The signal was two sentences long
In the first week of November 2025, Google Research published work describing satellites that carry tensor processing units into sun-synchronous orbit. The document contained numbers. Within seventy-two hours, a crypto news aggregator had compressed the whole thing into two sentences and a headline — Google plans to test an AI data center in space, and this will revolutionize energy efficiency and change global computing infrastructure.
The second sentence had no author attached to it. That is the tell.
Where liquidity hides, narrative finds its voice, and in this instance the voice arrived roughly two weeks ahead of the physics. Tokens with "compute," "orbit," and "space" in their names moved double digits on a paragraph containing no launch date, no power budget, no radiator area, and no named collaborator. I have watched this reflex long enough to recognize it instantly, and been wrong about it often enough to stay careful. The headline was noise. The paper underneath it was not — and the distance between those two things is where the actual trade sits.
Everything the headline deleted
The object behind the story is almost certainly Project Suncatcher, the moonshot Google Research put its name to in November 2025. Its facts are narrow and testable. Two prototype satellites, built with Planet Labs, targeted for early 2027. Sun-synchronous orbit at roughly 600 to 650 kilometers. TPU v6e accelerators that Google has already put through radiation testing, with results suggesting they can absorb a mission dose consistent with about five years on orbit. Photovoltaic generation in that regime running near eight times the yield of an equivalent panel area on the ground, because there is no night, no cloud, and no atmosphere to scatter the light. Heat rejection by infrared radiation into deep space, because there is no air to convect into.
That is the entire content of the announcement. No API. No SLA. No pricing page, no customer. Google's own stated commercial logic places the crossover point — where orbital compute becomes cheaper per unit of work than terrestrial compute — somewhere in the mid-2030s, contingent on launch costs continuing their descent and terrestrial electricity costs continuing their climb.
The aggregator's version kept a satellite and a promise and discarded everything else. No Planet Labs. No 2027. No mid-2030s. No radiation data. No admission that this is a research prototype, a feasibility check rather than a deployment. What remained was a headline-shaped object, which is precisely the shape markets price most efficiently and verify least.
Reading the silence between the blockchain blocks is a skill this industry claims to have and almost never practices. The deficit that hollowed out that news item — a fact stripped of every parameter that would let you test it — shows up in nearly every token whitepaper I have read since 2020. The difference is that in crypto the missing parameters are the product, while in an arXiv paper they are merely pending.
Why an energy-constrained hyperscaler looks up
I built my first liquidity heatmap in 2017, in a rented room in Chiang Mai, three weeks into a Python simulation of Uniswap's constant-product formula that I had written to model slippage during a Binance listing surge. That exercise taught me nothing about predicting price. It taught me that capital flow has a structure, and that the structure is usually legible before the price is. The same instinct applies here. Before asking whether orbital compute works, you ask what would make a company with a two-trillion-dollar market capitalization and a hundred-billion-dollar annual capital budget look upward.
The answer is that the binding constraint on AI expansion stopped being silicon more than two years ago. It is power, land, water, and the interconnection queue — and those four are not independent variables. A hyperscale campus drawing 100 megawatts needs a substation, a transmission upgrade, and a utility willing to prioritize it against residential and industrial load. In several large markets the queue for that prioritization is measured in years, not quarters. Evaporative cooling consumes millions of gallons in regions already rationing water. Local opposition has hardened from NIMBY noise into a genuine political constraint in Ireland, the Netherlands, parts of Virginia, and increasingly across Southeast Asia.
The illusion of control in a fluid world is the belief that a ten-year capital plan can be executed inside four-year political cycles. Google is not alone in discovering this. It is simply the first to write it down in a research blog. Microsoft signed power purchase agreements with nuclear operators. Amazon bought into small modular reactors. Meta went after geothermal and gas turbines. xAI put turbines behind a fence in Memphis and accepted the air-quality fight. Every one of those is a hedge against the same thing: the grid stopped being a utility and became a competitor.
Seen through that lens, Project Suncatcher is not a data center announcement. It is a strategic option purchased against the possibility that terrestrial energy never gets cheap again. And options, by construction, are priced for the possibility rather than the probability.
