The Supply Chain for a Dyson Swarm

Data Centers in Space: Why, How, and Where to Scale

Ross Centers & Brogan BamBrogan | Ethos Space | March 2026

Who will control the supply chain for compute at civilizational scale? The answer starts with a question that sounds simple, yet isn’t: where should a data center be?

The obvious answer is: wherever power is cheap, land is available, and regulatory friction is manageable. That answer has driven the last decade of hyperscaler expansion. It is the right answer for today’s scale of compute.

It is the wrong answer for tomorrow’s.

The demand for compute is not like the demand for food or automobiles or even electricity to produce those things. Human needs are finite and those markets saturate. Compute, in a world of strong AI and robotics, does not saturate. Compute generates its own demand: more intelligence produces more problems that require more intelligence to solve, more automation produces more systems that require more compute to run. The ceiling keeps moving.

This has a physical implication that the industry has not yet fully reckoned with. At the scale compute is heading toward: terawatts and eventually more, Earth is not the right place for it. Earth is a fishbowl at the bottom of a gravity well, with finite land, finite cooling, finite power, and a biosphere we cannot afford to sacrifice. The logic of compute growth, followed far enough, points off planet.

The natural conclusion to this is a Dyson swarm - a collection of solar panels orbiting the Sun collecting a substantial fraction of the energy it gives off. This will be built with materials that are abundantly available in space and trivial to move once you are no longer fighting Earth’s gravity well.

This is not science fiction. The physics are straightforward. The economics are compelling. The first steps of the supply chain are already being built.

What is not yet settled is who builds it, and on what terms.

Part I: Data Centers in Space Are a Good Idea, But Not for the Reason You Think

The case for data centers in space is usually made on cost. The argument goes: launch costs are falling, solar power in space is free and uninterrupted, you don’t need land or water rights, and you sidestep the regulatory friction that is increasingly throttling data center construction on Earth. Build it once, float it forever.

This argument is not wrong. But it is weak, and it is not the real argument.

Cost is the wrong frame because technology is inherently deflationary. Launch will get cheaper over time, but so will terrestrial solar & storage, nuclear power, stranded natural gas, and modular fission reactors that can be sited anywhere. Arguing that space compute wins on cost means betting on a specific rate of deflation in launch against a specific rate of deflation in terrestrial power and land, a bet that is genuinely uncertain and probably not necessary to make. 

The real argument is about scale. Not whether space compute is cheaper than terrestrial compute at the margin, but whether the machine economy of AI systems, robotics, and automated supply chains can grow to its natural size on Earth at all.

It cannot.

Consider what compute demand actually looks like when you remove the assumption that it behaves like other goods. Human demand saturates. A person can eat only so much food, occupy only so much housing, consume only so much energy. The wealthiest American and the wealthiest Norwegian have converged on roughly the same lifestyle, using about 10 kilowatts of continuous energy consumption embodied in a comfortable home, reliable transportation, and abundant food. There is a demand ceiling in the human economy, and we are approaching it across the developed world.

Compute has no such ceiling. Every increase in AI capability opens new applications that consume more compute. Every automation that frees human labor creates systems that require compute to manage. The outputs of AI systems become inputs to other AI systems. Compute is the one good in the machine economy that generates its own demand without satiation. The only tradeoff is between consuming compute now and building more infrastructure for compute later. This is not a speculative claim. It is already observable: software is writing itself, data centers are the fastest-growing category of industrial construction, and AI’s share of global electricity consumption is rising faster than efficiency gains can offset it.

Follow this trajectory far enough and you hit a hard physical ceiling that has nothing to do with the cost of solar panels or the price of rocket fuel. Earth’s primary energy limit where waste heat alone, independent of greenhouse gases, begins to meaningfully warm the planet, is roughly 1,000 terawatts. That is fifty times today’s production of 20 terawatts. The machine economy, growing at the rate AI investment is growing today, reaches that ceiling within the lifetime of infrastructure being built now.

At that scale, the question is not where to find cheap power. The question is where to find enough power, and where to put the waste heat. Space has a clean answer to both, and it’s important to develop that before we pave Earth in data centers.

A data center in space is a thermodynamic engine of almost platonic simplicity: a plane of photovoltaics facing the Sun, generating electricity that feeds a compute layer behind them, with waste heat radiated from the opposite face into the cold blackness of space. The Sun provides two billion times more energy than Earth receives from it. The universe provides infinite cold. In between, compute happens.

