Google’s Project Suncatcher Put 4 TPUs in Orbit. Launch Must Get 18x Cheaper Before Space Power Competes With Earth.

What went up

Project Suncatcher is a research moonshot at Google that asks whether AI compute could one day scale in space, where sunlight is nearly constant. Its first flight is a single prototype satellite built with Planet. It rode SpaceX’s Transporter-18 rideshare from Vandenberg on October 1, 2026, among about 130 payloads.

Over the coming weeks Google will watch how the chips handle radiation, temperature swings and launch stress in real conditions. Getting to orbit was already a test: Google says the ride lasts about 10 minutes at up to 10 g, and individual parts like a TPU chip can see 50 to 100 g.

Why put a data center in orbit at all

Energy. Google’s paper opens with the Sun’s output, 3.86 × 10²⁶ watts, more than 100 trillion times all the electricity humanity produces. In a dawn-dusk sun-synchronous orbit a satellite stays in sunlight almost all the time, with no night, no clouds and no atmosphere in the way. Google says panels there can collect up to 8 times the energy per year of a panel at mid-latitudes on Earth.

The long-term design in the paper is a cluster of 81 satellites flying within a 1 km radius at about 650 km altitude, with neighbors 100 to 200 meters apart. They would talk over free-space lasers at around 10 terabits per second per link, so the cluster works like one data center. Google’s bench demo hit 800 Gbps each way, 1.6 Tbps in total, with a single pair of transceivers.

The cost math in Google’s own paper

The interesting part of the paper is its honesty about money. It compares the cost of power in orbit, mostly the cost of launching the satellite that generates it, with what data centers pay for electricity on the ground.

At today’s launch price, the paper puts orbital power at about $14,700 per kW per year. That is roughly 5 to 26 times what a US data center pays ($570 to $3,000, from $0.06 to $0.25 per kWh). Only when launch falls to about $200 per kilogram does space reach $810 per kW-year and land inside the terrestrial range.

Getting from $3,600 to $200 is an 18x drop. The paper fits a learning curve of about 20% per doubling of launched mass and finds that reaching $200 per kilogram by about 2035 would take some 370,000 more tonnes in orbit, about 180 Starship launches a year. A small satellite buying a rideshare seat today pays even more, about $7,000 per kilogram.

Cold space doesn’t cool a chip

Space is cold, but a vacuum has no air or water to carry heat away. On Earth a data center moves heat with fans and liquid loops. In orbit the only way out is radiation from a hot surface, which is why Suncatcher uses heat pipes to move heat from the TPUs to radiators.

How much radiator does that take? A rough best case from the Stefan-Boltzmann law, our own physics rather than Google’s: a two-sided radiator at 27°C with emissivity 0.9, absorbing nothing from the Sun or the Earth, sheds about 830 watts per square meter.

Our math: area = heat ÷ (2 sides × 0.9 × σT⁴). Real radiators also absorb sunlight and Earth’s infrared, so they need more area than this.

A hotter radiator sheds heat much faster, because output grows with the fourth power of temperature. But chips can only run so hot. That is the trade-off the 15-minute compute bursts point to: per GCN, the TPUs run Gemini models until thermal limits force a cooldown. Google’s own paper calls thermal management “a critical optimization challenge” and gives no radiator numbers yet.

Radiation: better than expected

Before launch, Google put a Trillium TPU in a 67 MeV proton beam while it ran AI workloads. A shielded five-year mission in this orbit adds up to about 750 rad(Si) of total dose.

The high-bandwidth memory, the most sensitive part, showed its first irregularities at about 2.7 times the five-year dose. The chip itself took 20 times that dose without a hard failure. The softer problem is silent data corruption: the paper estimates about one event per 17 rad, or roughly one bad result per 3 million inferences at one inference per second. A small rate, but one that has to be designed around.

Fun facts

• The Sun puts out more than 100 trillion times humanity’s total electricity production.
• Google compares aiming the laser links between satellites to hitting a coin-sized target from miles away while both ends move.
• A Starlink v2 mini, the paper’s reference satellite, is estimated at 575 kg and about 28 kW, roughly 49 watts per kilogram by our math.
• At 100 reuses per Starship, the paper’s launch price drops below $15 per kilogram, 240 times below its current-price baseline.
• The same paper notes that Google cut the energy of a Gemini query 33x in one year. Efficiency on the ground is moving fast too.
• The prototype is designed to deorbit within about six years of launch rather than stay in orbit.

What this means for AI prices

Nothing changes this decade. On Google’s own numbers, orbital compute stays more expensive than ground compute until launch costs fall roughly 18x. The price you pay per token in 2026 and 2027 is set by data centers on Earth.

Power is the constraint everyone is chasing. A Google research team launching TPUs to collect sunlight shows how seriously labs take the energy limit on AI. The same pressure is why model makers keep cutting energy per query, and why cheap, efficient models keep getting better.

Watch the 2027 flight. The real test of Suncatcher is the pair of satellites planned for 2027 with laser links between them. A single satellite proves the chips survive. Two satellites sharing work at terabit speeds would prove the cluster idea.

FAQ

Where MegaBrain fits

Full disclosure: we sell an AI coding agent, and we don’t run anything in space. MegaBrain’s terminal coding agent and editor extensions run GLM, Kimi, MiniMax, DeepSeek, Qwen, Grok and Auto on the paid plans, Claude, GPT and Gemini paid from your credits from Pro, and any model through your own API key or local models, free on every plan. Whatever the next drop in compute costs brings, you can switch models without switching tools. See the plans and subscribe. MegaBrain is not affiliated with Google, Planet or SpaceX.

Sources

All pages opened October 27, 2026.

• Google on X — launch announcement with Planet on SpaceX Transporter-18, contact confirmed (October 1, 2026)
• Google — “Behind Project Suncatcher, our moonshot to put AI in space”, Travis Beals (September 24, 2026): launch loads, proton-beam tests, heat pipes and radiators, 2027 plan
• Google Research — “Exploring a space-based, scalable AI infrastructure system design” (November 4, 2025)
• Google — “Towards a future space-based, highly scalable AI infrastructure system design”, arXiv:2511.19468 (launch cost analysis, radiation results, 81-satellite cluster, optical links)
• GCN — “Google’s Project Suncatcher puts four AI chips into orbit aboard SpaceX Transporter-18” (October 4, 2026): TPU generation, power, burst length, mission life
• Runtime Wire — “Google puts four TPUs in orbit to test its space-compute plan”
• NewSpace Economy — “Satellite ridesharing market analysis 2026” (Transporter rideshare list price)