Frontier MechanicsField report / 077
Frontier Mechanics Week 77: SpaceX files for a million compute satellites, Teton powers nuclear simulation, Nebula becomes a hosted bus, and Claude writes a Mars drive
An orbital data-center proposal collided with thermal math, Idaho quadrupled laboratory compute, Mitsubishi booked a commercial spacecraft platform, and Perseverance tested AI-assisted command generation.
Compute moved into four different physical contexts this week. SpaceX filed an application for an orbital data-center constellation of up to one million satellites. Idaho National Laboratory commissioned Teton for modeling energy and materials. Terran Orbital announced a Nebula bus for Mitsubishi Electric. Anthropic and NASA described how Claude helped produce commands for a Perseverance traverse.
The stories were useful because none allowed software to float free. Compute needs watts and a heat sink, simulation needs validated models, a payload needs a bus, and generated commands need a flight team. Intelligence becomes consequential only after it meets a physical state that can disagree.
The week's systems all relocated computation: into orbit by the million, into a national laboratory's simulation loop, into a hosted spacecraft bus, and into a rover command plan. Relocation changes more than latency. It changes who supplies power, who verifies results, how failure is recovered, and which interface becomes scarce.
Space compute
A million orbital data centers meet the radiator equation
SpaceX filed with the FCC on January 30 for authority to operate a non-geostationary orbital data-center system of up to one million satellites. The scale made the application sound like satire, yet it exposes real tradeoffs. Orbit offers continuous solar input in some geometries and avoids terrestrial land and grid queues. It removes convection, makes repair difficult, and turns every watt of computation into heat that must radiate away.
Radiative cooling follows the Stefan-Boltzmann relation: emitted power scales with area, emissivity, and the fourth power of absolute temperature. A high-power processor cannot dump heat into cold vacuum by touching it; heat must conduct to large radiators with a clear view. Running hotter reduces radiator area and accelerates semiconductor degradation. Sunlight and Earth infrared add external heat loads.
Networking is another boundary. Training clusters move enormous volumes among accelerators with microsecond-scale latency. Satellites separated by kilometers can use optical links, but propagation, acquisition, routing, and topology differ from a data-center fabric. Orbital compute is more plausible for independent inference, preprocessing sensor data, or delay-tolerant jobs than tightly coupled frontier training.
Launch and replacement must enter cost per useful compute-hour. Radiation upsets memory, atomic oxygen attacks surfaces in low orbit, debris risk rises, and hardware generations age faster than spacecraft qualification cycles. Returning or servicing failed units is expensive; uncontrolled failure creates a disposal obligation.
An FCC filing asks for spectrum and orbital authority, not proof of economics. The proposal should be tested in orders of magnitude: watts per kilogram, radiator area per kilowatt, delivered bandwidth, useful operations per satellite, replacement rate, and conjunction burden. A million is not an architecture. It is a request to reserve optionality at planetary scale.
Orbital compute begins with an energy balance. Solar arrays produce intermittent power through pointing and eclipse; processors turn nearly every watt into heat; radiators can reject that heat only toward cold space and away from Earth or Sun. A one-gigawatt constellation would need generation and radiator area on the scale of industrial infrastructure, plus batteries and distribution. The launch mass is visible. The thermal geometry is the harder picture.
Networking creates a second constraint. Training needs high-bandwidth, low-latency collective communication with synchronized failures. Laser links across moving platforms face acquisition, pointing, weather at gateways, and changing topology. Embarrassingly parallel inference or preprocessing may tolerate distribution better than one tightly coupled training run. 'Data center' hides radically different workloads.
The million-satellite filing should therefore be read as an option claim, not a construction plan. Spectrum coordination, collision probability, debris, launch cadence, atmospheric impact, servicing, and deorbit scale nonlinearly. A credible path would demonstrate a small cluster with measured joules per useful token—including cooling, links, and replacement—and identify the workload whose Earth alternative is actually more expensive.
A million orbital data centers collide first with heat. In vacuum, waste heat leaves through radiation, proportional to area, emissivity, and the fourth power of temperature. Computing a megawatt and rejecting nearly a megawatt at electronics-safe temperature demands enormous radiator area. Raising temperature shrinks area but harms chips and power conversion.
Power collection adds another large surface and orientation constraint. Solar arrays, batteries for eclipse, structures, shielding, communications, and station-keeping all consume mass. Launch cost may fall, but deployment, collision avoidance, repair, and replenishment remain. Beaming data to Earth saves fiber, not thermodynamics.
The strongest niche would exploit a property unavailable on Earth: direct space-sensor processing, extreme security separation, or power and cooling economics that actually close. Proposals should publish watts of useful compute per launched kilogram, radiator assumptions, duty cycle, and replacement. Counting centers before closing one rack's heat balance is fiction with precise nouns.
Scientific computing
Teton gives a nuclear laboratory a faster uncertainty engine
Idaho National Laboratory detailed its Teton supercomputer on January 29, a 1,024-node system that roughly quadrupled the laboratory's computing capacity. The machine supports reactor physics, materials, grid, and energy research. Its useful output is not floating-point operations. It is a narrower and faster uncertainty range around experiments that are costly or impossible to repeat physically.