This architecture scales without limit for those who can master the supply chain. Which brings us to the constraint that matters.

There are two ways to build compute in space. The first is to manufacture it on Earth and launch it into orbit: the entire supply chain stays terrestrial, and finished data centers ride rockets to their destination. This approach scales linearly with launch capacity, and launch capacity has a ceiling of its own. The marginal gigawatt of space compute, packaged for launch, masses roughly 10,000 tons, about twice the entire Starlink constellation to date. At that mass-per-gigawatt, saturating even an aggressively scaled launch manifest gets you to perhaps 100 gigawatts per year of new space compute. That is not nothing. But it is not the machine economy’s natural size.

The second approach does not launch data centers. It launches the supply chain that builds them, and puts that supply chain somewhere with its own raw materials and a much easier path to orbit. The scaling in this case is not linear. It is exponential.

That somewhere is the Moon. And understanding why requires understanding what the launch bottleneck actually is and how to feed it correctly.

Part II: The Launch Bottleneck gets Broken by the Moon

The launch bottleneck is not simply that rockets are expensive. It is that Earth is an extraordinarily difficult place to leave.

To get one ton of payload into low Earth orbit, you need to burn roughly 50 tons of propellant. This is not a failure of engineering, it is a consequence of Earth’s mass. Our planet is large, its gravity well is deep, and its atmosphere is thick enough that you have to fight it the entire way up. A fully reusable rocket, perfected to the theoretical limit of the technology, changes the economics of launch dramatically. It does not change this ratio. The propellant still burns. The atmosphere still pushes back. Earth’s gravity still demands its tribute.

This means that everything launched into space carries an enormous embedded cost in mass: not just the mass of the payload, but the mass of the propellant needed to lift the payload, and the mass of the structure needed to contain the propellant, and the mass of the engines needed to burn it. Payload is what’s left over. On the best rockets flying today, payload is only around two percent of total vehicle mass at liftoff - 2%! This is why mass in the space economy is priced differently than anything else. On Earth, we abstract over material costs with dollars. In space, you cannot abstract over mass.

There are environmental costs to launch from Earth. Burning rocket fuel at a scale necessary to move the needle on global compute is inherently impactful. Less appreciated but more significant is the NOx production from hypersonic reentry of reusable rockets. You can’t get around either of these impacts.

Now consider the scale of compute the machine economy wants to build. The marginal gigawatt of space compute, packaged as satellites launched from Earth, masses on the order of 10,000 tons by the most optimistic engineering estimates. Even for optimistic scaling limits, you could only launch on the order of a hundred gigawatts per year of compute, if we collectively launched nothing else. The machine economy will want terawatts. Then it will want more.

This is the launch bottleneck. It is not a problem that cheaper rockets solve. It is a problem of physics.

The theory of constraints, developed in manufacturing, offers a precise way to think about this. In any system with a limiting constraint, the correct optimization is not to push more product through the constraint; it is to subordinate everything else to feeding the constraint as efficiently as possible. Pushing finished data centers through the launch bottleneck is the wrong optimization. The correct optimization is to push the supply chain itself through the bottleneck, and then let it operate where mass is cheap.

Mass is cheap on the Moon.

The Moon’s gravity is one-sixth of Earth’s. This is not a marginal difference. The energy required to escape a gravity well scales exponentially with the escape velocity, which means the Moon’s gravity well is dramatically shallower in the energy economics that matter for launch. Getting mass off Earth is almost impossible. Getting mass off the Moon is trivial.

This means that once you have a supply chain operating on the Moon, you can move as much mass as you can produce. The constraint shifts from launch capacity to manufacturing capacity, which is a constraint you can compound your way out of.

Good news - the Moon is made of what the compute supply chain needs. Lunar regolith, the loose rock and dust that covers the entire Moon, contains silicon, aluminum, other metals, and oxygen in abundance. Silicon is the basis of photovoltaics and semiconductors. Aluminum is a structural metal, electrical conductor, and the material for radiating waste heat. Oxygen, locked into the regolith as metal oxides, is the primary component of rocket propellant.

