The co-founder of a GPU cloud company put his Series A in the same breath as the Marshall Plan and the Apollo Program this week, and I had to read it twice to make sure it wasn’t satire. Jamie Cox, Chief Strategy Officer at Fluidstack, called it “a very overdue announcement”: $830 million raised at a $7.5 billion valuation, led by Situational Awareness. Then came the part that made me put my coffee down. The United States, he wrote, must add hundreds of gigawatts of compute, a buildout “larger than the Arsenal of Democracy, the Marshall Plan, and the Apollo Program combined,” so that “democracy wins and every human on Earth benefits from transformatively powerful AI.” My DMs are open, he added; please reach out.
That is a lot of freight to load onto a funding round that closed seven months ago.
Because here is the thing the grandeur is papering over: this money is old. Bloomberg reported the outline of this exact round back in December, roughly $700 million at a $7.5 billion valuation, Situational Awareness leading. Fluidstack never formally confirmed it at the time, which is why Cox is calling the announcement overdue. The number firmed up to $830 million once you fold in the $450 million tranche that landed in January and the strategic checks that followed. Fine. But by April, the same company was reported to be in talks for a fresh billion-dollar round at an $18 billion valuation, co-led by Jane Street and, again, Situational Awareness. So Fluidstack is out here formally confirming a $7.5 billion round while the market has, on paper, already repriced it at more than double that. They will officially own the small number and stay quiet on the big one. I find that funny.
From reselling other people’s GPUs to pouring their own concrete
Fluidstack started in 2017 as an Oxford spinout with a pitch people liked to call “Airbnb for GPUs”: aggregate idle graphics cards, rent them out by the hour, undercut the hyperscalers. That company does not really exist anymore. The one wearing its name now manages more than 100,000 GPUs and builds custom data centers for frontier AI labs, having moved its headquarters from London to New York in December and walked away from a roughly 10-billion-euro project in France in March to go all-in on the US. Revenue went from $1.8 million in 2022 to $66.2 million in 2024, which is a real trajectory until you set it next to an $18 billion whisper number and realize the valuation is being written against contracts and buildout, not last year’s income statement.
The customer list is the tell. Anthropic signed a $50 billion partnership in November for custom data centers in Texas and New York, with first capacity due online this year. Mistral, Meta, Poolside, Black Forest Labs, and Character.AI all show up too. When Cox calls the label “Series A,” after a $200 million round in early 2025 and that $450 million January raise, he is doing some retroactive tidying of a cap table that grew faster than its own naming conventions could keep up with. Nobody raises their “Series A” at a $7.5 billion valuation. The label is cosplay. The demand underneath it is not.
Google is sitting on every side of this table
This is the part I actually care about, and it is the part almost nobody frames correctly.
Fluidstack does not own most of the sites it operates. It leases them, and increasingly it leases them from bitcoin miners who have discovered that renting halls to an AI operator pays better than mining ever did. That is the same landlord-tenant flip I keep writing about from the miner side. Fluidstack is what the tenant looks like from the other end of the contract.
Take TeraWulf. Last August, Fluidstack signed to take more than 200 MW at TeraWulf’s Lake Mariner site in western New York, a ten-year deal worth at least $3.7 billion in contracted revenue, and the relationship has since grown to around $6.7 billion across two agreements. TeraWulf’s stock jumped as much as 38% on the news. But the part that matters is who guaranteed the rent: Google backstopped roughly $1.8 billion of the obligation, and in exchange took about 8% of TeraWulf through some 41 million warrants. Then Cipher Mining, September, near-identical structure: 168 MW of critical capacity (244 MW gross) at Barber Lake in Colorado City, Texas, ten years, about $3 billion and up to $7 billion with extensions, with Google backstopping roughly $1.4 billion and collecting about 5.4% of Cipher via 24 million-odd warrants. Cipher kept 100% ownership of the site. In November, they extended it so Fluidstack now leases the entire 300 MW campus, Google’s backstop climbing past $1.7 billion. That expansion alone added around $830 million of contracted revenue to Cipher, which is, by pure coincidence, almost exactly the size of the equity round I opened this post with. Two completely unrelated $830 million figures orbiting the same company in the same year should tell you how loosely these ten-figure numbers get thrown around right now.
