Flows, not stock
A Settlement Lives on Three Flows, Not Its Inventory
Metal, volume, spares and vehicles can wait; energy, water and computation pass through the site every hour.
Every settlement begins with an inventory - metal, shielding, pressure vessels, excavators, pumps, machine tools, suits and vehicles. A site can postpone a machine-shop expansion, defer a habitat module or cannibalize a rover and wait months for a bearing. It cannot stop energy, water and computation.
Energy comes first because power turns the other two into something useful: ice in regolith becomes water only after machines dig, heat, separate and purify it, and silicon computes only while power enters and heat leaves. Water is the material intermediary - drunk, split into hydrogen and oxygen, used as coolant and process mass, and stacked as shielding. Compute moves no mass or power itself but decides what the mass and power do: which load runs, which machine waits, which data crosses the link, and which decision is made locally because Earth is too far away.
| Item | Kind | What happens if production pauses for a day | Rated |
|---|---|---|---|
| Structural metal, habitat volume | Stock | Nothing; it is installed | Modeled |
| Spare parts, vehicles, machine tools | Stock | Nothing; governed by failure rates | Modeled |
| Electrical energy | Flow with a buffer | Batteries and fuel cells carry the loads that cannot be shed | Modeled |
| Water | Flow with an inventory | Tanks carry the site; recovery loops keep losses small | Modeled |
| Computation | Flow with no reserve | Unused cycles are gone; only precomputed work survives | Modeled |
Buffer, inventory, nothing
Energy Can Wait in a Battery and Water in a Tank; a Processor Cycle Cannot Wait
A kilowatt-hour can be stored and a ton of water can be stored. A cycle unused at 10:03 is gone at 10:04.
| Substrate | Reserve form | Governing discipline | Failure that matters | Rated |
|---|---|---|---|---|
| Energy | Buffer: batteries, regenerative fuel cells | Buffering | Generation below the unsheddable load for longer than the buffer lasts | Modeled |
| Water | Inventory: tanks and recovery loops | Allocation | Crew, propellant, storage and shielding drawing on one unprotected reserve | Modeled |
| Compute | None: only cached models, stored procedures and local autonomy | Placement and preparation | A decision needed faster than Earth can answer, with nothing prepared locally | Modeled |
Water is easy to store - the tank already exists and recovery loops return most of what people and machines use - so the water problem is allocation among crew, propellant, cooling, shielding, agriculture and industry. Compute has no tank behind the wall. Redundant processors supply spare capacity, but the only reserve against a lost link or a failed cluster is work done beforehand: cached models, stored procedures, precomputed trajectories, compressed data products and rules for what the site does when nobody on Earth can answer in time.
The triangle
Energy Extracts Water, Water Stores Energy, Compute Decides Where Both Go
The three convert into one another at exchange rates that can be written down.
| Conversion | Rate | Basis | Rated |
|---|---|---|---|
| Electricity to hydrogen and oxygen (electrolysis) | 4.4 kWh/kg water ideal; about 6.3 at 70% efficiency | Heat of formation of water, 286 kJ/mol | Observed |
| Hydrogen and oxygen back to electricity (fuel cell) | About 2 kWh per kg of water cycled | Lower heating value 3.7 kWh/kg at 55% fuel-cell efficiency | Modeled |
| Heat to water from icy regolith, 5 wt% ice | About 1.4 kWh/kg ideal; 3.6-5 realized | 19 kg of regolith warmed 200 K plus sublimation, 0.79 kWh/kg | Modeled |
| Same, 2 wt% ice | About 2.4 kWh/kg ideal; 6-8.5 realized | 49 kg of regolith per kg of water | Modeled |
| Water to oxygen by mass | 0.89 kg O2 per kg water | Stoichiometry; the other 0.11 kg is hydrogen | Observed |
| Electricity to compute heat | 1 W in, 1 W of heat out | Every watt into an NVIDIA H100 (700 W) leaves as heat | Observed |
| Heat rejection at 300 K | About 1.2 m2 of two-sided panel per kW | Emissivity 0.9, deep-space view; several times larger facing sunlit regolith | Modeled |
Power turns buried ice into water. Electrolysis splits the water into hydrogen and oxygen, and a regenerative fuel cell recombines them to return power when sunlight is gone or a reactor is offline. The same electrolysis chain makes breathing oxygen and propellant. Compute converts power to heat almost one for one, and saves power and bandwidth by deciding what not to run, what not to send and what can wait.