The sequence of development on the Moon follows industrial logic of staged development. You do not need a fully automated self-replicating factory on day one. You build it up in stages, each enabling the next. Instead of just sending mass from Earth to space, the amount of useful mass made available in space per unit of mass delivered to the Moon is the critical parameter. It is the scaling exponent for the machine economy through the launch bottleneck: mass return on investment (mROI).

The first step is infrastructure. Melt lunar regolith and recrystallize it into landing pads, roads, and foundations. A landing pad sounds modest, but it is the prerequisite for everything else: without one it is impossible to return to the same site multiple times and allow infrastructure to provide value. One hundred tons of supply chain equipment delivered to the Moon, used to build a landing pad, returns a 1000x mass return on investment. This is not theoretical. Ethos has developed this building material and tested it under rocket exhaust.

The second step is propellant. The same molten regolith processing that produces landing pad material can be upgraded to perform crude electrolysis, splitting metal oxides to liberate oxygen. Liquid oxygen is 80 percent of the mass of a bipropellant rocket’s propellant load. A lander that can refuel on the Moon does not need to carry return propellant from Earth. The mass freed by not carrying that propellant can instead be payload. This compounds the effective capacity of every subsequent launch dramatically — turning a modest lunar oxygen production facility into a multiplier on the entire cislunar logistics chain. Mass return on investment here can also reach the 1000x mROI range. The remaining 20% of propellant mass is hydrogen or methane, available from icy volatiles in permanently shadowed craters.

The third step is power and materials. Purifying silicon from lunar regolith and processing aluminum yields the basis for photovoltaic panels, electrical conductors and radiators. More power drives more processing. More processing yields more refined materials. The flywheel begins to turn. Better systems drive mROI toward full self-replication.

These three steps are how you build a mass driver on the Moon, manufacture the compute payloads for it to launch, and power the manufacturing and launch operation.

This is the insight that the launch bottleneck argument points toward, and that the data center in space argument obscures. The goal is not to put compute in orbit. The goal is to put the supply chain for compute on the Moon, where raw materials are available, where the gravity well is shallow, and where each investment in infrastructure compounds into the next. Data centers launched from Earth matter only insofar as they generate free cash flow to move the supply chain to the Moon. They are a means, not an end.

The end is a self-sustaining industrial base on the Moon, growing at a rate determined not by launch capacity from Earth but by the expanding energy and material resources of the Moon itself. Once that base reaches sufficient scale, it does not need Earth anymore.

This is when exponential growth begins in earnest. 

Part III: The Machine Economy Moves Sunward

When people imagine artificial intelligence going to space, they imagine it going out:  stations orbiting Jupiter, probes to Alpha Centauri, the romantic trajectory of human exploration extended to machines.

Physics points in the opposite direction.

The machine economy has one primary input: energy. And energy in the inner solar system follows a simple law: intensity scales with the inverse square of distance from the Sun. Every time you halve your distance to the Sun, you quadruple the power available per square meter.

Earth sits at one astronomical unit (AU) from the Sun and receives about 1.4 kilowatts per square meter. Mars, at 1.5 AU, receives less than half that. The asteroid belt is worse. The outer planets are negligible as power sources for any industrial civilization. Going outward means accepting a permanent energy penalty, paid in perpetuity.

Going inward is the opposite. At the orbit of Mercury, 0.4 AU from the Sun, the energy flux is roughly 9 kilowatts per square meter — more than six times what is achievable in Earth orbit. The math for a machine economy optimizing on energy is unambiguous: move toward the Sun to the limit of your thermal capabilities.

Mercury is where the logic of the machine economy finds its conclusion.

The planet is small: slightly larger than the Moon, with comparable surface gravity, which means escaping it is only marginally harder than escaping the Moon. It is geologically ancient, the remnant core of a planet that lost its mantle to an early collision, leaving it enriched in reduced metals at higher concentrations and in more useful forms than the lunar regolith. If the Moon is the quarry where the compute supply chain learns to walk, Mercury is the foundry where it runs.

The trajectory from Moon to Mercury is another sequence in supply chain expansion. The lunar supply chain develops the industrial techniques: regolith processing, in-situ power generation, autonomous construction, that Mercury demands at larger scale. Mass drivers on the Moon push another supply chain to Mercury, spinning up the flywheel anew.