Sit with Google’s position for a second, because it is the whole story. It guarantees the rent Fluidstack owes the miners, and takes equity in those same miners as its fee for doing so. It is also Fluidstack’s first and primary partner for distributing its own TPU accelerators, so some of the silicon going into these halls is Google’s answer to Nvidia rather than Nvidia itself. And it owns roughly 14% of Anthropic, the anchor customer on that $50 billion buildout. Guarantor, shareholder in the landlord, chip supplier, and part-owner of the tenant’s biggest customer, all at once, and never once obliged to put the data center on its own balance sheet or answer for the capex on an earnings call. It is one of the most elegant pieces of financial engineering in the whole AI infrastructure boom, and it barely gets discussed because it is spread across four companies’ filings instead of sitting in one headline.
Debt fills the rest of the gap. Macquarie arranged a facility of up to $10 billion secured against Fluidstack’s GPU assets, and there is an Icelandic site running on renewables for the workloads that can tolerate the latency. When Cox says “hundreds of gigawatts,” this is the machine he means: miners as landlords, hyperscalers as guarantors, private credit as the mortar, and the AI labs signing decade-long leases that make the debt bankable. It works beautifully as long as the demand curve keeps pointing up. I am not going to pretend I know what happens to a warrant-and-backstop lattice like this if a model generation disappoints and the lease renewals wobble. That is its own post, and I would rather write it after something actually cracks than speculate now.
The money came with a worldview, and you can hear it
Situational Awareness is not a normal fund, and Fluidstack is not quoting the Marshall Plan by accident.
Leopold Aschenbrenner is a former OpenAI superalignment researcher who published a long essay called “Situational Awareness” in June 2024, the gist of which is that AGI arrives around 2027, triggers an intelligence explosion, becomes a US-versus-China national security problem, and requires a trillion-dollar compute cluster where the binding constraint is not algorithms but megawatts. He turned that thesis into an investment firm backed by the Collison brothers, Nat Friedman, and Daniel Gross, reportedly compounding a $225 million book into something like $5.5 billion, and disclosed a 9.4% stake in Core Scientific along the way. When you take that man’s money, you are not just taking capital. You are buying into the frame.
And you can hear the frame in every word of Cox’s announcement. “Arsenal of Democracy.” “Democracy wins.” “Every human on Earth benefits.” That is Aschenbrenner’s worldview reflected back almost verbatim, right down to the conviction that pouring concrete around GPUs is a civilizational obligation rather than a bet on enterprise demand for Claude and Mistral. The bottleneck-is-megawatts argument is the one correct load-bearing idea in the whole essay, and it is the same argument I have been making about why bitcoin miners suddenly matter. Where I get off the train is the leap from “electricity is the constraint” to “therefore we are running the Apollo Program.” One of those is a supply-chain observation. The other is a recruiting pitch dressed as manifest destiny, and the question of who actually controls the compute, the models, and the off switch does not get more comfortable just because the people building it believe they are the good guys.
There is a real policy story tangled in here too, and it is not abstract. Fluidstack is building in New York, where the fights over how much power a single campus can pull have turned into actual legislation, and the 50 MW line New York just drew is exactly the kind of rule that decides whether a buildout at this scale is even legal. The wider politics of where these things get sited, and the incentives being dangled to land them, are moving as fast as the funding rounds. The money is arriving faster than the grid or the statute books can absorb it, and Fluidstack’s whole model assumes both catch up in time.
I keep coming back to the gap between the register and the reality. This is a company that resold spare GPUs eight years ago, now formally announcing a funding round that closed in December, in the language of wartime industrial mobilization, while it quietly negotiates a valuation more than twice as high that it is not ready to confirm. The demand underneath all of it is real, and so is the buildout, and the Google financing lattice holding it together is the most interesting corporate structure I have looked at all month. But if you need me to salute the Arsenal of Democracy over an $830 million Series A, I am going to need Fluidstack to first admit what it is actually worth today. They know the number. They just told us the old one.