TypeIITypeIITypeIITypeIITypeIITypeIITypeII

Thesis

Earth is not energy-scarce. It is heat-limited.

The ceiling on how much energy humanity can use on this planet is thermodynamic, not chemical — and it holds no matter how clean the source is. That single constraint is what turns industry in space from an aspiration into a requirement. What follows is the whole chain, from the physics to a robotics company.

Five links, none of them optional

Each step follows from the one above it. Any of them can be attacked on its merits; none of them requires believing a forecast.

  1. 01

    The master resource

    Every capability humanity has ever added — agriculture, industry, computation — shows up as a step change in energy throughput. Believing human capability keeps growing is already believing energy use keeps growing.

    Premise

  2. 02

    The ceiling

    Every watt consumed at the surface ends as heat in the surface system. That is true of fusion, of solar, of anything. At roughly 100× today’s consumption, waste heat alone produces more radiative forcing than doubling atmospheric carbon dioxide. A perfectly clean civilization that grows two orders of magnitude still cooks itself.

    Thermodynamics

  3. 03

    The consumption moves, not just the generation

    This is the step most people skip. Beaming orbital power down to Earth keeps the heat exactly where the problem is. The only version that works puts the industry and the computation off-planet too, radiating waste heat into a three-kelvin sky instead of into the biosphere.

    Consequence

  4. 04

    Industry means structures

    Everything in orbit today is built on Earth, folded into a fairing, launched, and unfolded. Two hard limits follow: nothing can be larger than a rocket, and nothing can be repaired. Every large-structure concept ever drawn reaches the same step and stops.

    Constraint

  5. 05

    The step is assembly

    Assembly in space has only ever meant astronaut hours — the most expensive labour humans have ever purchased. It does not scale by a factor of ten, let alone the factor required. Autonomous construction is the gate on everything above it, and it is a robotics problem.

    The company

The four numbers the argument rests on

All four are derivable from public constants — solar luminosity, orbital distance, Earth’s radius and albedo, and global primary energy consumption. Nothing here depends on a proprietary estimate.

3.8 × 1026 W The sun’s output

Radiating continuously in every direction, whether or not anything catches it.

1 in 2.2 billion What Earth intercepts

The planet’s entire share of its own star. Everything alive has always divided this.

1 in 20 trillion What civilization uses

All human primary energy consumption, as a fraction of one ordinary star.

3.9 W/m2 Waste heat at 100×

Forcing from human waste heat alone at 100× today’s use. Doubling CO2 is 3.7.

The limit cheap launch does not move

Falling launch cost solves delivery. It does nothing about geometry. A payload bay is a fixed cylinder, and no price reduction makes it wider: nine metres is nine metres at any dollar per kilogram.

The clearest case is orbital compute. SpaceX has said publicly that it will build data centres in orbit by scaling up laser-linked Starlink satellites — free-flying units, each launched complete, needing no assembly whatsoever. That architecture works, right up to the point where a single facility needs gigawatt-class power.

A gigawatt of computing must reject a gigawatt of heat, and in vacuum the only route is radiating it. At a radiator temperature realistic for electronics, that is a surface close to a kilometre on a side.

1 GW One facility

Roughly where a distributed mesh of free-flyers stops substituting for a tightly coupled cluster.

653,000 m2 Radiator required

Two-sided, emissivity 0.9, radiator at 350 K. A square 808 m on a side.

9 × 18 m Starship payload bay

A footprint of 162 m2, and a hard geometric cap on any single launched object.

4,000× The gap

What assembly has to close, and what no launch price closes on its own.

Folding is why this is not needed yet

The area depends on radiator temperature and almost nothing else: 383,000 m2 at 400 K, 1.21 million at 300 K. Published figures in the low thousands of square metres are not physical — they imply a radiator running above 1,200 K, and the electronics being cooled run near 350.

The obvious escape is folding, and it has worked so far. But deployables scale badly: every hinge is a single point of failure with no repair path, which is part of why a twenty-five square metre mirror took decades and billions. A square kilometre is not a deployable. It is a construction project.

Starship makes mass cheap. It does not make the fairing bigger.

