Satellite InferenceAI Inference in OrbitGet in touch DESIGN STUDY / A ONE-MEGAWATT CLUSTER
We flew 1,080 one-megawatt clusters through the same full-force simulation as our formation page: ten altitudes, nine node powers, four spacings and three formations, each with the shielding its altitude needs. Every candidate is checked against ten requirements, from collision safety to the disposal rules. Below is the cheapest design that meets them all, and what breaks when one choice changes.
Change one thing
12 nodes of 100 kW at 650 km in a lattice tilted 18° out of the orbit plane, 142 m apart: every requirement met, at $11.93 a GPU-hour in 2030.
| Requirement | Our design |
|---|---|
| Neighbours at least five node widths apartat least 142 m | 142 m |
| Every pair kept apart radially and cross-trackat least the 85.2 m keep-out distance | 93.7 m |
| A dead node can be avoided in timethe neighbour can move once the failure is noticed | 89.4 m, no dodge |
| Neighbours shade the arraysunder 1% on the worst day | 0.00% |
| Twice the five-year dose, memoryunder the 2 krad where tests saw irregularities | 1.99 krad |
| A dead node re-enters on its ownwithin 5 years (FCC and ESA) | 1.2 years |
| Starlink and Kuiper shellsstay 10 km clear | 650 km: clear |
| Atomic oxygen on a 25 µm filmat most 25 µm in five years | 17 µm |
| Spare nodes2 for a full megawatt after five years, 95% sure | 2 |
| GPU-hour in 2030lowest that meets every requirement | $11.93 |
| Propellant over five years | 173 m/s |
| Narrowest cut through the network | 57.6 Tb/s |
01 / OUR DESIGN
Each node carries 100 kW of compute on 655 m² of arrays and 151 m² of radiators, 3,440 kg dry and 28 m across. In Earth's shadow it keeps 40% of its compute running on its battery, whose heat keeps the coolant warm, and its radiators are 15% larger than average conditions need (power and heat). If nodes stop manoeuvring at 1% a year and 1.5% are lost in commissioning, 2 spares keep a full megawatt to the end of five years with 95% confidence.
Dawn–dusk and sun-synchronous at 650 km, in sunlight 96% of the year. Starlink flies from 480 to 570 km and Kuiper is authorised from 590 to 630 km. Below them, atomic oxygen rules the orbit out; above, every 50 km costs more shielding.
Nodes sit on a square lattice, 142 m apart at their closest, turned out of the orbit plane so that no pair is ever level in height and across the track at once: at least 94 m apart, beyond the 85 m keep-out distance, with nothing to steer. Holding the tilt against Earth's bulge costs the outermost nodes 12.0 m/s a year.
At 650 km the memory receives 199 rad a year behind 12 mm at solar minimum, the worse half of the cycle for trapped protons. Twice the five-year dose stays under the 2 krad at which high-bandwidth memory showed irregularities in Google's proton tests.
Drag 2.0 m/s a year on average and 5.2 at solar maximum, the tilt 12.0, collision avoidance 1.2, and 97 m/s at the end to lower the orbit: 41 kg of propellant per node with electric thrusters.
On the central launch path with mass-produced satellites, 5.9 times a US data centre's $2.04. Unlike our economics page, this counts the spare nodes, the fixed mass of every node, the memory's shielding and the propellant. Replacing failed nodes with later launches instead of carrying spares would lower it.
02 / ALTITUDE
At 450 km a dead node falls within weeks and drag is cheap to cancel, but in five years atomic oxygen erodes 228 µm of polymer film facing the direction of flight, 9 times a standard 25 µm film, and drag at solar maximum costs 64 m/s a year.
Starlink and Kuiper fill these shells with thousands of satellites. At 500 km ESA counts 29 close approaches a year with a collision probability above one in a million for a satellite of 6 m², about three times as many as at 600 or 700 km (8.5 and 10.3).
Every 50 km higher costs shielding: the memory needs 20 mm of aluminium at 700 km and 40 mm at 800 km. At 800 km a dead node stays up 8.7 years, past the five-year rule of the FCC and ESA, and debris larger than 1 cm hits 2.3 times as often as at 650 km.
03 / NODE SIZE AND FORMATION
Node power in kW; under it, working and spare nodes for 1 MW. Small nodes repeat 300 kg of fixed mass each (33 kg per kW at 10 kW, 3.6 at 100 kW). Large nodes need spares as large as themselves. The steps come from whole numbers of nodes.
Share of one orbit along the bottom; for each formation, the neighbouring pair that comes closest. In the flat lattice a pair is level radially and cross-track twice an orbit, kept apart only along the track; tilting the lattice 18° keeps every pair at least 93.7 m apart in those two directions.
In a dawn–dusk orbit the Sun shines almost along the normal to the orbit plane, so nodes in that plane never shade each other, and a free-fall lattice repeats itself every orbit. Seen by a camera that turns with the orbit, as on our formation page, the lattice breathes in its plane; seen without turning, as in Google's animation, the same motion looks like a dance.
