to mind.
All the old arguments about distributing datacenters through the world will apply to spaceports just as well: spacefaring companies might want to go jurisdiction-shopping for favourable terms, might want to have some way to escape the U.S. government in particular, and there’s some geopolitical appeal to reducing the concentration of oversight. I suspect that the few democratic countries with space and geography to build a launch site would be well-served by doing so.
In careful defence of weaponising space
Bear with me on this final stretch; it’s out there, but I can’t help myself. If enough compute to run powerful AI ends up in space, that puts geopolitical stability at risk. If space datacenters exist, but there are no ways to destroy them, there is no deterrence that keeps anyone from doing very dangerous things with their datacenters. In the limit, starting a reckless recursive self-improvement loop could be equivalent to launching an ambiguous intercontinental missile: enough to invite a second strike from another nation state, because the risk is too high that you’re doing something that could upset the gameboard decisively. Reckless RSI—if it is a discrete event, which seems unclear at best—is therefore something nation states might be deterred from.3 But only if second-strike capability exists.
Fortunately, space, by default, is offence-dominant: you can’t really put space datacenters into an impenetrable bunker or hide them anywhere, so second-strike capability would seem likely to remain intact.
At least, so long as geopolitical rivals are equally able to hit space-based assets at an acceptable cost. China might reach technical and commercial parity before long, but that doesn’t help the rest of the world; and even between superpowers, the balance could be quite volatile. First, because you could imagine one country pulling ahead in launch capacity quickly, outproducing their rivals’ anti-satellite weapon capacity quickly and consolidating its lead through monopolising space businesses at least, and through monopolising space warfare at worst. And second, because dislodging and then replacing a space monopolist once entrenched is extremely difficult: blowing up a bunch of enemy satellites leaves a great deal of debris in orbit that in turn makes it harder to launch your own vehicles. Adversaries might have to choose between accepting the monopolist’s terms, or being barred from going to space themselves by the self-reinforcing debris clouds called Kessler syndrome.
This sort of entrenchment is already concerning if Starlink is the most impressive space platform. It’s a serious threat to other nations’ sovereignty if you have enough compute in space to make artificial superintelligence.
If we were sure all this was going to happen, I think it might be prudent to start weaponising space. Not in the sense of proliferating earth-focused weapons in space, but the opposite: proliferating space-focused weapons on earth. When I draw the trendlines out, I just don’t really see another way. Too much of geopolitical stability depends on nations getting to stop each other from doing whatever they want. If only a few nations get to go to space, this mutual system of checks and balances ceases to work.
So if you’re really sold on space compute and you want to be able to credibly deter destabilising deployments of AI systems, I’m not sure there’s a way around building ASAT—anti-satellite weaponry—at scale. That weaponry is imprecise, escalatory, and sometimes suicidal. But mutually-assured Kessler syndrome still makes for very effective deterrence, especially since it hits the incumbent hardest.
It’ so effective, in fact, that countries deploying ASAT at scale might ruin the prospect of space compute altogether. It makes insuring satellites much harder, and puts the outer space commons at considerable risk of becoming inaccessible in the long run. Still, middle powers might mostly see the trend toward space compute as a way to make them obsolete; so why comply in keeping the insurance rates low? As with most sovereignty-brained gambits, the Gaullists in France are once again already on it.
Who will be the Taiwan of space?
To be an effective analyst of AI policy, you have to be bitter-lesson-pilled: believe that ultimately, this entire thing comes down to compute. Thinking about what happens to the chips gets you close to figuring out what happens to the world. If that’s true, then this upcoming shift in how chips are deployed would spell a paradigm shift: it should change how we think about long-run datacenter politics, frontier AI economics, and the incentives around exports, imports and controls. The next few years will make it seem like all AI policy is about earthbound datacenters—but if the space compute thesis is right, that view will ultimately be mistaken.
We’re still extremely early to that conversation, so the opportunity for trajectory changes is still massive. Today’s policy terrain was locked in by buildout decisions and semiconductor economics decades ago: what a strange artefact of history that Taiwan, South Korea and the Netherlands are that important, or even that California de facto gets to make federal AI policy. If the space compute era comes, its material preconditions will be locked in in the next few years: when the chips that are supposed to go into the satellites are designed, and when the sites that launch the rockets are dug.
Anywhere in the world, I talk to people asking me for some arcane bit of AI alpha that can help them find a winning national AI strategy. I feel like they’re trying to be early to a conversation that’s been going on for too long: the labs are scouring the world for leapfrogs, the markets are pricing in even the most niche bottleneck assets. But these days, in AI, it’s already all priced in.
So here’s your space compute essay—how’s that for alpha? Yes, it sounds strange. Arguably, it might never work. More likely than not, you’ll look silly in a decade if you go all-in on my advice now. Definitely, you couldn’t sell acting on any of this to a parliament. But if it did happen, you’d feel just a bit stupid for not having taken it seriously earlier. Betting on the next phase of the scaling hypothesis always feels a little like that.
A thorough report by Forethought and a good piece by Semianalysis go into much more detail.
The compute supply chain, of course, persists and remains important as ever; compute governance isn’t going anywhere, even if datacenter policy is.
AI 2040 and the MAIM doctrine rest on a similar premise.