Here is a contradiction worth sitting with. BlackRock's technology investors looked at the same power crunch everyone else is panicking about and concluded, flatly, "no" — power constraints will not derail the AI buildout. Their April note calls the frictions a "self-regulating mechanism" that caps overbuilding and prioritizes the best projects.
Then look at Essex.
A £2bn data center, stuffed with Nvidia GPUs, backed by more than $5bn from Nokia, Nvidia, Dell and Blue Owl, being built to serve Microsoft — and it cannot open on schedule because it is missing one 90-megawatt wire.
Nscale's flagship UK site was told the grid connection won't be ready for its planned 2027 opening. So the company is now in talks with California's Bloom Energy to route around the grid entirely, powering the site with solid oxide fuel cells while it waits for a connection the utility can't deliver on time.
That is the whole thesis in one building. Capital arrived years ahead of electrons, and the only fix on offer is to build your own power plant in the parking lot.
Capital arrived years ahead of electrons, and the only fix on offer is to build your own power plant in the parking lot.
Last week, on July 1, we asked whether hyperscaler capex would hold. This week the question sharpens into something more physical: not whether the money keeps coming, but whether the power and the debt structures underneath it can survive the collision.
Because the money is emphatically still coming. Start there.
Nscale's Essex problem is not a one-off. Sightline Climate estimates 26% of data-center capacity was delayed in 2025, and that 30–50% of the global capacity expected in 2026 could slip on power constraints, equipment shortages, or local opposition.
Half of next year's promised compute might not show up on time. Hold that number, because it is the fact everything else has to be priced against.
Now the financing. When Morgan Stanley brought a $3.1 billion loan for CoreWeave to market, investors put in more than $19 billion of orders — what reportedly amounts to roughly a six-times oversubscription for AI infrastructure debt.
The uncomfortable line, from the reporting on that deal: "the loans may outlast the chips and customer contracts supporting them."
Read that slowly. The debt is being written against depreciating chips and finite customer contracts, but structured to sit on balance sheets longer than either lasts. You are lending against collateral that expires before the loan does.
Mistral shows the same shape in Europe. The French lab raised $830 million in debt to build a Paris data center around 13,800 of Nvidia's GB300 GPUs, backed by a consortium from Bpifrance to HSBC.
But GPUs are a wasting asset — and the debt is a multi-year claim. Banks are betting that AI infrastructure is the next great asset class. They are underwriting it as if it ages like a toll road, not like a laptop.
This is the structural point Columbia's Stijn Van Nieuwerburgh makes in his March paper on financing the buildout: the user of the compute is increasingly not the owner of the asset, and risk is sliced across "a layered set of claims held by different investors."
Off-balance-sheet structures, long-term leases, project finance, asset-backed lending — all of it lets firms scale fast while keeping equity valuations high and the true distribution of risk opaque. He compares it, pointedly, to railroads and to prior episodes of complex structured finance. Both are cautionary tales.
And now the regulators have arrived, which is the clearest sign that power access has stopped being an engineering problem and become a bond-market one.
On June 18, 2026, FERC issued Section 206 show-cause orders to all six grid operators under its jurisdiction — PJM, MISO, SPP, CAISO, ISO-NE, NYISO — demanding they justify or rewrite their tariffs for loads over 50MW.
The legal significance, per the framing around those orders, is that large-load interconnection is now "a tariff-design problem, not just an engineering queue problem." The federal government is redesigning the rules of grid access because the queue has become a financial risk.
How big is the queue? In Texas, ERCOT is sitting on a 438 GW large-load request queue — nearly 90% of it data centers.
For scale: BlackRock estimates U.S. data centers consumed roughly 42GW in 2025, though its own chart puts the 2025 figure at 19GW. Texas alone reportedly has requests on the order of ten times the entire country's current data-center draw. Most of that will never energize, and the unresolved question in every docket is the same — who pays when the projected load never fully materializes.
Two debt channels are now converging on one physical bottleneck. Utilities are borrowing to fund the grid capex; hyperscalers are borrowing to fund the chips and buildings. Both claims rest on the same wire that, in Essex, isn't there.
So watch how the chip vendors are reacting to the squeeze, because it is telling.
Google is pitching its TPUs as a cheaper, more reliable alternative to Nvidia's GPUs, targeting exactly the cloud providers that built their businesses on Nvidia hardware.
Nvidia's counter is not just faster silicon. It's financial guarantees — backstops that could help those same customers fund new data centers.
Sit with the circularity there. The dominant chip supplier is now helping finance the buildings that will house its chips, to defend against a rival offering to lower the cost of the compute. The vendor is becoming the lender to its own demand. That deepens exactly the layered, opaque risk stack Van Nieuwerburgh warns about.
Which brings us to the escape hatch nobody can actually build.
If electrons won't arrive on the ground fast enough, the pitch goes, put the compute in orbit — continuous solar, natural cooling, no grid queue. BlackRock lists space data centers alongside nuclear as a genuine 2030s solution. SpaceX's proposed orbital data centers are, according to TechCrunch, reportedly the main driver behind its $2 trillion valuation.
Then, over one weekend in July, Sam Altman and Elon Musk traded insults, and Altman said the quiet part: "homeboy you're the one sellling public market investors on short-term space datacenters."
The striking thing is that Altman is right, and the experts agree with him. Talk to the entrepreneurs, to Google's own orbital-compute team, to engineers who ran the numbers for fun, and you get the same answer: not soon.
The blocker is launch economics. SpaceX reportedly conceded during its IPO roadshow that Starship may not be fully reusable near-term and may have to throw away each second stage per launch — which, as one assessment put it, "would put a kibosh on economical space data centers."
Musk's rejoinder was "We start flying them next year." Sure. But launching one satellite is not manufacturing and launching them at scale, and that — everyone agrees — is a question for the 2030s. The escape hatch opens a decade after the debt comes due.
Notice, too, where the talent is quietly pointing. Anthropic just recruited Monzo cofounder Tom Blomfield onto its compute team, where he wrote that "availability of compute becomes one of the most important issues to solve." When your bankers and your best operators are both fixated on compute supply, that's the constraint.
So here is the synthesis, and it is uncomfortable.
The AI buildout is a bet that capital can outrun physics. Money can be raised in weeks — $19 billion of orders for a $3 billion loan — but a 90MW connection in Essex takes years the grid doesn't have.
The financing structures being written today assume a smoothness that the interconnection queues and the delayed-capacity numbers flatly contradict. Debt is being priced as multi-decade claims against chips that depreciate and connections that may slip, on the assumption that 30–50% of next year's capacity somehow shows up anyway.
BlackRock may yet be right that the system self-regulates. But self-regulation, in their own telling, means capacity gets capped and marginal projects die.
The debt does not price for projects dying. It prices for the electrons arriving on schedule. And in Essex, they didn't — so a data center is shopping for fuel cells to paper over the gap between what capital promised and what the grid can deliver.
Watch the connection dates. That's where this either holds or breaks.