Written By | Torch Media Africa
WASHINGTON, D.C.
For twenty years the American defence establishment has worried about aircraft carriers, hypersonic missiles and rare-earth magnets. This week it turns out to be worrying about something far less glamorous: a steel box full of copper windings and mineral oil, roughly the size of a shipping container, that nobody outside the electricity industry has ever thought about twice.
The Wall Street Journal reported on Wednesday, September 10, that the Pentagon is in talks to lend roughly $5 billion to Fluidstack, a nine-year-old AI cloud company that owns almost no computer chips of its own and was founded by three friends who met at Oxford University. The money would come from the Department of Defense’s Office of Strategic Capital — a lending shop most Americans have never heard of, and one whose entire loan authority two years ago was less than a fifth of this single deal.
And here is the twist that has infrastructure executives passing the story around in group chats: the billions are not, according to the reporting, to build more data centres. They are to build the electrical guts that data centres cannot run without. Transformers. Switchgear. Circuit breakers. The unsexy apparatus that stands between a signed compute contract and an actual, live, humming megawatt.
“The Department does not comment on transactions that are either pending or in the middle of negotiations until the press release comes out.” — Pentagon spokesman, to The Wall Street Journal
The deal has not been signed. It may yet change shape or collapse altogether. But the fact that it is on the table at all tells you something enormous about where the AI boom has actually hit the wall — and it is not where most people assumed.
The lending office nobody had heard of
The Office of Strategic Capital, or OSC, was set up in December 2022 by then-Defense Secretary Lloyd Austin. Its logic was simple and slightly un-American: some technologies matter enormously for national security but are too capital-hungry and too slow-returning for venture capital, and too commercial for traditional defence procurement. So the Pentagon would do what the Department of Energy already did for clean energy — lend money rather than buy product.
For its first years, the office was small. Congress appropriated enough in fiscal 2024 for it to lend a total of about $984 million, in individual tickets ranging from roughly $10 million to $150 million, to companies working on any of the 31 critical technologies listed in that year’s National Defense Authorization Act — microelectronics, quantum computing, mesh networks, autonomous robotics, cybersecurity, and more. Previous borrowers have been rare-earth suppliers and drone manufacturers. Serious companies, modest cheques.
A single $5 billion loan would be, by an enormous margin, the largest thing the office has ever done. It would be roughly five times the office’s entire original loan authority, extended to one borrower, in one transaction. And it would not be a wild outlier for long: in April, the Pentagon asked Congress for more than $20 billion for the strategic capital loan programme in its 2027 budget request. The small lending desk is being rebuilt into something closer to a state development bank.
The startup that owns no chips
Fluidstack is one of the strangest success stories of the AI era. It was founded in Oxford, England in 2017 by Gary Wu, now the chief executive, alongside César Maklary and Jamie Cox. Cox was reading classics; Wu was studying economics. All three were in their twenties. The company began by scraping together spare graphics-card capacity wherever it could be found and renting it out — a kind of Airbnb for compute.
Nine years later it has relocated its headquarters to New York and is building or contracting data centre capacity across New York, Texas, Louisiana and Indiana. In early September it closed a $1.5 billion funding round led by the quantitative trading firm Jane Street, valuing the company at more than $18 billion and pushing its total funding past $2.6 billion. As recently as late 2025 it was valued at a fraction of that.
What makes Fluidstack unusual is that it sits in the middle of everyone else’s deals rather than owning the assets at the ends of them. In November 2025 the AI company Anthropic announced a $50 billion plan to build custom American data centres, starting with Fluidstack sites in Texas and New York. Google, meanwhile, provides a financial backstop on Fluidstack’s data centre leases — if Fluidstack fails to perform, Google assumes the lease or pays the termination fee — which lets a young company borrow at rates a young company has no business getting. Anthropic then leases the compute, including Google’s own TPU chips.
A company that owns essentially no chips has become a load-bearing wall of the American AI build-out — and now, possibly, a borrower from the United States Department of Defense.
That structure — creditworthiness borrowed from bigger balance sheets, revenue reportedly leaping from under $2 million to a projected figure in the hundreds of millions — is exactly what makes the Pentagon’s interest legible. Fluidstack is not being funded because it is a defence contractor. It is being funded because it has become a chokepoint.
The grey boxes that broke the boom
To understand why the money is going into manufacturing rather than buildings, you have to understand what has actually gone wrong with the electricity supply chain — and it is genuinely startling.
A large high-voltage power transformer used to take about a year to order and receive. In 2020 and 2021, that was still roughly true. Today, lead times for high-capacity transformers run to three or four years, and industry trackers report that a major high-voltage unit from a tier-one manufacturer can now quote at 60 months or more. Five years. Order a transformer today and you might energise it in 2031.