At sufficient scale, the machine economy stops being a collection of discrete installations and becomes a continuous industrial system spanning the inner solar system: processing material from Mercury, generating power from the Sun, building compute wherever the thermal architecture allows, and radiating waste heat into space. This is what a Dyson swarm actually is. Not a single structure, not a rigid shell, but a diffuse cloud of independently operating compute nodes, each a thermodynamic sandwich of solar collector, compute layer, and radiator, collectively harvesting a significant fraction of the Sun’s total output.

The Sun produces 3.8 × 10²⁶ watts. Earth intercepts approximately two billionths of that. A Dyson swarm harvesting even a modest fraction of the Sun’s total output represents an economic opportunity so far beyond anything achievable on Earth that the comparison becomes almost meaningless. If you equate the machine economy’s output to energy consumed, which is a reasonable approximation when energy is the master resource, then the economic scale of a Dyson swarm is literally billions of times larger than the sum of everything humans have ever built or could build on Earth.

This is the context in which to understand "data centers in space". It is about positioning for this. About being the entity that controls the first rungs of the ladder: the launch capacity, the lunar supply chain, the early orbital compute, when the machine economy begins its move toward the Sun in earnest.

The competitive dynamics of this transition are severe. Exponential growth is extraordinarily sensitive to initial conditions. A supply chain that begins compounding on the Moon five years before its nearest competitor does not end up five years ahead. It ends up orders of magnitude larger. First mover advantage in an exponential system is not a head start, it is a structural divergence that the laggard may never close.

The players in this compete to participate in the machine economy at Kardashev scale. The Moon is the first industrial node. 

From this perspective, every rocket company without a compute strategy and every AI lab without a launch strategy is making the same mistake: optimizing for the human economy at the precise moment when the machine economy is beginning to diverge from it.

The window to correct that mistake is open. It will not stay open.

Part IV: Why This Is Good: Earth is for Humans and Nature

There is an objection that runs underneath everything written so far. It goes like this: if we are describing a machine economy that grows without bound, consumes the resources of the inner solar system, and builds structures of civilizational scale, who is this good for?

This is a fair question which deserves a direct answer. The idea of a machine economy growing in space is actually the best solution for Earth, for humans and nature. 

First know that the machine economy is not optional. Growth is a dominant strategy in any resource development competition, and we already know that the largest economic actors are committed to growth beyond terrestrial scale. The logic driving growth:compute generating demand for more compute, automation enabling faster automation, capital flowing toward the highest returns, is not a policy choice that can be voted away. It is the aggregate behavior of millions of actors, each making individually rational decisions, compounding into a system dynamic that no single government or coalition has the leverage to stop. The question is not whether the machine economy grows. It is where.

If it grows on Earth, the consequences are already visible in outline. Data centers are the fastest-growing consumer of land, water, and electricity in the developed world. The regulatory and political backlash is real and intensifying because the human and machine economies are in conflict for resources. The machine economy is optimizing on a single variable at exponential speed. It is not a competitor that human institutions are well-designed to manage.

The waste heat problem alone is clarifying. Every watt of power consumed by a data center becomes waste heat that must go somewhere. On Earth, it goes into the atmosphere, the water table, or the local environment. Greenhouse gas emissions can in principle be addressed by decarbonizing the power supply: run data centers on nuclear or solar and the carbon problem is manageable. The waste heat problem cannot be decarbonized away. It is thermodynamically unavoidable. At roughly 1,000 terawatts of global primary energy consumption, waste heat alone begins to warm the planet by approximately one degree Celsius, independent of what fuels the power. That is fifty times current levels, well within the range the machine economy reaches in the coming decades if it stays on Earth. Our home planet has a hard physical ceiling for industrial activity, and the machine economy will hit it.

In space, waste heat is not a problem. The radiator facing cold space is using a truly infinite and free resource. Moving the machine economy off Earth relieves pressure on the biosphere and removes the machine economy from a thermodynamic regime where its growth is fundamentally in tension with human habitation.

This reframing matters for how we think about our precious Earth. If the machine economy moves off Earth, then Earth stops being the site of an unresolvable conflict between industrial growth and human flourishing. It remains the place that is perfect for humans and nature while machines have better options elsewhere.

Humans evolved here. We are adapted to this gravity, this atmosphere, this circadian rhythm and temperature and biological richness. We did not evolve for microgravity or radiation or the harsh void of space. We need Earth. Machines do not, and are in fact, better suited elsewhere.

