A standing premise first, because everything below depends on it: no magnetic fusion device has ever produced more fusion energy than the heating power put into it. The plant in our numbers is a peer-reviewed design, not a machine. With that stated, the question this page answers is simple to ask. A ship that lands on Mars has to make its fuel there to come home. How much power does that take — and at what point, if ever, does fusion become the way to supply it?
Getting somewhere is a propulsion problem. Getting back is a supply problem. A Starship-class vehicle leaving the Martian surface for Earth needs, in the published plan, full tanks: about 1,200 tonnes of liquid methane and oxygen.2 None of it can be carried from Earth economically, so all of it is made on site — water mined from ice and split, carbon dioxide pulled from the air and combined with the hydrogen. That is a chemical plant, and a chemical plant is an electrical load. We sized the load.
Per returning ship: ≈ 10 GWh of electricity (4.6–16.9), of which three quarters goes to splitting water.6
Spread over the roughly 500 days a landed ship waits for its window home, that is about one megawatt of plant capacity per ship (0.5–1.7), running most of the time — an average draw of about 0.8 megawatts. Five published estimates, none derived from ours, bracket the same range.1234 With today’s reference reactors, the power source and the propellant plant for one returning ship weigh roughly 230–275 tonnes on the ground — more than the ship carries as payload. So the first result is plain: coming back is an energy problem before it is a rocket problem.
At fleet scale the numbers stack. A thousand ships per window, the publicly stated ambition,12 needs about a gigawatt of power plant on the surface. That is the largest surface electrical load in anything we reviewed for Mars — and it lands, uninvited, in exactly the power class of a fusion plant.
Before running this, we wrote down what we expected, dated and sealed, together with the result that would prove us wrong. Our claim was that fusion is the return technology of the surface frontier. The computation did not support it.
Fission has no fixed limit for this job. A reference fission unit is designed for 40 kilowatts5 — a design target; none has flown. Twenty-four of them refuel one ship per window, and twenty-four more refuel the next. The practical limits are landings, radiator area and spacing, not physics. Fusion is never the only way home. It can only be the lighter one — and whether it is depends on which fission you compare it with.
| Compared with | Fusion lighter above | Reading |
|---|---|---|
| Today’s fission units (0.0067 kW per kg) | ≈ 560 ships per window280–990; ≈ 330 counting replacements | Fusion winsinside the stated fleet ambition |
| Megawatt-class fission (0.02–0.06 kW per kg) | never, within one plant’s capacitytwo independent design studies; none built | Fission winsunless the plant is partly built on Mars |
| Solar with storage | ≈ 250–830 ships per window | Fusion winsdust derates solar; it does not kill it |
| The Moon, any cadence | never | Not neededsolar and fission land 5–27× lighter |
One more honest line. Dust storms are the usual argument for nuclear power on Mars, and they hold against solar — the 2018 storm cut array output by 97 %.7 But making propellant is a load you can defer and store, so solar is slowed by dust, not stopped. And the argument says nothing at all about fission versus fusion: both run through a storm unchanged. We will not use it for more than it proves.
NASA reached the same starting point from the other side. In December 2024 it chose fission as the primary power source for crewed Mars missions, noting that making propellant could need megawatt-class power.17 Its 2025 model of a Mars fuel plant found fission lighter than solar panels with batteries, and the power system alone up to 59 % of the plant’s landed mass.18 Both documents are recent; our numbers were computed independently of them, and they agree in direction.
What survives is narrower and, we think, more useful. One reference fusion plant — about 850 megawatts net11 — refuels roughly 890 ships per window. A fusion plant cannot be built small; its nuclear shielding does not shrink with it. Its smallest sensible size happens to match the fleet people talk about. That is a coincidence of scale, and it decides nothing on its own.
What decides it are two open questions, both checkable: how light megawatt-class fission becomes — two independent design studies put it at 0.026 to 0.06 kilowatts per kilogram, none built1516 — and whether the heaviest precision parts of a fusion plant — its magnet plate steel — can be made on the surface instead of shipped. If the second is yes, fusion becomes lighter above a few hundred ships even against the better fission. If both go the other way, fusion is not the return technology at all, and this page will say so.
The order does not depend on either answer. Fission arrives with the first ships and adds capacity with every unit landed. A fusion plant delivers nothing until it and the factory that builds it are standing. Fission carries the first several hundred ships home, whatever happens after.
From the lunar surface the climb is shorter, and no lunar carbon resource has been shown to be usable for making methane at scale, so only the oxygen is made locally. A return leg costs 1–2.5 GWh. Fusion’s power class is reached only at around a thousand return flights a year, and at that pace polar solar and megawatt fission both land five to twenty-seven times lighter than a fusion plant. On the Moon, fusion is a plant to prove, not a way home. We had argued the first; this computation removes any temptation to claim the second.
Fusion on a planetary surface is far away, but the distance can be listed. Before any of this matters, a magnetic device has to exceed breakeven, which none has; the best result stands at about two-thirds.10 A breeding blanket has to be measured making its own fuel. A reactor of the hundred-kilowatt class has to run on another world — the only one tested so far ran on the ground, in Nevada, at about a kilowatt.13 Machines have to dig and build off Earth at the tonne scale, which has never been done; the deepest hole beyond Earth is 2.25 metres.14 Then the fleet has to exist, and then the factory that makes the plant. Each is a condition. None has a date. Several have funded programmes behind them. That is what we mean by far ahead, and not never.
Fusion’s hard problems split in two. Some need hardware no model can replace: a burning plasma, a neutron source to test materials, a divertor under real load. Those move at the speed of machines. Others need only analysis that nobody had done — what a plant weighs off Earth, how it rejects heat without air, what it costs to come home. Those now move at the speed of computation. Neural surrogates already run fusion turbulence simulations about 100,000 times faster than the codes they learn from,8 and AI-driven digital twins are part of at least one plant programme on Earth.9 Earth fusion is held up by hardware. Off-world fusion is held up by analysis. The second half is where the acceleration is real, and it is the half this page is made of.
Download the technical memorandum — FUNNY-TM-2026-001 Rev G, Electric Power Requirement for Surface Production of Return Propellant (PDF, 12 pages) → · the calculation script (Python)
Every step, assumption and range, each number tagged as measured, calculated, estimated or single-source. The input that decides the most — how light megawatt-class fission gets — is bracketed by two independent design studies, and the memorandum says plainly that no such reactor has been built. Revision B incorporated an external review: nine corrections accepted, three rejected with the reason shown, three confirmed unchanged, and three further errors we found ourselves while recomputing. Revision D adds NASA’s 2024 decision and 2025 plant model, a process diagram with the energy of every step, and the fusion plant’s itemised mass; Revision F carries its corrected radiator field. Every table in the memorandum is produced by the published script. The revision record lists all of it.
A standing premise of this page: no magnetic-confinement device has yet demonstrated net fusion gain, and the fusion plant in these figures is a peer-reviewed design, not built hardware. Fission and solar figures rest on published programme values. Computed September 2026 in our internal research record; one external cross-model review received and incorporated, no human peer review yet.
Every number above traces to a source or to our own arithmetic, and the memorandum shows the arithmetic. Click a number in the text to land here.
Full derivation, assumptions and sensitivity: FUNNY-TM-2026-001 (download above). Internal research record, September 2026; not externally reviewed.