Switchgear — the assemblies of breakers and disconnects that route and protect power inside a facility — is no better. Medium-voltage switchgear carries multi-year backlogs and is described in several supply channels as effectively sold out through 2028. High-voltage circuit breakers are quoting around 125 weeks. Across the category, lead times sit at two to four times pre-pandemic norms, and manufacturers describe the demand shift not as a spike but as a permanent step change they are still tooling up to meet.
The consequences are already measurable in gigawatts. More than 6.4 gigawatts of data centre capacity that was expected to come online in the first quarter of 2026 slipped, adding to a backlog now measured at roughly 53 gigawatts of delayed capacity. That is not a shortage of ambition, or land, or even chips. That is a shortage of electrical hardware.
Put crudely: you can buy the world’s most advanced accelerators, sign a multi-billion-dollar lease, break ground in Texas, and still sit in the dark for years because the transformer queue is longer than the chip queue.
The emergency order that changed the definition of “defence”
There is a second piece of context without which the Pentagon’s involvement makes little sense. On August 26, 2026, President Trump signed an executive order declaring a national emergency over the security of the United States bulk-power system, invoking the International Emergency Economic Powers Act and the National Emergencies Act.
The order empowers the federal government to bar or condition transactions involving foreign-produced bulk-power equipment — transformers, control systems, and the associated software and firmware — where officials judge there is a risk of sabotage, unauthorised access, malicious remote action, or catastrophic effects on critical infrastructure. Strikingly, it also contemplates equipment already installed: the Secretary of Energy may impose conditions on the continued use, operation, maintenance, servicing or updating of foreign-made grid gear bought before the order was signed. The Department of Energy is required to publish implementing rules within 120 days — by around December 24 of this year.
That order is the hinge. The moment foreign-made grid componentry is formally a national security risk, domestic manufacturing capacity for that componentry becomes a national security asset — and therefore, on the Office of Strategic Capital’s own terms, something the Pentagon can lend against. A loan that would have looked in 2024 like the Defense Department subsidising the cloud industry looks in September 2026 like the Defense Department onshoring the electrical supply chain.
Expert perspective: what the analysts are actually saying
Infrastructure analysts have been blunt that the bottleneck has moved. Industry assessments through 2026 have converged on the same short list of constraints — AI compute demand, government chip policy, dependence on mature-node components, advanced packaging, and power semiconductors — with electrical equipment increasingly named as the binding one. Build.inc, which advises on data centre development, has gone as far as describing transformer procurement as a development constraint in its own right rather than a line item in a schedule.
The manufacturers’ order books tell the same story from the other side. One power systems supplier reported third-quarter 2026 orders up 268 per cent year over year against a $1.5 billion backlog — roughly eighteen months of production already sold.
Sceptics raise two fair objections. The first is moral hazard: Fluidstack is an $18 billion private company that just raised $1.5 billion from one of the most sophisticated trading firms on earth and enjoys a lease backstop from Google. If that company cannot raise capital on commercial terms, who can? The second is precedent. Once the Pentagon’s balance sheet is available for AI-adjacent infrastructure, the queue at the door will not be short — and the $20 billion-plus request for 2027 suggests the government expects exactly that.
What this means for you
If you have ever wondered why your utility bill keeps climbing while your habits stay the same, this story is a large part of the answer. The same transformers and switchgear that data centres are queuing for are the ones your local utility needs to replace ageing equipment, connect new housing, harden lines against storms and hook up solar and battery projects. When a handful of enormously well-funded buyers absorb years of manufacturing output, everyone else waits longer and pays more.
There is also a quieter consumer story. Grid interconnection queues are now a routine reason that new housing developments, factory expansions and electric vehicle charging hubs slip by years. The bottleneck that delays a data centre in Louisiana is the same bottleneck that delays a substation upgrade in your suburb.
And if the executive order’s implementing rules land as written in December, utilities may face conditions on foreign-made equipment they have already installed — costs that, historically, flow to ratepayers.
The bottom line
Nothing here is final. The Pentagon has confirmed nothing, Fluidstack has not commented publicly, and a deal of this size in this political environment could easily be restructured, shrunk or shelved. Treat the $5 billion figure as a negotiating position reported by people familiar with the talks, not a signed cheque.
But the direction of travel is unmistakable. The artificial intelligence race has stopped being a race about algorithms, and stopped being purely a race about chips. It has become a race about steel, copper, oil-filled tanks and factory floor space — about whether America can build the electrical hardware to plug its own intelligence in. The Pentagon appears to have concluded that it cannot, and that this is now a defence problem.
Watch three dates. December 24, when the Department of Energy’s grid-equipment rules are due. The 2027 budget cycle, when Congress decides whether the Office of Strategic Capital gets its $20 billion. And whenever a Pentagon press release finally appears — because, as the spokesman said, that is the only moment the Department will admit any of this is happening at all.