This natural divergence of interests, with machines wanting energy which means going toward the Sun and humans wanting Earth which means staying here, is an exit from the conflict that would otherwise define the coming century. Geopolitical realism, applied to the relationship between human and machine civilization, suggests that the most stable outcome is one where their interests are genuinely separated, not one where they are forced to compete for the same finite resources on the same finite planet.

The vision that emerges from following this argument to its conclusion is a conserved Earth. A planet where the pressure of exponential industrial growth has been relieved not by constraining growth, which is probably impossible, but by redirecting it toward an environment where it causes no harm and faces no ceiling. A planet where rewilding is possible not as a marginal concession to environmentalism but as the natural consequence of industry having found a better home. Machines, given the choice, would rather be somewhere else.

Earth is for humans and nature. Sunward of Earth is for industry and compute. Mars and beyond is for exploration.

Conclusion: Competitive Dynamics of Exponential Scaling

SpaceX is the first organization to attempt vertical integration of the entire stack required to participate in the machine economy at Kardashev scale: reusable launch, cislunar logistics, orbital infrastructure, and now via xAI and the planned Terafab, the compute layer that justifies building all of it. 

This is a rational strategy. It is probably the most important corporate strategy in the history of industrial civilization. And it deserves competition.

Not because SpaceX will fail, but because they will probably succeed. A Dyson swarm built by a single actor is a significantly worse outcome than one built by many. The machine economy growing in space is good for humans and nature. The machine economy growing in space under a single point of control is a concentration of civilizational power with no precedent and no obvious check. The diversity of interests, values, and approaches that makes human civilization resilient on Earth does not automatically replicate itself in space. It has to be deliberately built there, by multiple actors, competing and cooperating across a genuinely open frontier.

This is not an abstract concern. The economics of exponential growth in a constrained competitive environment strongly favor whoever moves first and fastest. Once the supply chain achieves economic closure, no longer needing inputs from Earth to sustain and grow itself, the window for terrestrial capital to buy meaningful ownership closes permanently. You cannot participate late. The economics do not allow it.

This means the decision that matters is not whether to eventually pursue this strategy. It is whether to pursue it now, while the supply chain is still being built from Earth, while the first rungs of the ladder are still accessible to outside capital and outside talent, while the frontier is still genuinely open.

For rocket companies, the implication is straightforward. Launch capacity is necessary but not sufficient. A rocket company that only launches payloads for other people’s missions is a contractor in someone else’s supply chain. It will not be the entity that captures the economic returns of Kardashev-scale growth. Rocket companies are valuable because they provide access through the launch bottleneck, but the scaling exponent of the machine economy is mass returned on investment beyond the launch bottleneck. Rockets are the base of the exponent, not its scaling factor. The entities that capture Kardashev scale returns will master the scaling exponent. Every serious rocket company should be asking what version of this stack they can assemble, and who they need to partner with to assemble it.

For AI labs and hyperscalers, the implication is less obvious but equally urgent. You are already building the demand side of a Dyson swarm. Every data center you construct, every gigawatt of power you contract, every chip you procure is a node in a machine economy that has a natural endpoint in space. The question is whether you have any ownership of the supply side scaling in space beyond the launch bottleneck, or whether you will be a customer of somebody who does. 

The investment required to participate in establishing a supply chain beyond the launch bottleneck is large by the standards of venture capital and modest by the standards of hyperscaler capital expenditure. The returns are measurable in a position in the economy that replaces all current economies.

It is worth saying plainly: it is infinitely better for humanity and for nature that even just one player dominates this than if no one does. Earth does not want to be in resource conflict with the machine economy. 

But the best outcome is a solar system where no player dominates, with multiple supply chains, multiple power bases, multiple visions of what machine civilization should look like and what relationship it should maintain with the human economy and Earth’s natural environment. The solar system should be made as diverse and resilient as Earth.

That future has to be chosen. It has to be funded and built by people and organizations that understand what they are building and why it matters. The return on capital will be astronomical, but what really matters is what kind of new civilization we are building.

The supply chain for a Dyson swarm is being assembled right now. The Moon is the first node. The window is open.

The question is who shows up.

Ross Centers is CEO of Ethos Space and founder of the Planetary Sunshade Institute. Brogan BamBrogan is cofounder of Ethos Space and was employee #23 at SpaceX.

ross@ethos-space.com