Google is about to put TPUs into orbit as the first step toward solar-powered machine-learning clusters in space. The launch post is short. The engineering argument is in the paper behind it. This course lays out that argument: why the satellites must fly a few hundred metres apart, what the radiation tests showed, and how launch prices would have to fall. It separates what is demonstrated from what is projected, and names what the comparison with data centres leaves out.
Built on Towards a future space-based, highly scalable AI infrastructure system design (Google, arXiv v2, Jun 2026) and Behind Project Suncatcher (24 Sep 2026)
The authors’ own claim is modest and precise: the core concepts are “not precluded by fundamental physics or insurmountable economic barriers.” That is a feasibility argument, not a business case. The paper says of its cost comparison, “The below does not constitute a full economic analysis.” The course holds them to that framing in both directions. It does not dismiss a feasibility result for lacking economics, and it does not read a feasibility result as a forecast.
Short on time? Do Module 2 (why they fly so close) and Module 5 (load-bearing versus speculative).
If AI is a general-purpose technology, demand for compute and energy will keep growing. The Sun emits more than 100 trillion times humanity’s total electricity production. The paper argues that “at some point in the future, the best way to power AI will likely thus be to more directly tap into that power.” The practical hook is orbit. In the right low orbit, satellites see near-constant sunlight, “generating up to eight times more solar power than on Earth,” with little need for batteries.
No giant orbital data centre assembled in space. Instead a swarm of small satellites, each with solar arrays, TPUs and optical links. They fly in a dawn–dusk sun-synchronous low Earth orbit, which keeps them over the day–night line and in near-continuous sun. They fly close together so that laser links between them can match the bandwidth of the fibre inside a terrestrial TPU pod. Each satellite would eventually carry “dozens of TPU chips.”
| When | What was said |
|---|---|
| Nov 2025 (research blog + paper) | A learning mission with Planet, “slated to launch two prototype satellites by early 2027.” |
| Sep 2026 (launch post) | A first satellite on SpaceX’s Transporter-18 rideshare, built with Planet, to test TPUs in spaceflight, radiation and vacuum cooling. The laser links wait for two satellites in 2027. |
The sequencing is sensible: prove the chip survives before proving the chips can talk. But it means the defining capability of the design, high-bandwidth links between close-flying satellites, is not what launches now.
Large ML jobs are spread across many chips that exchange data constantly. Inside a terrestrial TPU pod, chips talk over optical links at hundreds of gigabits per second per chip. Commercial satellite-to-satellite optical links run at roughly 1–100 Gbps. The paper sets a target of about 10 Tbps per link, one to four orders of magnitude beyond what space lasers do today. The 2026 post puts it plainly: existing systems are “optimized for low bandwidth across large distances,” while these need “very high bandwidth over extremely short distances.”
Received optical power falls with the square of distance. Long-range space links make do with about a microwatt at the receiver, which limits them to a few simple channels. The commercial dense wavelength-division multiplexing (DWDM) transceivers used in data centres, which put many wavelengths on one fibre, need hundreds of microwatts. Fly the satellites close enough and there is enough light for data-centre optics to work in free space.
The paper’s illustrative cluster is 81 satellites in one plane, at a mean altitude of 650 km, with a cluster radius of 1 km. Neighbouring satellites oscillate between roughly 100 and 200 m apart over each orbit. That is far closer than any existing constellation flies. Modelling with the Hill–Clohessy–Wiltshire equations plus a differentiable model in JAX suggests the cluster’s shape stays predictable. The main perturbation, Earth’s oblateness, can largely be designed around, so formation flight should need only “modest” manoeuvring beyond ordinary station-keeping.
Closeness buys bandwidth with commodity optics, but it spends safety margin. A hundred satellites a hundred metres apart is a new operating regime. Precision navigation, collision avoidance and the behaviour of a cluster when one satellite fails all become central. The paper models the orbits. It does not yet demonstrate close-formation operations, and that is what 2027 is for.
Google tested Trillium (v6e) TPUs in a 67 MeV proton beam at UC Davis’s Crocker Nuclear Laboratory while they ran ML workloads. It reports this as the first published radiation test of such a device. With typical shielding, the expected dose in orbit is about 150 rad(Si) a year, so a five-year mission needs survival to about 750 rad(Si).
| Effect | Result | Against the requirement |
|---|---|---|
| Total dose: long-term damage | HBM memory showed irregularities from 2 krad(Si). Everything else ran correctly to 15 krad(Si). No hard failures. | HBM margin ~3×; logic margin 20× |
| Silent data corruption: single-particle bit-flips in logic and SRAM | About 1 event per 17 rad | On the order of 1 per 3 million inferences, at 1 inference a second |
| Functional interrupts: host crashes and reboots | About 1 per 450 rad (CPU) and 1 per 400 rad (RAM) | Less than one host reset per system per year |
The authors’ own judgement on silent corruption: the rate “is likely acceptable for inference, the impact of SEEs on training jobs, and the efficacy of system-level mitigations, requires further study.”