The arithmetic that decides the outcome
This is where the crypto coverage and the physics diverge completely, and where I would rather do the math than repeat the adjectives.
Heat rejection in vacuum follows the Stefan-Boltzmann relation: radiated power scales with the fourth power of absolute temperature and linearly with radiating area. At a radiator temperature near 300 Kelvin, a double-sided surface with high emissivity rejects on the order of 800 watts per square meter. Move one hundred kilowatts of IT load, then add the waste heat from power conversion and conditioning, and you are already north of two hundred square meters of panel before counting the structure that carries it. Push toward a megawatt and the required area crosses into the thousands of square meters.
Generation is constrained in the same geometry. The solar constant at Earth's orbit is about 1,360 watts per square meter. At realistic cell efficiency you collect a few hundred watts per square meter of array. A megawatt-class orbital platform therefore needs thousands of square meters of photovoltaic surface — roughly the scale of the International Space Station's arrays, which is a useful anchor, because it tells you what such an object costs, weighs, and how rarely anyone builds one.
A single satellite is not a data center. It is, at generous sizing, a few hundred kilowatts of IT load. A mid-size terrestrial hyperscale campus runs 50 to 150 megawatts. The ratio between those figures is two to three orders of magnitude, and no quantity of orbital enthusiasm compresses it. The real object Google describes is a constellation — hundreds of platforms launched and replaced on a rolling basis — and what reached crypto readers was a satellite.
Interconnect is the harder problem. Distributed training of large models depends on dense gradient synchronization: All-Reduce operations that move the same tensors across every worker at high frequency. Inside a modern training node, accelerator-to-accelerator fabrics move hundreds of gigabytes per second at microsecond latency. Between nodes, the best terrestrial fabric available moves 400 to 800 gigabits per second. The stated target for inter-satellite optical links is roughly 100 gigabits per second, with public demonstrations clustered an order of magnitude below even that. The deficit between an orbital constellation and a terrestrial cluster is not a tuning problem. It is a number, and the number is roughly five to ten times against the inter-node fabric and four to five orders of magnitude against the intra-node fabric.
You can run inference on that. You can run embarrassingly parallel workloads on that. You cannot run frontier-scale pre-training on that without a synchronization strategy nobody has yet published in convincing form.
Radiation tolerance is the one constraint where Google holds data rather than a plan. The TPU v6e result supports a five-year mission life, and that is real. It also carries a hidden cost: radiation-hardened design typically trades silicon area, clock headroom, and yield for survivability, so effective compute per watt on orbit is not compute per watt in a data hall in Iowa.
Underneath all of it sits launch cost. Falcon 9 pricing has hovered in the low thousands of dollars per kilogram to low Earth orbit. The Starship-era figures that make orbital arithmetic close are aspirational, not contracted. Google's mid-2030s crossover estimate is functionally a call option on someone else's rocket. That is a strategic vulnerability worth naming plainly: if cheap, high-cadence launch concentrates in a single provider, then orbital compute concentrates its supply chain in a counterparty that is also a competitor's affiliate.
The governance blank spot nobody is pricing
There is a layer here the crypto coverage cannot see and the mainstream coverage has not reached. A constellation of compute platforms worsens conjunction risk, and the debris cascade scenario has stopped being a thought experiment and become an actuarial question. Astronomers already fought Starlink over optical and radio interference; a fleet of satellites whose entire purpose is to reject waste heat into the sky will draw the same opposition with less public sympathy. The Outer Space Treaty of 1967 says nothing useful about jurisdiction over a rack of accelerators owned by a Delaware corporation orbiting above Indonesia. The ITU coordinates spectrum, not computation. Orbital compute introduces a sovereignty question no existing body has standing to answer: whose data protection law governs a tensor that never touches soil? Markets are currently pricing this at zero, which is the correct price only if the project never scales.
