Blue Origin

Background

A semester-long team recommendation written for Blue Origin, out of a technology strategy course at Berkeley. Blue Origin owns the launch vehicle, the orbital transfer stage, and the destination, which makes it one of very few companies that could hold orbital compute end to end instead of one slice of it. We spent the term on where that integration is worth the most, and came back with orbital data centers.

The premise is that AI compute has run into limits that are physical rather than financial. Terrestrial data centers are rationed by grid interconnects and by water for cooling. Orbit has uninterrupted sunlight and a very cold sink to radiate into. Analysts put orbital compute at $1.77B in 2029 and near $39B by 2035, which places it just before the steep part of the adoption curve rather than after it.

We recommended entering as an infrastructure provider rather than a cloud: compute modules in low Earth orbit, archival storage on the lunar surface, sold to hyperscalers and government.

Choosing the orbit

Most of the architecture follows from one number: we put the shell at roughly 1,000 miles, about 1,600 km. Lower is cheaper to reach and worse to stay in. The ISS orbits near 250 miles, low enough that residual atmosphere drags on it and the station has to be reboosted. At 1,600 km drag is negligible, so a pod holds its slot for years on very little propellant, and fuel not spent station-keeping is mass spent on compute instead. Orbital velocity there is also lower, so New Glenn keeps about 70% of its nominal payload.

Going higher costs latency and radiation. Round trip stays between 3 and 40 ms for most populated regions, inside the budget for inference workloads, and each cluster holds line of sight to roughly a quarter of Earth's surface at once. The real price is that this altitude sits in the inner trapped radiation belt, a harsher environment than ISS altitude, which is why the compute design spends most of its complexity on tolerating errors rather than preventing them.

The constellation

Ten clusters across five orbital planes, up to twelve pods each, with pods inside a cluster joined by a laser optical mesh so no single node partitions the cluster when it fails and traffic can route around congestion. Every pod carries ion propulsion for station-keeping and collision avoidance.

The five inclinations are picked for coverage rather than symmetry: 28.5° for equatorial regions, 43° and 51.6° across the northern population centers, 66° for high latitudes, and a 97.8° sun-synchronous orbit for global reach.

Deployment rides entirely on New Glenn. At a conservative 30 metric tons of usable payload that is one pod per launch: twenty foundational missions to establish the mesh before 2028, then two launches a month to reach all ten clusters by 2030, split between Cape Canaveral for the mid-inclination planes and Vandenberg for polar. We carried 10% payload redundancy in the manifest so that losing a launch does not move the date.

Keeping the hardware alive

Space is an excellent heat sink and a terrible coolant. There is no air to move, so heat has to be conducted out and then radiated away. Each compute tile mounts to a coldplate and connects through loop heat pipes to deployable radiators, which is what high-power spacecraft already do and scales with payload power. Power is multi-junction gallium arsenide with lithium-ion sized for eclipse, and surfaces are conductively coated and bonded to a common potential, since the plasma up there will otherwise charge an isolated structure until it discharges through something that mattered.

For the computers we argued against radiation-hardened parts. Rad-hard silicon runs generations behind commercial, and the entire business is selling current-generation compute, so hardening the chips would undercut the product. The recommendation is dense commercial hardware with tolerance layered around it: ECC memory, periodic scrubbing of RAM and flash so single-bit flips cannot accumulate, watchdog recovery, and triple-modular redundancy applied selectively in reconfigurable logic. NASA's Spaceborne Computer on the ISS is the existence proof that commercial parts survive up there when the system around them is built expecting bit flips.

Downlink is optical, where recent missions have demonstrated over 100 Gbps, with Ka-band RF as the fallback for weather that closes an optical link.

Colocation, not cloud

The strategy call matters more than the architecture. Blue Origin should not sell cloud services. Competing with AWS and Azure at the software layer would put it against their core competence with none of its own, and would tie a launch company's returns to a hardware refresh cycle it does not control.

Colocation instead. Blue Origin provides the orbital environment, meaning power, thermal, structure, networking and servicing, and the customer brings their own compute. That is the Equinix model moved up 1,600 km. It sidesteps obsolescence risk, because when a customer's accelerators age out that is their capital expense, and it turns revenue into long-term leases rather than usage.

Servicing is what makes those leases defensible, and it is the part nobody else can offer. Blue Ring handles rendezvous, refueling and repositioning, and Orbital Reef acts as a human-tended hub for anything needing hands, which together support a hardware refresh every five to six years and a service life past fifteen. We also recommended buying the ruggedized compute layer rather than building it, with Ramon.Space and OrbitsEdge as targets, keeping Blue Origin's scope on heavy lift and servicing.

The model ramps from a 2 MW pilot in 2026 to roughly 100 MW by 2030, utilization climbing from 50% to 90%, at a blended price near $25M per megawatt-year. The base case is about $2.3B of revenue and $1.1B of free cash flow in 2030, and the conservative case still clears $400M.