In the flat lattice every neighbouring pair is level in height and across the track twice an orbit, kept apart only along the track. When a node dies it tumbles, its drag grows and it drifts along the track: in our simulation it comes within 3.2 m of a neighbour 4.0 hours later. The neighbour can move aside, but only once the failure is noticed.
Each node also swings across the track, 0.66 times its radial swing and in step with its motion along the track, which turns the lattice 18° out of the plane. Every pair is then separated radially and cross-track in turn, as in the TanDEM-X and PRISMA formations (D'Amico and Montenbruck, 2006). The Sun stays at least 58° above the orbit plane, so the tilted lattice never shades itself.
A helix, with every node on its own ring around the centre, is just as safe and needs 17 m/s less propellant over five years, but it is a chain: each node has two neighbours instead of four, the narrowest cut through the network carries 38.4 Tb/s instead of 57.6, and the cluster is 3.5 km long instead of 1.2. For a 100 kW node, the lattice costs the same per GPU-hour as the helix.
04 / LINKS AND OPERATIONS
Each link uses 10 cm telescopes at 1550 nm. Over a few hundred metres a telescope can be split into many beams: 16 at our longest link of 299 m, each carrying 24 wavelengths of 400 Gb/s as in the Suncatcher paper. That is 154 Tb/s a link, more than a node can power: 76.8 Tb/s a node with 5% of its power for 26 W transceivers. Safety and propellant set the spacing, not the lasers.
A megawatt of arrays and radiators is a large target. Scaling NASA's rates of approaches that need a manoeuvre by area, the cluster would make about 148 a year, 12 for each node; at 0.1 m/s each, the return to the slot included, that is 1.2 m/s a node a year. Debris larger than 1 cm would hit somewhere in the cluster about once every 3 years.
At solar minimum the upper atmosphere shrinks: a node that dies then could stay up 18 years instead of 1.2. The five-year rule is checked at the average of the solar cycle, as lifetime estimates usually are; a node that can still steer lowers its own orbit at the end.
05 / METHOD AND LIMITS
Every candidate flies in the cluster simulation of our formation page: each node propagated with Earth's gravity to J4, drag from NRLMSIS 2.1 along the dawn–dusk orbit and solar radiation pressure, from mean relative orbital elements matched over one revolution, on the day of the year when the Sun sits lowest over the orbit plane. The centre node is made to fail and tumble, and every close approach over the next twelve hours is measured. Nodes are sized with our power and heat model for 2030 technology, keeping 40% of their compute running in Earth's shadow and with 15% more radiator than average conditions need, plus 300 kg of fixed mass each and the shielding their altitude needs; propellant follows from the thrust budget at a specific impulse of 1,500 s. Cost uses the formulas of our economics page on its central path for 2030. Our design is the cheapest that meets every requirement; designs within 0.5% of its price count as equal, and among them the one with the widest cut through its network is chosen.
Radiation along the dawn–dusk orbit from 450 to 1,000 km uses the models of our radiation page (AP-8 and AE-8 through IRBEM, cosmic rays fitted to PAMELA and AMS-02, SHIELDOSE-2), recomputed at every altitude. Debris flux comes from NASA's ORDEM 3.2 for a 98.5° orbit, close approaches from ESA's Space Environment Report 2026 and NASA's conjunction assessment handbook, scaled by area. Atomic oxygen is NRLMSIS 2.1 along the same orbit, with the erosion yield of Kapton measured in flight. Links follow the Suncatcher paper's optics and transceivers.
Not modelled: the structure and pointing of 28 m panels; heavy ions; the exact debris and conjunction rates of a cluster, which are scaled from single satellites; atomic oxygen on surfaces that do not face the direction of flight; replacing failed nodes with later launches; and operations beyond the cost model's allowance. The failure rates are assumptions informed by Starlink's record and published spacecraft reliability, not measurements of our nodes. A design study for communication, not a mission design.
Sources: Google Research, Project Suncatcher, arXiv:2511.19468 v2 (2026); D. A. B. Miller, Applied Optics 39 (2000); S. D'Amico and O. Montenbruck, J. Guidance, Control and Dynamics 29 (2006); NASA ORDEM 3.2 via the NASA Conjunction Assessment Handbook Vol. 2 Rev. 2 (2026); ESA Space Environment Report 2026; FCC 22-74 (2022); ESA ESSB-ST-U-007 (2023); NRLMSIS 2.1 via pymsis; NASA AP-8 and AE-8 via IRBEM; SHIELDOSE-2; Banks et al., NASA/TM-2004-213223 (Kapton erosion yield); S. G. Turyshev, arXiv:2608.27499 (2026, fixed mass per node); J. McDowell, Starlink statistics (2026); Castet and Saleh, J. Spacecraft and Rockets (2009); FCC authorisation of the Kuiper system (2020). The SI-P1 pathfinder is a hosted payload; this study is about future clusters of SI Nodes.