Some functions do not decompose

The obvious reply to a fairing limit is to stop building one big thing. Starlink is the proof that this works: for coverage, aggregate bandwidth, and graceful degradation, a thousand small satellites beat one large one and always will.

So the question is not large against small. It is which functions can be faked by adding units. The test is whether the thing you want scales with total quantity or with a single continuous dimension. Coverage, throughput and redundancy are quantities, and quantities distribute. Diameter, radius, rigidity and coupling are dimensions, and no number of small units sums to one large one.

What a constellation can and cannot substitute for
Function What sets its size Needs one structure
Coverage How many units are in the shell. More units, more sky, immediately. No
Downlink How many units are in the shell, again. This is precisely what a constellation is for. No
Optical resolution Aperture diameter, by λ/D — the largest single dimension, not the total collecting area. Yes
Artificial gravity Rotating radius, by a = ω2r. A formation cannot produce centripetal acceleration. Yes
Frontier training Interconnect between accelerators, which falls off a cliff the moment they stop sharing a chassis. Yes
Propellant storage Volume against boundary area. Boil-off is a surface effect; capacity is a volume. Yes

What those dimensions actually cost

Artificial gravity. At two revolutions per minute — the usual ceiling before people get sick — one gravity needs a 224 m radius, a structure 450 m across. Even an unpleasant four rpm still needs 112 m. There is no arrangement of free-flyers that produces this.

Resolution. A thousand one-metre telescopes have the light-gathering of a thirty-metre mirror and the resolution of a one-metre one. Radio interferometry genuinely escapes this, by recording phase and correlating later. Optical does not: the light has to be physically combined with path lengths matched to a fraction of 500 nm, which means nanometre control across kilometres.

Coupling. In-rack interconnect runs near 14 Tbps per accelerator; a generous optical link between satellites is 100 Gbps. That is 144× less bandwidth, and 333 µs of light time over 100 km against roughly one microsecond in a chassis. Inference distributes comfortably. A training run that synchronises gradients at every step does not.

Storage. Boil-off scales with surface area and capacity scales with volume, so a 20 m tank loses a tenth the fraction per unit stored that a 2 m tank does. For depots, many small units is not neutral. It is the worst available option.

A constellation is a fleet. What comes after it is a facility.

Why this is a company now and was not in 2015

Launch cost is collapsing. When delivery gets cheap the constraint moves to the next link in the chain. Cheap launch converts “we cannot afford to put mass up there” into “we cannot do anything with the mass once it arrives.” That shift is happening in real time, which is why this plan gets more valuable as launch improves and never requires anyone else to stumble.

Autonomous manipulation became real. Learned manipulation moved from scripted, fixture-dependent motion to something approaching general within about the last four years. The core technical capability this requires did not exist when the previous generation of orbital-robotics companies was funded.

Compute demand is hitting physical limits. Datacentre siting is already constrained by power and cooling. That is a concrete near-term pull toward off-planet industry, on a timeline measured in decades rather than centuries.

Why a company and not an agency

The problem is repetition, not exquisiteness. Agencies and primes are structurally excellent at building one magnificent object on cost-plus, and structurally incapable of making the ten-thousandth unit cheap — which is the only thing that matters here.

A swarm is not an engineering marvel. It is a manufacturing and logistics problem wearing an engineering marvel’s clothes, and that is commercial competence.

What is genuinely uncertain

Timeline. The waste-heat ceiling determines where growth has to go, not when. It is a direction-setting argument and should not be read as a date.

Whether autonomy is good enough yet. Manipulation improved dramatically; nobody has shown it is sufficient for unrecoverable assembly of novel structures. That is the technical bet, stated as a bet.

The first market. The sequencing logic is strong — the right beachhead is the one that already shares orbit’s binding constraint. Which specific terrestrial environment that turns out to be is still open.

Argue with this Why any of it matters

Current status, stated plainly. There is no legal entity, no hardware, no prototype, no contract, no grant, and no team. Nothing on this page is in progress.

The reasoning is published so it can be attacked before anything is built on top of it. At this stage the reasoning is the only asset.