Every cycle, a story arrives that the crypto crowd mistakes for someone else's story, and the mistake is always expensive. This quarter it is Nscale — an AI infrastructure firm, incubated years ago by a distressed-debt fund, now moving toward a public listing on the strength of a roughly $6.2 billion multi-year compute agreement with Microsoft, executed on GB200 racks in a converted industrial site in Loughton, England, plus a slice of OpenAI's Stargate Norway programme alongside Aker. That a company most readers have never heard of can anchor the single largest sovereign-adjacent compute deal in Britain should tell you the plumbing has changed.
The financial press will file this under artificial intelligence. The people who watch bitcoin hashrate should file it somewhere else entirely. Because the physical substrate of Nscale — the grid interconnects, the transformers, the liquid-cooled halls, the brownfield shells — is the same substrate the mining industry spent a decade assembling, and it is now being repriced by a liquidity flow that has nothing to do with mining and everything to do with where capital wants to sit. Tracing the liquidity ghost in the machine has never been more literal.
To read Nscale you first need to know what a neocloud is. Strip away the branding and it is a company that rents GPU clusters under long-duration, take-or-pay contracts to hyperscalers and AI labs, financing the hardware with debt. The scarce input is not the chip and it is not the capital; it is the permission to draw power from a grid and the building in which that power becomes compute. In the United States and Britain, interconnection queues now run five years and beyond, and some regional grids have no spare capacity until the 2030s. A decommissioned bitcoin mine, by contrast, already holds an energized high-voltage connection and a shell engineered for continuous, high-density load.
Europe adds a policy tailwind the American neoclouds lack. Britain's AI Growth Zones and the continent's broader anxiety about hosting its critical compute on American hyperscaler infrastructure have created a class of government buyers for whom price is not the primary variable; sovereignty is.
For a decade the crypto industry was the marginal buyer of exactly these assets, and in doing so it accumulated, almost by accident, the most valuable industrial option of the AI age. Nscale's strategy is not ingenious; it is simply to acquire the brownfield, rewire it for liquid-cooled GB200 density, and sell the resulting compute forward. I have spent years modelling liquidity for institutions that prefer to remain unnamed, and I can tell you the structure that results looks nothing like a software company and everything like a regulated utility or a data-centre REIT — high margins, heavy leverage, and a timing mismatch between accounting income and real cash that financiers love to litigate. The ETF wave washed away the retail tide in crypto; in compute, the institutional bid has done the same, and the marginal buyer of power is now a pensions data-centre sleeve rather than a hash-rate speculator.
That mismatch is where I want to slow down, because it is where the whole edifice is most fragile. The GAAP net loss at a neocloud is not the product of failure; it is the arithmetic of three lines — GPU depreciation, debt interest, and stock compensation. The comparable operator CoreWeave reported roughly $1.92 billion of 2024 revenue against adjusted EBITDA near $1.2 billion and a net loss in the billions, and the gap between those two figures is the true battlefield of the valuation. A company can post dazzling adjusted EBITDA and still be consuming capital at a rate that only works while the debt markets stay open and the rental price of compute stays high.
The debt itself tells a story. GPU-collateralized asset-backed securities were priced, through 2025, in the nine-to-ten percent range, and that spread is not a rounding error — it is the market's priced opinion on the residual value of silicon and the credit of the counterparty. If Nscale's cost of borrowing sits materially below that band, ask yourself who is subsidizing it. If it sits above, ask what the lenders know about the collateral that the equity story omits.
When the incubating fund is a distressed-debt specialist, the capital structure rarely arrives clean. Related-party debt at above-market rates, preferred-stock liquidation preferences, and performance ratchets are the genre, and they are typically disclosed only deep in the related-party section of the filing — read the footnotes before the headline.
Then there is the circularity. Nvidia appears simultaneously as supplier, as ecosystem partner, and — by most public accounts — as an investor, which produces a loop in which the vendor funds the customer, the customer buys the vendor's product, and the product is pledged to borrow against. This loop is harmless while demand is real, and it contracts from three directions at once when demand softens. History rhymes in the ledger, and this particular rhyme has been sung before — in telecom equipment financing, in aircraft leasing, in shale.
The other question the marketing will not surface is the depreciation schedule. If the filing assumes six years of useful life for a GPU whose economically productive lifespan is three, then current earnings are being systematically flattered and the reckoning is merely deferred. It is the single line I would read word for word before touching the equity.
And then the geopolitical layer, which the coverage ignores entirely. A British company, buying American chips, servicing American and European clients, sits inside two regulatory regimes at once. Cross-border compute export controls, the EU AI Act's obligations on high-risk systems, and data-residency rules are not footnotes; they are the terms on which a sovereign-AI vendor keeps its contracts. The government customer is also the regulator, which makes him a moat and a liability at the same time.
Now the contrarian turn, because this is where the crypto reader's instinct goes wrong. The seductive inference — mines become AI data centres, therefore mining assets are undervalued, therefore crypto is levered to AI — is a decoupling thesis that flatters both sides and satisfies neither. The truth is subtler and colder. The repricing of brownfield power is real, but it does not lift the crypto complex; it partially cannibalizes it. Every energized site diverted from hashing to inference is a site removed from the bitcoin network's supply of cheap marginal power, which means the miner-to-AI conversion the market reads as bullish is, on a longer horizon, a slow compression of hashrate growth and a quiet redistribution of the industry's industrial base.
We were told the merge was a fever dream for liquidity. Something similar is happening here, in reverse — a fever dream for power — and the crypto industry is watching its own real estate walk out the door and into an AI prospectus.
So where does this leave the cycle? Nscale's listing, if it lands at the multiple its underwriters want, becomes a thermometer rather than a story: it tells you how deep the capital cycle has pushed asset-backed, highly concentrated, long-contract infrastructure into the public market, and it primes a queue of imitators. Watch the depreciation line, watch the customer concentration, watch the spread on GPU-backed debt. The question that should keep you awake is not whether Nscale succeeds. It is what the world looks like in 2028, when silicon is abundant, hyperscalers build their own halls, the contracts come up for renewal, and the ghost in the grid finally sends its invoice.
