Cold Compute at the Poles

The case for cold compute

Why the Idea Is Compelling: Free Cooling and Free Vacuum

Two of the most expensive things in cryogenic computing are the cryostat and the vacuum chamber. The Moon provides both for free.

On Earth, cooling a superconducting quantum computer to millikelvin temperatures requires a dilution refrigerator that costs roughly $600,000, consumes continuous electrical power, and occupies a room-sized support stack. Reaching even the intermediate 77 K stage requires liquid nitrogen infrastructure. The vacuum chamber that houses the experiment costs more. Every watt dissipated inside the cryostat costs hundreds of watts of room-temperature power to remove — a terrestrial cryostat operating at 4 K draws 200–1,000 watts from the power supply for every watt consumed inside it.

The lunar PSR offers two of those three things for free. The floor of Shackleton crater sits at roughly 88 K. The deepest micro-traps in Hermite and Haworth reach approximately 25 K. The lunar exosphere at ∼10⁻⁹ bar is a better vacuum than most terrestrial cryostats achieve. A computer placed in a PSR starts at 40 K in hard vacuum without building or powering anything. The question is whether that head start is worth the cost of getting the hardware there.

Three tiers of cryogenic computing

77 K, 4 K, and 10 mK: Different Technologies, Different Requirements

The PSR gives you 40 K for free. That is enough for one tier, a head start on a second, and nowhere near the third.

Cryo-CMOS is the tier where the PSR case is strongest. A 2024 Nature Electronics perspective showed that cooling silicon transistors to 77 K breaks the fundamental thermionic switching limit, reduces thermal noise, improves carrier mobility, and enables tighter voltage margins. The PSR floor is colder than 77 K, so the environment does the work a liquid-nitrogen system does on Earth, with no consumable.

SFQ logic (Rapid Single Flux Quantum and its variants) operates at roughly 4 K with switching energies of 10⁻¹⁸ J per gate — roughly a million times lower than CMOS. An RSFQ chip with one million gates running at 25 GHz dissipates approximately 25 mW. That is an astonishingly low heat load. But 4 K is not 40 K, and a cryocooler is still needed for the final cooling step.

Quantum computing requires 10–20 mK — three orders of magnitude colder than the PSR floor. The 40 K ambient reduces the work the dilution refrigerator must do for its upper stages, but the millikelvin step is where the difficulty and the He-3 consumption sit. The PSR helps but does not solve the hard part. Modeled

The thermal budget kills scale

Heat In, Cold Out: Why You Cannot Run a Data Center in a Cold Trap

The same constraint that limits ISRU and the SMES limits compute. The cold trap has a thermal budget measured in fractions of a watt per square meter.

A surface at 40 K radiates 0.145 W/m². At 25 K, 0.022 W/m². Those numbers are the thermal budget of the cold trap — the rate at which it can reject heat to space and maintain its temperature. Any heat source placed in the trap must stay within that budget or the trap warms.

At the SFQ scale, the numbers look benign. 25 mW for a million-gate chip is invisible against even a modest thermal budget. But scale it. Ten chips: 250 mW. A thousand chips: 25 W. Add the cryocooler waste heat (even a high-efficiency 4 K cooler rejects several watts per watt of cooling at the warm end), the power-conversion electronics, and the communications equipment. A modest compute installation quickly reaches kilowatts of total thermal dissipation.

A kilowatt radiated into a 40 K floor that passively rejects 0.145 W/m² requires roughly 6,900 m² of floor area as a heat sink — a circle 47 meters in radius — just to absorb the heat without warming. At any industrial compute scale, the cold trap warms. The environment you came to exploit is destroyed by using it.

This is the same constraint that governs ISRU extraction and the mass-driver SMES. Early analysis explored siting both SMES and compute in the cold trap. Both were rejected for the same reason: the trap is a finite thermal resource, not a free cooling utility. Observed

The latency constraint

2.56 Seconds Round Trip: What Workloads Survive the Light-Time Delay

The Moon is 1.28 light-seconds away. That rules out interactive computing, real-time inference, and anything with a human in the loop on Earth.

The Earth–Moon distance imposes a minimum round-trip latency of approximately 2.56 seconds. No amount of bandwidth or processing speed reduces this — it is the speed of light. Any workload that requires a response within that window cannot be served from the Moon.