An inference is short and independent. A rare wrong answer affects one request. A training run is one enormous computation spread across thousands of chips for weeks. An undetected bit-flip can propagate into the weights and quietly damage the whole run. Per chip, 150 rad a year at one event per 17 rad is about nine silent errors a year. Multiply by the number of chips in a cluster and the count is no longer rare. Silent data corruption is already a known problem on the ground. Space raises the rate, and so raises the importance of the detection and correction the paper says still need study.
SpaceX’s price history, from Falcon 1 to Falcon Heavy, fits a learning rate of about 20%: the price per kilogram falls about 20% for every doubling of cumulative mass launched. If that continues, which “would require ~180 Starship launches/year,” launch to low Earth orbit could fall below $200/kg by about 2035. With about 72% less mass launched, the same curve gives about $300/kg. A bottom-up look at Starship’s published specifications points the same way: roughly $60/kg to SpaceX with 10× reuse, under $15/kg with 100× reuse, and a propellant floor near $8/kg. Today’s reference is $3,600/kg on a reusable Falcon 9.
To compare with a data centre, the paper asks what it costs to put a kilowatt of solar power in orbit, per year of life. For a Starlink v2 mini (575 kg, about 28 kW, 5-year life): at $3,600/kg that is about $14,700 per kW per year; at $200/kg about $810. Across other satellite designs, the $200/kg range is $810–7,500. US data-centre power costs about $570–3,000 per kW per year ($0.06–0.25/kWh, PUE 1.09–1.4). So at $200/kg, launched power could be “roughly comparable” to terrestrial energy spend.
Launched power price = launch price × mass ÷ kW ÷ life, as in the paper. Grid power = price × 8,760 h × PUE. Like the paper, this excludes chips, satellite manufacture and data-centre buildings.
| Claim | Status | Why |
|---|---|---|
| Up to 8× more solar energy in dawn–dusk orbit | Physics | Orbital geometry and insolation; not in dispute |
| TPUs survive a five-year dose | Tested | Proton-beam tests with margin; in-orbit data from the 2026 mission |
| Terabit free-space optics with commodity parts | Bench | 1.6 Tbps on the bench; not between moving spacecraft |
| Stable 100–200 m formation of 81 satellites | Modelled | Orbital-dynamics modelling; not flown |
| Silent errors acceptable for training | Open | Authors say it “requires further study” |
| Launch <$200/kg by ~2035 | Projection | Extrapolated learning curve; depends on one vendor’s volume |
| Cost-competitive with terrestrial compute | Not claimed | Power-only comparison; “not a full economic analysis” |
A communications satellite spends much of its power on radio and sends some of it away as signal. A compute satellite turns nearly all of its power into heat in a small area, and in vacuum “you can only diffuse heat via radiators.” Radiator area grows with power, which pushes up kg per kW. The launched-power arithmetic borrows a Starlink’s mass-to-power ratio, which may flatter a compute design.
“Currently, failed TPUs are manually replaced by technicians,” the authors note, which is “obviously impracticable in space.” Their answer is redundancy. Spare capacity is paid for at launch prices and carried for the whole mission.
The launched power price amortises over five years. On the ground, operators replace accelerators when newer chips are enough better per watt. In orbit you cannot refresh chips without launching new satellites, so obsolescence either shortens useful life, which raises the price per kW-year, or strands old compute in orbit.
Useful work needs inputs up and results down. The best demonstrated optical ground link is NASA’s TBIRD at 200 Gbps (2023). The paper names atmospheric turbulence and beam tracking as challenges. This favours workloads that are compute-heavy and data-light, like some training or batch inference, over interactive serving.
The post frames the first mission modestly: “This first launch is about seeing what works, identifying points of failure, and applying those findings to future missions.” Before launch, the hardware already survived vibration testing on all three axes, with loads of up to 10 g for the vehicle and 50–100 g for components like the chips. “Tests like this rarely go as planned, so we were pleasantly surprised that the hardware held up.”
For each open question, decide whether the 2026 single-satellite mission can largely answer it, whether it waits for the 2027 two-satellite mission, or whether neither can.