Hook
Over the past eighteen months, the marginal buyer of Nvidia's GB200 has not been a hyperscaler. It has been a neocloud โ a rented-compute vendor whose loan book is dressed as a balance sheet. Nscale, the London-based operator that has now filed for an IPO, is the clearest specimen of the species yet. And the most important line in its filing will not be the revenue figure. It will be the depreciation schedule.
That is the detail that decides whether an "AI infrastructure" asset lives or dies. Everything else โ the roughly $6.2 billion multi-year Microsoft capacity agreement, the role in OpenAI's Stargate Norway alongside Aker, the 230MW buildout ambition โ is narrative. Narrative is liquid. The depreciation assumption is solid.
Context
Nscale's lineage matters more than its marketing. It was incubated by Sandton Capital, a distressed-debt shop. That single fact reframes the entire company. This is not a research lab that stumbled into a business model. It is a capital structure that acquired the right assets: grid interconnection rights at brownfield sites โ former bitcoin mines among them โ that already hold high-voltage access.
In the US and UK, queuing for new grid interconnection now runs five years, with some regions penciled into the 2030s. A brownfield site with existing power can be delivering sold compute in twelve to twenty-four months. That gap is the whole business. Not GPUs. Not models. Electricity and permission.
The anchoring customers are what make the debt work. Microsoft and OpenAI sit at investment-grade or near-investment-grade credit. That counterparty quality lowers the cost of GPU-collateralized borrowing, which lowers the effective yield Nscale must clear. It is a chain: contract creates credit quality, credit quality creates cheap debt, cheap debt funds more GPUs.
Elegant โ until it isn't. When I audited Golem's reward-distribution mechanism back in 2017, the failure mode was identical in shape: a structure that looked self-reinforcing only as long as one exogenous variable held. Then it broke.
Core
Here is the mechanism, stripped of adjectives. Take-or-pay contracts with hyperscalers generate predictable contracted revenue. That stream is pledged against GPUs. The spread between cost of capital and contracted yield is the profit. Functionally, this is a REIT with a liquid-cooling system bolted on. It should be valued like a utility or a data-center trust.
The market keeps pricing these companies like software.
Look at the comparable. CoreWeave in 2024: roughly $1.92 billion revenue, roughly $1.2 billion adjusted EBITDA, roughly $860 million net loss. That gap is not mismanagement. It is arithmetic โ GPU depreciation, debt interest, stock compensation. Expect the same silhouette in Nscale's S-1. The headline phrase "stunning economics" cuts in both directions at once: stunning margins, stunning losses, simultaneously. Math does not care about your conviction that high EBITDA means profitability.
The real signals are buried in the footnotes.
First, the depreciation life assumed for GPUs. If the company amortizes over six years while the economic life is three, reported profit is systematically inflated. Every AI-infrastructure filing deserves to be read one line at a time, and this is the line.
Second, the definition of backlog. Contracted-but-not-yet-energized capacity is not revenue. The ratio of backlog to recognized revenue tells you whether "amazing growth" is banked or merely promised. The furthest-dated portion depends on data centers being delivered on schedule โ a high-variance assumption in UK and Norwegian grid conditions.
Third, related-party leverage. A distressed-debt incubator does not build clean capital structures. Expect preferred liquidation preferences, above-market intercompany debt, performance ratchets. These live in the related-transactions section โ usually page ninety-something โ and they silently reshape common shareholders' economics. In summer 2020, tracking capital velocity between Compound and Aave taught me the same lesson: the contract terms, not the headline APY, determine who actually eats.
Then there is Nvidia. Supplier, ecosystem partner, and reportedly investor. Follow the loop: Nvidia invests in Nscale, Nscale buys Nvidia GPUs, those GPUs collateralize debt, the debt funds more GPU purchases. Harmless while demand is real. Contractile when it isn't. In the chaos, look for the invariant โ and the invariant here is who bears GPU residual risk at the end of the chain. It is the lender, not the equity holder, not the vendor.
Add a regulatory layer that the retail framing ignores. AI chip export and re-export fall under US BIS rules. A British company buying American silicon to serve American clients sits inside two policy regimes at once. If "compute service delivery" is ever pulled into export-control scope, the cross-border model is exposed in ways no revenue line captures. The sovereign-AI demand that underwrites the European positioning is real โ but the customer is also the regulator, and that is a moat and a liability in the same contract.
Contrarian
The consensus framing is that neoclouds compete with one another. CoreWeave versus Nebius versus Crusoe versus Nscale. That is the wrong axis entirely.
The threat is upstream. Microsoft, Amazon, Google and Meta are building AI capacity at a scale no neocloud can match. The neocloud exists for exactly one reason: a temporary capacity deficit at the hyperscalers. When self-built capacity catches demand โ or when AI capex growth merely decelerates โ outsourced compute is the first line item to shrink. Contract renewal, not peer competition, is the risk.
And the crypto-native audience holds a second blind spot. The story being sold is "bitcoin mines are becoming AI data centers, therefore mining assets are undervalued." Partially true โ brownfield sites carry genuine optionality. But "has power" is not "has a sellable cluster." Liquid-cooled GB200 racks draw more than 120kW per cabinet and demand rebuilt cooling, medium-voltage distribution, and network topology. The conversion cost is routinely underweighted. Miners who pivot and stumble will still have stopped hashing โ which quietly compresses bitcoin's hashrate growth. That is a second-order effect nobody is pricing.
The crypto connection is why this filing appeared where it did. Crypto-native media has a structural incentive to amplify the "mining assets re-rate into AI" thesis. Treat the optimism as a tell, not a signal.
Takeaway
Nscale's IPO is not really a company story. It is a thermometer. If a highly levered, customer-concentrated asset backed by long-dated contracts can list at a rich multiple, the AI capital cycle is still in its self-reinforcing phase, and more of these issuances will follow. If it breaks issue, it becomes a marker of the cycle turning.
Either way, the filing itself is the information: the depreciation assumption, the backlog definition, the lock-up structure, the third customer โ if one exists. Solitude is the price of clear vision, and the crowd is watching the price. Quietly positioned while the world shouts, the analyst should be reading page ninety-something, where the physics of the balance sheet actually lives.