One inventory
The Water Is Fungible; the Reserve Cannot Be
One lunar water inventory is energy storage, propellant, oxygen, coolant and shielding at once - leverage and common-mode risk together.
| Claim | Annual draw | Consumed or held | Reserve rule | Rated |
|---|---|---|---|---|
| Breathing oxygen by electrolysis | About 1.4 t water | Consumed (hydrogen partly recovered) | Protected; never transferable | Modeled |
| Drinking and hygiene make-up at 98% recovery | About 0.1 t | Consumed | Protected; the ISS Water Recovery System reached 98% in 2023 | Observed |
| Propellant oxygen, 50 t a year | About 56 t water | Consumed | Production stock; first to be paused | Modeled |
| Night storage, 25 kW for 100 h | About 1.3 t cycled | Held; returns to the tank | Separate tank and transfer path | Modeled |
| Storm-shelter shielding, 10 cm wall | About 100 kg/m2 | Held in place | Physically isolated; not drawn down | Modeled |
| Thermal-loop coolant | Under 1 t standing | Held | Fixed in the loop | Modeled |
A ton of water cannot be run through a fuel cell, loaded into a lander and left in place as an emergency reserve at the same time. Terrestrial utilities rarely draw physical lines through a fungible commodity; an off-Earth site has to: separate tanks, separate transfer paths, protected reserves, and rules operations cannot override. Breathing oxygen alone draws more water each year than drinking-water losses do at 98% recovery.
One bus
Installed Capability Is Not Usable Capability: 204 kW of Peak Demand on a 150 kW Bus
An early lunar site owns oxygen production, ice extraction and fabrication, and cannot run all of them at once.
| Load | Average | Peak | Can it pause? | Rated |
|---|---|---|---|---|
| Life support and habitat, four crew | 15 kW | 15 kW | No | Modeled |
| Thermal control, avionics, communications | 10 kW | 10 kW | No | Modeled |
| Oxygen by electrolysis, 50 t a year | 40 kW | 40 kW | Yes - production campaign | Modeled |
| Ice extraction at 5 wt% (55 kW at 2 wt%) | 32 kW | 32 kW | Yes | Modeled |
| Excavation and haul | 5 kW | 5 kW | Yes | Modeled |
| Local compute cluster, eight H100-class | 10 kW | 10 kW | Partly - shift to margin | Modeled |
| Rover charging | 10 kW | 10 kW | Yes | Modeled |
| Machine shop | 6.7 kW | 31.6 kW | Yes - except an urgent repair | Modeled |
| Regolith sintering campaign | - | 50 kW | Yes - campaign | Modeled |
| Total | 129 kW | 204 kW | Firm load: 25 kW | Modeled |
Illustrative loads sized from the exchange rates above; the machine-shop figures come from The Machine Shop MVP. The 150 kW bus matches the order of an early lunar site with solar on a lit ridge plus a first reactor such as NASA Lunar Reactor-1.
A settlement is a schedule for a finite bus, not a catalog of machines. Oxygen production slows while a fabrication run is underway; a noncritical refining step waits through a long eclipse; sintered pads are made in campaigns; compute shifts to hours with generation and radiator margin. The cheapest storage at a frontier site is sometimes neither battery nor tank but permission to stop making something. The minimum viable outpost is the smallest architecture that keeps the firm 25 kW alive and schedules everything else. Landing equipment unlocks a rung of the industrial ladder; allocating power makes it real this week.
| Darkness | Energy | Water cycled (fuel cell) | Li-ion at 175 Wh/kg | Rated |
|---|---|---|---|---|
| Best south-pole ridge outage, about 100 h | 2,500 kWh | 1.3 t | 14 t | Modeled |
| Equatorial lunar night, 354 h | 8,850 kWh | 4.4 t | 51 t | Observed |
Storing a ridge outage as water in a regenerative fuel cell cycles about 1.3 t of water against 14 t of lithium-ion cells, before tanks and stacks. The same tank also holds oxygen and propellant - which is why the reserve rules above exist.
Co-location
Off Earth, the Map That Matters Is Electrical
Two machines on one bus are one site at any distance; two machines with separate power are separate sites at a hundred meters.
A solar array on a lit ridge of Shackleton crater and an ice plant in permanent shadow kilometers below act as one industrial site if they share generation, storage and distribution. The best place to mine ice is rarely the best place to refine it, the best place to generate power is rarely the best place to reject heat, and the best place to house people is rarely the best place to run dense compute. A mature site is a network of specialized nodes joined by power lines, pipes, vehicles and links - an infrastructure topology, not a compact Moon base or Mars base. The corridor and node maps describe the same structure at solar-system scale.