Where this actually touches crypto
Here is what I find instructive. In 2020, during the DeFi Summer, I was part of a small DAO building a cross-chain bridge aggregator. I wrote the first interface to the smart contract and spent evenings tracing Curve's emission mechanics. When the bridge was exploited, I did what an analyst does and pivoted — not to fixing code, but to modeling the governance token's volatility. That failure taught me the most durable thing I know: yield is a function of liquidity incentives far more than it is a function of protocol utility. Everything I have mapped since — the correlation between TVL inflows and token price elasticity, the emission schedules that convert a subsidy into a price — descends from that lesson.
Apply it to compute. The tokens that moved on Google's two-sentence headline are not building orbital anything. In the overwhelming majority of cases they are terrestrial hardware networks with a subsidy structure bolted on top: node operators paid in emissions, not in revenue, with unit economics that invert the moment the incentive curve steepens. Chasing ghosts in the algorithmic machine, the market bought the most physically implausible version of the story and ignored the version that actually has a paper behind it. That is not a coincidence; it is a selection effect. Real research has parameters, and parameters can be wrong. A tweet has none, and therefore cannot be falsified until the price says so.
In 2021, while coordinating a campaign for a mid-tier NFT project, I built a dashboard tracking USDT supply changes against OpenSea volume and found a roughly fourteen-day lag in the market's response. That lag was not sentiment. It was settlement — new stablecoin units entering circulation before finding their way into risk. The same lag runs through the compute complex today. When stablecoin net issuance contracts, narrative-driven hardware tokens bleed before the narrative adjusts. In a bear market that lag is the difference between exiting at a price and exiting at a mark.
Which brings the question to the only one that matters right now. Not which protocol wins. Which one is bleeding. The honest answer, for most of the DePIN and distributed-compute cohort, is: look at what fraction of node operator compensation comes from token issuance versus external paying demand. If the answer is above ninety percent, the network is not a network, it is a transfer. Volatility is just information wearing a mask, and in a drawdown it stops wearing it.
The NVIDIA angle deserves a line too. Google's vertical integration — custom silicon, custom interconnect, now potentially custom orbit — has been quietly eroding accelerator dependency for years. If orbital compute ever becomes real, it is marginally bearish for merchant silicon demand and marginally bullish for anyone selling radiation-tolerant parts, spatial thermal control, or optical terminals. None of those beneficiaries are the tokens that ran on the headline.
The contrarian read: relocation, not decoupling
The consensus version of this story is that compute is about to decouple from geography, which is why the space narrative and the crypto narrative fused so quickly. That framing is wrong in a specific and useful way. Orbital compute does not remove the constraint. It relocates it — from electricity, land, and water to radiator area, launch cadence, spectrum, and orbital governance. Every one of those new constraints is harder to price, harder to hedge, and governed by institutions that do not answer to shareholders.
And there is a second-order effect almost nobody has raised: if orbital compute works at any meaningful scale, it is bearish for the terrestrial distributed-compute tokens currently selling scarcity. Their entire premium rests on the premise that compute is locally scarce and must be aggregated from fragmented supply. A constellation of Google-owned TPUs above the atmosphere is the least fragmented compute supply imaginable. The narrative that lifted these tokens is the same narrative that, taken seriously, threatens them.
Meanwhile, the more defensible contrarian position is that the interesting thing Google did in November 2025 was not the satellite. It was the signal to utilities, regulators, and capital markets that the grid is now a strategic risk rather than a service. That signal is being read, correctly, inside every hyperscaler planning office. It is being misread, expensively, inside every trading desk that turned it into a ticker.
Takeaway
The testable milestones are not far away and they are specific. Do the two prototype satellites launch in early 2027. Does Google publish single-platform power, thermal rejection area, and compute parameters. Does anyone demonstrate inter-satellite synchronization at a scale that looks like training rather than a file transfer. Does the launch cost curve actually cross the power price curve, or does the crossover slide another five years into the future the way every moonshot timeline does.
Everything else is noise wearing a chart. And the question I would leave on the table is not whether computing escapes the planet. It is whether, by the time the first two prototypes reach orbit, any of the tokens that traded on their announcement still have a bid — or whether the market will have already rotated to the next two sentences, leaving behind the wreckage of a story it never bothered to read past the headline.