Nuclear simulation couples scales. Neutron transport determines where fissions deposit heat; thermal hydraulics moves that heat through fuel and coolant; structural models predict stress and deformation; materials evolve under irradiation. Fully resolving every particle and grain is impossible, so codes use discretization, reduced models, and empirical closures.
High-performance computing increases resolution and ensemble size. Engineers can sample uncertain cross sections, boundary conditions, manufacturing tolerances, and failures. A single detailed run produces a scenario; an ensemble produces a distribution useful for safety. Parallel efficiency matters because communication between subdomains can dominate as more nodes join.
AI can accelerate surrogate models or detect defects in instrument data, but it has to respect conservation and known bounds. A learned surrogate should report where training data are sparse and be checked against higher-fidelity codes. Faster wrong physics only tightens confidence around an error.
Teton's institutional value is shared capacity with expert support and validated software. Laboratories preserve benchmark problems, experimental comparisons, and configuration history across machine generations. The supercomputer will become infrastructure when researchers can reproduce a conclusion after the hardware itself has been replaced.
Nuclear simulation couples physics with different time and length scales: neutronics, fuel behavior, heat transfer, structural stress, coolant flow, and chemistry. MOOSE organizes these models around shared numerical infrastructure, while Teton-class compute expands mesh resolution, ensembles, and parameter sweeps. More FLOPs are valuable only if solvers converge and uncertainty is carried across coupled modules.
Verification and validation remain separate. Verification asks whether code solves the equations correctly; validation asks whether those equations reproduce experiments. Faster hardware can multiply a modeling error with impressive resolution. Regression suites, manufactured solutions, benchmark experiments, and versioned material data have to scale alongside the machine.
The practical payoff may be decision speed. Engineers can explore accident sequences, manufacturing tolerances, and operating envelopes before committing to expensive tests. But a regulator needs reproducible inputs and uncertainty, not only a rendered temperature field. The computer becomes nuclear infrastructure when it shortens design loops without making evidence harder to audit.
Teton accelerates uncertainty analysis by approximating expensive nuclear simulations, but a surrogate's confidence is meaningful only inside validated regimes. Reactor states near transients, unusual material conditions, or geometry changes may sit outside training. The system should detect extrapolation and route those cases to high-fidelity physics rather than produce a fast confident value.
Uncertainty has several sources: uncertain inputs, numerical approximation, model-form error, and surrogate error. Combining them into one interval can hide which experiment would reduce it. Calibration against withheld simulations and physical measurements is essential, with coverage tested across operating regimes rather than averaged globally.
The practical value is more design and safety scenarios per engineer-day. Teton can screen candidates, guide sensitivity analysis, and prioritize full calculations while preserving traceability to accepted codes. A regulator will care less that inference is fast than that every surrogate result carries domain, version, error bound, and a reproducible fallback.
Satellite manufacturing
Terran Orbital sells Mitsubishi a platform and an integration boundary
Terran Orbital announced on January 28 that it would provide a Nebula satellite bus for a Mitsubishi Electric low-Earth-orbit demonstration. A commercial bus packages structure, power, thermal control, avionics, attitude control, communications, and propulsion so the customer can focus on its payload. Standardization is valuable only when mission-specific requirements do not consume the standard.
Payload accommodation begins with budgets: mass, center of gravity, watts by mode, heat dissipation, pointing, data volume, electromagnetic emissions, and launch loads. These quantities interact. A high-power payload needs larger arrays and radiators, which change attitude dynamics and structure. The bus owns the coupled margins.
The interface control document is a living contract. Connectors and dimensions are the simple portion; fault response, timing, coordinate frames, software commands, and test responsibilities create most late surprises. Hosted demonstrations benefit from hardware-in-the-loop testing with real payload electronics before shipment.
Terran Orbital, now part of Lockheed Martin, can draw on production systems and mission assurance while serving a commercial Japanese customer. Cross-border programs add export controls and supply-chain coordination. The payload must receive enough design visibility to trust the bus without learning every proprietary detail.
A platform business compounds when anomalies and improvements transfer across missions. Custom work should become a new option or interface rather than a permanent fork. The Mitsubishi mission is small in orbit and strategically useful on the factory floor: it tests whether Nebula is a product or merely the starting point for another bespoke spacecraft.
A hosted bus sells boundaries: payload mass, volume, power, thermal rejection, pointing, data, orbit, and contact schedule. Mitsubishi can focus on its mission if Terran Orbital makes those numbers dependable. The bus provider then owns integration risk when a payload's peak power, thermal transient, or electromagnetic emissions violate assumptions that a one-line interface limit concealed.
Standard buses create production learning. Repeated structures and avionics can share tests, spares, supplier lots, and flight software. Yet mission customization accumulates quietly through harnesses, apertures, propulsion, and ground interfaces. Configuration control must preserve what is common enough to deserve heritage and what changed enough to require new qualification.
