Orbital Compute & Space Infrastructure

The path opens by interrogating its own premise, because this is a field where the promotional material substantially outruns the flown hardware. The arguments actually made are: the downlink is a bottleneck so process at the source; sunlight is continuous so power is free; vacuum is cold so cooling is free; and orbit is outside any single jurisdiction. Two of those are largely wrong as stated, one is conditional, and one is real. Establishing which is which up front is what stops the rest of the path becoming advocacy.

What is physically different about the place. Hard vacuum with no convection and outgassing from ordinary materials, thermal cycling driven by orbital day and night, charged particles and heavy ions, atomic oxygen in low orbit, plasma charging, micrometeoroid and debris flux, and the launch itself as the most violent mechanical event the hardware ever survives. Each of these eliminates design options that are free on the ground, and the following units take them one at a time.

🔴 The constraint that makes this a different engineering discipline rather than datacenter hardware in a different postcode. Accumulated dose gradually shifts device parameters until the part stops meeting spec; single particles flip bits, corrupt logic transiently, or trigger a latch-up that destroys the device if power is not removed. The distinction between recoverable and destructive events is the one that governs every architectural decision that follows, and it is the concept most software people have no prior model for.

The design response to A3, and the live argument in the field. Radiation-hardened parts are reliable, expensive and generations behind, which is fatal for a compute mission; commercial parts are fast and cheap and fail in orbit unless the system is built to expect it. Covers error-correcting memory, redundancy and voting, memory scrubbing, watchdogs, latch-up detection with power cycling, and checkpointing — the architecture that lets unreliable parts host a reliable system.

The unit that dismantles the 'space is cold, cooling is free' claim, which is the most common misconception in the whole subject. With no air, heat leaves only by conduction to a radiating surface and then by radiation, and radiator area scales with the heat you must reject. A high-power processor in orbit is therefore a thermal design problem long before it is a power one, and the honest conclusion — that thermal rejection, not silicon, is the practical ceiling on orbital compute density — has to be reached here rather than asserted.

Solar array sizing against orbital average rather than peak, eclipse fraction and the battery mass it forces, degradation over mission life, distribution and regulation, and duty cycling as the honest answer when continuous operation cannot be afforded. Read together with A5, this unit produces the number that actually bounds a mission: how much compute can be sustained, which is almost always far less than the marketing figure.

The economics that constrain the physics. Cost per kilogram to orbit and how much it has and has not fallen, the standard dispenser and rideshare form factors that hardware must fit, structural design for launch loads rather than operational ones, and the fact that radiators, shielding and batteries are all mass. Every kilogram of mitigation from the previous units is priced here.

🔴 The unit the whole thesis turns on. Available downlink is a function of link budget, contact time and spectrum — and for a low-orbit satellite talking to a handful of ground stations it is far smaller than the data a modern sensor generates. That gap is the honest case for onboard processing. Covers radio and optical downlink, inter-satellite links, and the quantitative comparison of transmitting raw data versus spending power and thermal budget to reduce it first.

Having established the constraint, the applications that actually clear it: cloud screening so unusable imagery is never transmitted, event detection and alerting where latency matters more than fidelity, compression and feature extraction, and sensor fusion across a constellation. Taught with the discipline of stating what each saves and what it costs in power and thermal, because that ratio is the whole argument.

Operating a computer that cannot be physically touched for its whole life. Fault-tolerant boot and redundant images, a recovery path that survives a failed update, over-the-air updates over a link measured in kilobits and available for minutes a day, autonomy because the ground is not always reachable, and telemetry designed so a fault can be diagnosed from what was downlinked rather than from what you wish you had logged. This is the unit where terrestrial software instincts are most actively harmful.

Once there is more than one node, orbit becomes a network with unusual properties: links that are predictably intermittent rather than randomly lossy, long and variable delay, and topology that changes deterministically with the orbits. Covers inter-satellite links, delay-tolerant networking and store-and-forward, and constellation routing — and why ordinary terrestrial protocols behave badly under predictable disconnection.

The non-engineering constraints that are genuinely binding. Command authentication and encryption, supply chain assurance for parts that cannot be inspected after launch, the data-residency and jurisdiction claims made for orbital infrastructure and how well they survive contact with actual law, and export control — which restricts what may be built, discussed and shared, and is a real Risk element for anyone working in this area rather than a compliance footnote.

How anyone gains confidence in hardware that cannot be repaired. Derating, redundancy and cross-strapping, failure modes and effects analysis, parts screening, and the qualification campaign — vibration, thermal vacuum, radiation testing — that consumes much of a programme's schedule and budget. Placed late because it only makes sense once the failure mechanisms in A2 through A6 are understood as the things being qualified against.

The bracket that closes the path opened in A1. Who is actually flying compute today rather than announcing it, what has been demonstrated versus what has been claimed, the real cost per unit of work delivered, and where the near-term applications sit. Given how much of the public material on this subject is fundraising rather than engineering, the assessable skill is reading a claim and identifying what it did not say.

The capstone Skill element, because there is no certification to sit and no lab to run. Take a stated workload and produce a defensible design: power budget, thermal rejection area, radiation mitigation architecture, downlink budget, mass and launch envelope, and an honest conclusion about whether it is worth flying at all. The artifact is the argument, and 'this should not be in orbit' is a passing answer if it is properly supported.