Two prototype satellites. Launch window: early 2027. Orbit: 600 to 650 kilometers. Payload: Google's own TPU v6e silicon, radiation-tested for a five-year mission life. Partner: Planet Labs. Project: Suncatcher.
That is the entire verifiable payload of the story that crossed my feed this week under a headline about Google testing AI data centers in space. The source was a crypto outlet. No author. No date. No orbital parameters. No link to the underlying research. Two of the four claims in the piece — "revolutionizing energy efficiency" and "transforming global compute infrastructure" — carried no attribution at all. They were weather, not information.
I run a verification protocol before I read anything. List the data sources. List the on-chain metrics. List the parameters. When a story arrives with none of the three, I file it as a signal, not a fact. Signals are useful. They just don't get a position size.
The underlying object is real, and it is not a data center. Google Research published Suncatcher in November 2025 as a moonshot-class investigation into placing TPUs in sun-synchronous orbit. The physics motivation is clean. A sun-synchronous orbit sees no night, no cloud cover, and no atmospheric attenuation; photovoltaic yield runs roughly eight times what an equivalent panel produces on the ground. That eliminates the two dominant operating cost lines of a terrestrial facility — electricity and cooling — plus the three that don't show up on the P&L until they do: land, water, and the interconnection queue.
That is a coherent thesis. It is also, right now, a paper.
I understand why a blockchain outlet ran it. The shape is familiar to anyone who has watched DePIN raise capital. Physical infrastructure. Enormous addressable market. No token yet. A validator set presumably arriving in a later post. Same template, no asset attached.
And I have a stake here that has nothing to do with satellites. I price yield. Every yield number I have ever computed sits on top of an energy number. Validator margins, mining hashprice, staking APR net of hardware depreciation — all of it traces back to watts. When the largest buyers of compute on earth start treating energy as a binding constraint, that is not a story about orbit. It is a story about the denominator of my book.
The physics wall is the radiator, not the rocket.
Run the numbers. Stefan-Boltzmann: radiated power scales with emissivity, area, and the fourth power of temperature. At 300 Kelvin, σT⁴ is about 459 watts per square meter. With an emissivity of 0.9 and a double-sided panel, you shed roughly 826 watts per square meter of radiator. One megawatt of waste heat therefore requires about 1,200 square meters of panel — roughly a third of an acre, per megawatt, just to avoid cooking the silicon.
Scale that to a training cluster worth talking about and the radiator area scales linearly with power. There is no Moore's Law for this. You can shrink the area by running the radiator hotter, since area falls with the fourth power of temperature — but that requires the electronics to survive at those temperatures, and they cannot. Emissivity coatings buy single-digit percentages. The constraint is thermodynamic and it does not negotiate.
Now compare launch cost. Dollars per kilogram have fallen on a visible, forecastable curve for a decade. Launch is a variable that improves on schedule. Radiator area is a variable that does not improve at all. Everyone is watching the rocket. The wall is the panel.
Interconnect is where the training case dies.
Google's published target for inter-satellite optical links is 100 Gbps. Current demonstrations sit near 10 Gbps. Inside a terrestrial facility, accelerator-to-accelerator interconnect runs hundreds of gigabytes per second per link, and the aggregate switching fabric runs into terabytes per second.
Take a 175-billion parameter model. One full gradient in bf16 is roughly 350 gigabytes. A ring all-reduce moves that payload approximately twice. At 100 Gbps, and assuming a perfect link with zero contention — generous — a single synchronization pass takes on the order of a minute. A terrestrial cluster completes the same operation in milliseconds.
That is three to four orders of magnitude, depending on whether you count per-link or aggregate. Training is not throughput-bound. It is synchronization-bound. You cannot close a gap that wide by adding satellites; adding nodes to a ring all-reduce makes latency worse, not better. The topology works against you.
Inference is a different conversation. Embarrassingly parallel, tolerant of latency, tolerant of jitter. If Suncatcher ever ships something commercial, watch for inference. Not training.
Radiation and the depreciation treadmill.
Google has radiation-tested TPU v6e and concluded it can absorb the dose profile of a 600–650 kilometer orbit across a five-year mission. That is real engineering and it is credible. But radiation-hardened design is paid for in performance or cost, usually both. And a five-year mission life means a constellation is a depreciating fleet with a replacement schedule attached — the opposite of the load profile a terrestrial facility offers: seven years of uninterrupted service with incremental refresh.
Anyone modeling orbital compute as "a data center that happens to be up" is modeling it wrong. It is a fleet with a mortality curve. Capex recurs.
Unit economics: there is nothing to discount.
Current stage: research. Pre-proof-of-concept. No API. No SLA. No pricing. No customers. Google's own framing places cost parity with terrestrial compute in the mid-2030s, contingent on launch costs continuing to fall while grid power costs continue to rise. Two contingent curves, ten years out, with no measurable milestone before 2027.
So there is no revenue model to stress-test. The project is not a product line. It is a hedge against Google's own energy bottleneck — a strategic option on the power curve, purchased with research budget. That is a coherent thing for a hyperscaler to own. It is not a thing you can underwrite.
This is where the crypto read gets interesting and where it gets lazy. Interesting: the real constraint is energy, and energy is the same constraint that sets mining hashprice, validator margins, and the sustainable floor under staking yield. If the largest compute buyers on earth are seriously modeling an escape from the grid, the political economy of every grid-constrained jurisdiction shifts — and crypto miners sit in exactly those jurisdictions, holding exactly those interconnection agreements. Lazy: inviting retail to buy "space compute" exposure that does not exist. There is no orbital asset. There is no token. There is a research blog and two satellites scheduled three years out. Efficiency is the only morality in the machine, and there is nothing here yet to measure.
The consensus read is that Google is betting on space. That is the wrong frame. Trace the capital. Google's actual energy dollars are going to nuclear power agreements, geothermal, and long-duration storage — terrestrial assets with signed offtake and delivery dates. Space is the narrative that makes the terrestrial ask look moderate. When a company publicly explores the most aggressive option available, it moves the Overton window for everything less aggressive. The orbital program is cheap. The cover it provides is not.
Second blind spot: the source. The story reached me through an outlet with no byline, publishing to a crypto audience about a project with no crypto component. That format has a shape. Physical infrastructure, no token yet, sovereign-scale ambition — that is the DePIN pitch template minus the token. The publishing decision was not editorial curiosity. It was inventory preparation. Somewhere right now someone is drafting a space-compute whitepaper with a validator set and a 40% APY. I have audited fifty of these. The exit is the position. Trust is a variable I no longer solve for.
Three numbers to watch, and only three. Per-satellite watts. Radiator area per satellite. Measured inter-satellite bandwidth under a training-like synchronization load. If Google publishes all three in a peer-reviewed follow-up, the program is real and the mid-2030s crossover deserves a model. If the next update is another blog post with a rendering and no parameters, it is a public relations instrument with a launch date attached.
The trade is not in orbit. The trade is in the grid — in who gets to consume power, at what price, and under whose jurisdiction. That question is already live. Space just gave it a longer deadline.
