office of strategic capital
The state has started underwriting compute. Not by building it — by lending against it, which is a much harder thing to be right about.
The Wall Street Journal reported on 11 September that the Pentagon is in talks to lend about $5 billion to Fluidstack, an AI cloud company founded in London in 2017 and since relocated to the United States. Reuters carried the report. The deal is not signed, and both outlets say the terms could change, so treat every number here as reported rather than settled.
The money would come from the Office of Strategic Capital, which is a credit shop inside the Department of Defense. It makes loans to companies doing work the department considers critical to national security. Prior recipients include rare-earth suppliers and drone manufacturers.
Read that list again, because it tells you what the office thinks it is for. Rare earths and drones are both things you cannot buy in a hurry. You can buy a lot of software in a hurry. You cannot buy a separation facility in a hurry, and the whole theory of the office is that some capacity has a lead time long enough that markets will not produce it in time to matter.
Compute has now been added to that list. That is the news, and it is real news.
Here is the detail that almost every write-up buried, and it is the one that changes the argument.
The loan is not for a data center. Per the reporting, Fluidstack would use it to shore up U.S. supply chain and manufacturing capacity for certain data-center-related components — not to fund a new AI facility outright. Fluidstack is building capacity in New York, Indiana, Louisiana and Texas, but that is not what the five billion is nominally against.
So the Pentagon is not buying compute. It is not even financing the buildings that hold compute. It is financing the upstream manufacturing of some specific set of parts that go into the buildings that hold compute, through a company that operates the buildings.
That is a very particular bet, and it requires knowing something quite hard: which component is the one that actually gates capacity.
This is the oldest unanswered objection in economics, and it is worth stating precisely, because the casual version of it is wrong.
Hayek’s knowledge problem is usually rendered as “central planners have too much data to process.” That version is weak, it was weak in 1945, and modern computation plausibly answers it. The strong version is the one that survives. Hayek’s claim in “The Use of Knowledge in Society” is that the decisive knowledge is knowledge of the particular circumstances of time and place — this supplier is slipping, this transformer order is quietly eighteen months out, this crew is the only one that has commissioned this class of equipment — and that much of it is tacit, unstated, and held in fragments by people on the spot.
The part that bites hardest is his later formulation: competition is a discovery procedure. A great deal of the relevant knowledge does not exist before the process that elicits it. It is not sitting in a database waiting to be queried. It is generated by people making decisions under real stakes and finding out what breaks.
Which bottleneck matters is exactly that kind of knowledge. You find out by watching who fails to deliver. A loan front-runs the discovery. It has to name the bottleneck in a credit agreement, before the market has run the experiment that would reveal it.
I want to be fair to the other side, because this is not a knockdown. The counterargument is that we already know the answer — that power equipment and high-voltage gear and advanced packaging have been visibly constrained for two years, that this is not a mystery requiring discovery, and that waiting for the market to confirm what everyone in the industry already says is a luxury a defense department does not have. That is a serious reply. It may well be correct. It is also exactly the kind of confidence Hayek called the pretence of knowledge, and the way you would find out it was wrong is by being wrong expensively.
There is a second problem, and it is about what a lender can see.
James Scott’s argument in Seeing Like a State is that a state cannot govern what it cannot read, so it remakes the world into forms it can read — cadastral maps, standardized measures, permanent surnames. The simplification is not neutral. The map gets pressed back onto the territory, and the things the map left out turn out to have been load-bearing.
A credit facility is a legibility instrument. To lend against a supply chain you need the supply chain to have the properties a lender can underwrite: a factory with an address, a bill of materials, a lead time, a purchase order, a depreciation schedule. Those are real things. They are also the legible things.
The thing that most often gates a data center coming online is not on that list. It is an interconnection queue position, a utility’s willingness to energize, a substation commissioning crew, a county commission meeting. It is a hundred people who have done this before and know which vendor actually ships. Scott’s word for that is mētis — practical knowledge that lives in a setting and cannot be written down without ceasing to be what it is. You cannot pledge mētis as collateral. So the instrument reaches for what it can hold, and what it can hold is steel.
Scott is careful and I will be too: he says plainly that many state simplifications were enormous goods, and that legibility on its own is usually benign. This is not a catastrophe argument. It is a narrower claim — that the shape of the tool selects the target, and the target it selects is the part of the problem that fits in a loan document.
Now the asymmetry that I think is the actual point.
There is an older way to do this, and the United States has done it: the state builds the thing. Arsenals, the interstates, the Strategic Petroleum Reserve, a fair amount of the early semiconductor industry bought into existence by simply being the customer.
Compare the failure modes. If you build the wrong thing, you own the wrong thing. A mothballed plant is a bad outcome and it is also an asset, sitting on public books, available to be repurposed by someone with a better idea in fifteen years. Half the industrial base of the last century is made of facilities built for a reason that stopped applying.
If you lend against the wrong thing, you own a default. There is no building. There is a workout, a recovery percentage, and a line item. Construction is a bet you can be wrong about and still own something. Credit is a bet where being wrong leaves you holding paper.
And the upside is asymmetric in the other direction. If the loan works — if Fluidstack builds the capacity and repays at a modest coupon — the public gets its money back plus interest, and the private equity holders get the appreciation on a strategically indispensable asset that the government identified and de-risked. The taxpayer takes a slice of the downside and rents out the upside.
The steelman here is genuinely strong and I will not skip it. Credit is cheap. A loan that gets repaid costs far less than an appropriation that gets spent, it does not require Congress to re-fight the same battle every fiscal year, and it crowds in private capital at a ratio no direct program achieves. The Office of Strategic Capital exists because appropriations are slow and politically expensive, and that is not a bad reason for a thing to exist. If you think the alternative to this loan is a well-designed public buildout, the alternative to this loan is probably nothing.
So: the name.
Someone chose the words Office of Strategic Capital, and the words are exact. Not the Office of Strategic Production. Not Strategic Industry. Capital. The state has concluded that its comparative advantage in the industrial base is not making things and not buying things, but deciding who is creditworthy.
That is a claim about knowledge before it is a claim about money. It says the government can identify, ahead of the market, which private balance sheet is load-bearing for national security, and can be confident enough in that judgment to write a nine-figure check against it.
It might be right. Rare earths were not a hard call, and power equipment may not be either.
But it is the one claim the entire argument against central planning was about, and we have now decided to test it in the least reversible form available — not a factory you can walk into and repurpose, but a note that either performs or does not.
Sources: Reuters via The Star ↗ · Data Center Dynamics ↗ · Reuters via Yahoo Finance ↗ · Hayek, “The Use of Knowledge in Society” (1945) ↗ · Hayek, “The Pretence of Knowledge” (Nobel lecture, 1974) ↗ · Scott, Seeing Like a State (1998) ↗