Hosted-service economics reward schedule credibility. A payload often decays in value while waiting for a custom satellite, especially in Earth observation or communications. If a platform shortens delivery and provides known operations, it can charge for time saved. The second and third missions will show whether Nebula is a product line or merely a reusable name on bespoke engineering.
Terran Orbital selling Mitsubishi a platform defines who owns interfaces and mission risk. A standard bus can shorten schedule if payload mass, power, thermal, pointing, and data stay inside validated envelopes. Custom requests should be priced visibly; otherwise the platform accumulates exceptions until every spacecraft is bespoke.
Integration begins with coupled budgets. A payload's heat affects radiator size, its pointing affects solar generation, and its data affects storage and downlink. Joint tests need hardware and flight-software timing. A clean contractual boundary cannot replace engineering feedback across it.
The strategic value for Mitsubishi is a repeatable supply chain and perhaps sovereignty over missions; for Terran, volume and production learning. Report unit cycle, first-pass yield, on-orbit checkout, and change orders. A platform is proven by the second customer receiving the promised schedule without inheriting the first customer's hidden design.
Planetary robotics
Claude writes rover commands inside a very human safety case
Anthropic described on January 30 how Claude assisted NASA's Perseverance team in planning a roughly 400-meter Mars drive. The model analyzed terrain and mission materials, then generated commands in Rover Markup Language. Engineers reviewed the plan before transmission. The milestone was not an autonomous language model driving Mars; it was a constrained coding assistant entering a mature planetary-operations workflow.
Rover command sequences encode waypoints, timing, instrument state, and safeguards. Mars-Earth delay prevents joystick control, so Perseverance already executes onboard navigation within approved goals. Ground teams combine orbital maps, rover imagery, slope, wheel hazards, energy, thermal limits, and science priorities. A language model can reduce translation and search effort across these artifacts.
Generated commands are unusually verifiable. Parsers reject syntax; simulators exercise sequences; resource models check power and time; experienced operators inspect intent. The model should be denied authority to redefine constraints. Its value is proposing a candidate inside a test harness where failure is cheap.
Training-data leakage and stale documentation remain risks. Mission commands use specialized semantics where a plausible token can be dangerous. Retrieval should bind the model to current flight rules and cite the source for each non-obvious choice. Logs need to preserve prompts, model version, output, edits, and approvals for anomaly review.
The collaboration is a strong template for high-consequence AI: choose a task with expensive human translation, machine-checkable structure, simulation, and a responsible expert. Claude did not replace the flight team. It made a portion of the team's reasoning inspectable in a new tool.
Mars command generation is a constrained compilation problem. Scientists express an observation goal; planners convert it into rover activities with duration, power, temperature, data volume, instrument state, and geometric dependencies. Commands are checked, simulated, and packaged for an uplink window. A language model can assist with the translation, but it should never erase those typed resource constraints.
The difficult cases are not syntax. They are conflicts: a drive leaves insufficient energy for overnight heaters, an arm pose blocks an instrument, terrain uncertainty invalidates a planned placement, or a downlink cannot return all data. The model needs authoritative state and must expose the chain from intent to command. A plausible plan that refers to stale rover state is dangerous precisely because it reads well.
Claude's contribution should be measured in planner time and caught errors across real sequences, with every output passing existing deterministic checks. The long-term opportunity is a copilot that explains tradeoffs and searches alternatives while humans retain release authority. Space operations are a good agent benchmark because the world state is rich, permissions are strict, and mistakes cannot be patched interactively from 225 million kilometers away.
Claude writing rover commands sits inside a safety case where humans define intent and constraints. The model can translate science goals into candidate sequences, check documentation, and identify conflicts. Deterministic planning tools should validate energy, thermal, communication, geometry, and command syntax before anything reaches the spacecraft.
Mars operations have delayed feedback. A mistaken command may not be understood until the next communication cycle, and the rover cannot be physically reset. Simulation, digital twins, and conservative state estimation are therefore part of authoring. The model should cite which telemetry and rule supports each precondition and abstain when state is stale.
Productivity is measured by valid sequences per planner-hour and errors caught before uplink. Human approval remains a technical control, not ceremonial liability. The best system makes experts faster at exploring alternatives while leaving an auditable chain from science request through checked command. Language fluency earns no direct authority over a distant robot.
The week's compute systems all sat behind an adapter: radiators and optical links, validated multiphysics codes, a satellite bus, or a rover test harness. The adapter determines which outputs can touch reality.
Software is easiest to trust when the physical system can reject it cheaply. A thermal limit, conserved quantity, interface budget, or simulator acts as an external critic. Without that critic, more computation can manufacture certainty faster than evidence.
This week put intelligence in orbit, a laboratory, a bus, and a rover workflow. The valuable part was the boundary that said no.
Moving computation changes its support system. Orbit needs power and radiators, simulation needs validation, a hosted payload needs contractual interfaces, and a rover agent needs resource-safe compilation.
Week 77 made compute feel abundant. The more revealing question was which scarce physical or institutional resource each additional operation consumed.