What survives: batch computation (training runs, Monte Carlo simulations, protein folding, climate modeling), asynchronous data processing, archival storage and disaster recovery, and autonomous systems that process data locally and return results. What does not survive: interactive queries, real-time API serving, latency-sensitive AI inference, financial trading, gaming, videoconferencing, or any workflow where a human waits for the answer.

Lonestar Data Holdings launched its Freedom data center payload on an Intuitive Machines lander in February 2025, targeting disaster-recovery and data-archival use cases — exactly the latency-tolerant workload class. That mission demonstrated the concept but did not depend on cold-trap cooling; it landed near the equator. Building

The science instruments that do close

Ultrastable Lasers and Precision Instruments: The PSR Cases That Work

The case that closes is not a data center. It is a precision instrument that needs exactly what the PSR provides.

In February 2026, a team from JILA/NIST, JPL/Caltech, PTB (Germany), and Lunetronic Inc. published a design for a Lunar Silicon Cavity — an ultrastable optical cavity to be installed in a PSR. The PSR environment provides three things the cavity needs simultaneously: cryogenic temperature (for zero-thermal-expansion silicon at 124 K or below), ultra-high vacuum (no gas-molecule perturbation), and exceptionally low vibrational noise (no weather, no traffic, no tectonic activity). The resulting laser would have fractional frequency stability surpassing the best terrestrial systems, forming the basis for gravitational-wave detection, optical atomic clocks, and quantum-entangled satellite networks.

This case closes because the instrument is small (watts, not kilowatts), its heat load is negligible against the trap’s thermal budget, it does not need to communicate at low latency, and the PSR environment is not just helpful but irreplaceable — no terrestrial facility can match the combination of cold, vacuum, and seismic stillness simultaneously and indefinitely. Building

Cold compute as an incremental benefit

If You Are Already There for Other Reasons, the Cold Is Free

Cold compute does not justify going to the Moon. But if you are already on the Moon, it is a free benefit.

If SpaceX builds a lunar manufacturing facility to produce AI satellites (as described on the mass-driver page), some of those satellites will need testing and calibration before launch. Running them on the surface in the ambient cold before committing them to orbit is an incremental use of the environment, not a standalone business case. The compute hardware was going to be manufactured there anyway. The cold is a bonus.

Similarly, if a crewed outpost operates at the south pole for ISRU, science, and logistics, some local compute is needed regardless — mission control, autonomous vehicle management, ISRU process control, communications processing. Placing the compute hardware in a shaded annex rather than in the sunlit habitat saves cooling power. That is a real engineering benefit. It is not a reason to build the outpost.

The cold-compute case is incremental, not foundational. It rides on top of an outpost or factory built for other reasons. It does not pull the first dollar of investment to the Moon. Projected

Adjudication

Does Cold Compute Close? Three Verdicts by Use Case

Standalone lunar data center — computing for Earth customers served from a cold-trap facility. Does not close. Delivery cost overwhelms cooling savings. The 2.56-second latency rules out interactive workloads. The thermal budget limits scale — any industrial compute installation warms the trap it came to exploit. And terrestrial cryocoolers are getting cheaper faster than lunar delivery is getting cheaper. There is no price point at which shipping servers to the Moon beats buying a better refrigerator on Earth for general-purpose computing. Modeled

Precision science instruments — ultrastable cavities, atomic clocks, gravitational-wave detectors. Closes. The heat load is watts, not kilowatts — negligible against the trap’s thermal budget. The PSR environment is irreplaceable: cold, vacuum, and seismic quiet simultaneously and indefinitely, with no terrestrial equivalent. Published designs exist (Lunar Silicon Cavity, Ye et al. Feb 2026). These instruments do not need low latency, do not need to scale, and do need exactly what the crater floor provides. Building

Incremental compute at an existing lunar outpost or factory — process control, satellite testing, batch AI. Closes as a free benefit, not as a driver. If the outpost exists for ISRU, manufacturing, or crew, placing compute hardware in a shaded annex rather than in the sunlit habitat saves cooling power. Real savings, zero additional investment justification required. But it does not pull the first dollar of investment to the Moon. The cold-compute case rides on infrastructure built for other reasons. Projected

The honest summary: the cold is real, the vacuum is real, and the cooling savings are real. But they are not large enough to justify sending hardware to the Moon for computing alone. The cold-compute case closes only when it rides on infrastructure built for other reasons — ISRU, manufacturing, science, or crew presence — or when the instrument requires the specific PSR environment in a way no terrestrial facility can match.

Record

Provenance and Data Sources