| Destination | Sunlight vs Earth | Water source | Heat sink | Light time, one way | Rated |
|---|---|---|---|---|---|
| Earth orbit | 100% | Shipped from Earth | Deep space, sunlit half the orbit | Under 0.15 s | Observed |
| Moon, south pole | 100%, grazing; ridges lit 80-90% of the year | Ice in permanently shadowed craters; LCROSS measured 5.6 wt% at Cabeus | Shadowed craters near 40 K | 1.3 s | Observed |
| Mars, mid-latitudes | 43%, cut for weeks by dust storms | Buried ground ice mapped by the SWIM project; hydrated minerals | Thin CO2 air, about 210 K | 3-22 min | Observed |
| Ceres | 13% | Near-surface ice at high latitudes (Dawn) | About 160 K surface | 15-31 min | Observed |
| Callisto | 3.7% | Surface ice-rock mix, about half ice | About 120 K surface | 33-53 min | Observed |
Sunlight sets the energy answer. On the Moon, ridges at the south pole see the Sun most of the year and solar carries the site until a reactor arrives. At Mars, dust storms and 43% sunlight push every serious plan toward fission. At Ceres and Callisto solar is too weak to run industry, and the site exists only with a reactor. Light time sets the compute answer: at 1.3 s the Moon can be teleoperated from Earth; at 3-22 min Mars cannot, and every decision needed faster than the 6-44 minute round trip is made on site.
China
China Is Building on All Three Substrates
A lunar reactor with Russia, a polar ice survey, ISRU demonstrations and a flying compute constellation.
| Program | Substrate | Status | Rated |
|---|---|---|---|
| ILRS lunar nuclear power station · CNSA and Roscosmos | Energy | Agreement May 2025; delivery 2035, completion around 2036 | Projected |
| Chang'e-7 · CNSA | Water | South-pole ice survey with a hopping probe into a shadowed crater; launch postponed 24 August 2026 | Building |
| Chang'e-8 · CNSA | Water and materials | ISRU demonstrations at the south pole, planned around 2029 | Projected |
| Three-Body Computing Constellation · ADA Space and Zhejiang Lab | Compute | 12 laser-linked computing satellites flown since May 2025 | Observed |
The CNSA-Roscosmos reactor targets the same south-pole region as Artemis and the same decade as NASA's Lunar Reactor-1. Chang'e-7 is the first mission built to measure ice inside a shadowed crater rather than infer it from orbit; its result sets the extraction grade - 5 wt% or 2 wt% - that changes the bus by 23 kW at the site above.
The stack
Every Layer Above Assumes the Three Flows Already Exist
Materials, Ship or Make, transportation, habitats, the human layer and the machine shop all draw on the same bus, tanks and processors.
| Layer | Energy | Water | Compute | Rated |
|---|---|---|---|---|
| Materials | Smelting and electrolysis power | Process water and reagent | Process control | Modeled |
| Ship or Make | Sets which rungs can run at once | Feedstock for propellant and oxygen rungs | Design and CAM loop | Modeled |
| Transportation | Charging and propellant production | Propellant oxygen and hydrogen | Guidance and traffic | Modeled |
| Habitats and life support | Firm load | Breathing oxygen, drinking water, shielding | Monitoring and alarms | Modeled |
| The Human Layer | Crew-hour power | Crew water budget | What crews no longer have to do | Modeled |
| The Machine Shop MVP | 6.7 kW average, 31.6 kW peak | Coolant (or dry cutting) | Validated part library | Modeled |
A site can hold rich ore and still be poor if extraction costs too much power. It can hold abundant ice and still be fragile if every critical system draws from one unprotected tank. It can hold large compute and still be blind if it cannot reject the heat or if its decisions live minutes away. The operating system under the hardware is three flows, many conversions and one bus.
Contested
Deliberately Not Claimed
The load-bearing uncertainties, tagged with the confidence they actually earn.
Lunar ice grade rests on one impact. LCROSS measured 5.6 wt% at one spot in Cabeus in 2009. Extraction power nearly doubles between 5 and 2 wt%, and no lander has yet sampled ice in a shadowed crater.
Regenerative fuel cells have not run at site scale off Earth. Round-trip efficiency, stack life over hundreds of cycles and cryogenic or high-pressure gas storage at the Moon are untested; the 2 kWh/kg figure assumes 55% fuel-cell efficiency.
98% water recovery is an ISS figure. It was reached in low Earth orbit with resupply as a backstop and does not include agriculture or industrial process water.
Scheduling as storage assumes processes can stop. Cryogenic plants, furnaces and electrolyzers lose efficiency and life when cycled; how often each can pause is unmeasured for off-Earth hardware.
Record
Provenance and Data Sources
- Confidence
- Observed for flown, fielded or measured - Building for hardware and programs in development - Projected for announced plans - Modeled for original analysis
- Status date
- Q4 2026
- Physical data
- Water heat of formation and heating values; LCROSS Cabeus result (2009); NASA ISS Water Recovery System (2023); Mars SWIM ice maps; Dawn at Ceres
- Programs
- NASA Lunar Reactor-1; CNSA-Roscosmos lunar power station agreement (May 2025); CNSA Chang'e-7 postponement (24 August 2026); Chang'e-8 plans; Three-Body Computing Constellation
- Weakest link
- Lunar ice grade and regenerative fuel-cell performance at site scale