Nscale's IPO Is a Depreciation Bet, Not a Compute Story

I read the depreciation schedule before I read the revenue. It is the quietest line in any AI infrastructure filing, and it is the one that decides whether "stunning economics" means a sixty percent adjusted EBITDA margin or an eight-point-six percent net loss margin stretching toward 2030. Nscale's IPO paperwork is not yet the thing itself. What surfaced in early reporting is a signal, not a dataset. A Microsoft contract near $6.2 billion. A role in OpenAI's Stargate Norway footprint, alongside Aker. A balance sheet built on GPU-collateralized debt. No revenue figure, no net loss, no customer concentration ratio made it into the coverage. Five facts, three of them metadata. So I trace the shadow before it casts. Before the numbers arrive, the structure already tells you what deserves auditing. What Nscale actually sells is not compute. It sells the financialization of two scarce assets: grid interconnection rights and long-dated take-or-pay contracts. The mechanics fit on one line. A hyperscaler signs a multi-year agreement to buy capacity it may never fully use — take-or-pay. Nscale uses that contract and the GPUs behind it as collateral, raises debt against both, and earns the spread between its cost of capital and the contract's yield. The GPUs are not a product. They are a mortgage. This is a quasi-REIT wearing a software company's valuation. The revenue is contractual and concentrated, the assets are physical and depreciating, the returns are levered. Every structural feature points at power and property, not at algorithms. CoreWeave printed the template: $1.92 billion of 2024 revenue with roughly sixty-two percent from a single counterparty, Microsoft. Concentration here is not a management failure. It is the endogenous property of the model, and it appears again in Nscale's dependence on Microsoft and OpenAI. Nscale adds two variables. First, anchor tenants with investment-grade or near-investment-grade credit, which lowers the cost of the debt raised against those GPUs relative to peers with no anchor. Second, a European posture: Loughton in the UK, Norway's hydroelectric capacity, a sovereign-AI narrative that governments find politically attractive and price-insensitive. Then there is the origin story. Nscale was incubated with Sandton Capital, a distressed-debt shop. A company born in workout capital will carry workout capital's fingerprints: related-party notes, liquidation preferences, performance covenants. Those terms decide what common equity actually owns. They are usually buried in the related-transaction section, and they are the first place I would look. The most consequential number in the filing — when it appears — will be the depreciation horizon assigned to the GPUs. Here is the arithmetic. CoreWeave's 2024: $1.92 billion revenue, roughly $1.2 billion adjusted EBITDA, about $860 million net loss. The gulf between those last two figures is not operational decay. It is depreciation plus interest plus stock compensation. When a company reports a sixty percent adjusted EBITDA margin and a negative net margin in the same breath, both numbers are true and mutually explanatory. The adjusted figure strips out the cost of the hardware dying. The GAAP figure puts it back. So the question is not whether Nscale is profitable on an adjusted basis. It is how long the company believes a GB200 holds value. If the filing assumes six years and the economic life is three — the pace at which NVIDIA itself obsoletes its own silicon — then current-period earnings are systematically overstated, and the correction arrives as a concentrated impairment the moment the cycle turns. The bug hides in the beauty. A beautiful EBITDA margin and a fragile depreciation assumption are the same sentence read twice. I listen to what the compiler ignores, and it says: check the collateral's half-life. GPU-collateralized debt is a new instrument, and its pricing is the market's most honest opinion. CoreWeave's GPU ABS priced in the nine-to-ten percent range in 2025. That spread is not a financing accident. It is the market pricing residual-value risk and counterparty risk in one number. If Nscale's debt clears materially below that band, the difference is either better collateral or a related party subsidizing the rate. If it clears above, the collateral is suspect. Either deviation is a disclosure you should demand. The second structural question is circularity. NVIDIA appears in this story three times: as the supplier of the GPUs, as a reported investor in the equity, and as a beneficiary of the demand that equity funds. That loop is harmless while demand is real, and contracts in both directions when it is not. Vendor financing has a long history, and its failure mode is synchronized — when end-demand softens, the vendor's orders and the customer's ability to pay weaken together. Regulators in France and at the US Department of Justice are already studying vertical lock-in across compute supply. A filing that makes this loop explicit could validate or embarrass the structure, depending on how the related-party section is written. Then there is time-to-power, the one moat that is real. In the UK and the US, a new data center can wait five years or more for a grid interconnection, with some queues stretching into the 2030s. A brownfield industrial site — a retired plant, a former Bitcoin mine — already holds the high-voltage connection. It can deliver in twelve to twenty-four months. That difference is Nscale's entire asymmetric advantage, and it is the reason miners like IREN, TeraWulf, and Cipher have become AI-hosting candidates. But "has power" is not "has a data room." A liquid-cooled GB200 rack draws north of 120 kilowatts per cabinet, and it demands rebuilt cooling, mid-voltage DC power architecture, and a fresh network topology. The brownfield's electrical hookup is the beginning of the conversion cost, not the end of it. External observers consistently underestimate this gap. The number that would resolve it — cluster utilization, or model FLOP utilization — almost never appears in these filings. Vulnerability is just a question unasked. Ask what percentage of nameplate capacity is actually selling. There is also a jurisdictional layer that crypto-native coverage tends to skip. Nscale is a UK company buying American silicon to serve American customers, with assets in Norway. That places it under US export-control rules still being written, European data-sovereignty obligations, and the EU AI Act's high-risk provisions. As a sovereign-AI supplier, its customer is a government — and that government is also its regulator. Stability and regulatory exposure arrive in the same envelope. None of this is a prediction of failure. I trace the shadow before it casts. The obvious reading of this IPO is a bet on AI compute demand. That is the wrong variable. Demand for GPUs is not in question. What is in question is whether the value of a deployed GPU three years from now supports the debt raised against it today. This is a collateral question dressed as a growth question. When a company borrows against an asset whose resale value halves every hardware generation, the lender is underwriting a depreciation curve, not a revenue line. The second contrarian point: an IPO is a timing decision, not a validation. A distressed-debt sponsor and a management team choose the public window when they believe it is the best one available. Filing into late 2025 or early 2026 — before the first wave of self-built hyperscaler capacity lands and before GPU supply normalizes — is a statement about the future cost of capital. If Nscale's owners believed the 2027 environment would price these contracts more generously, they would wait. And note what the coverage left out entirely. No pricing range. No underwriters. No use of proceeds. No third or fourth customer. In the void, the bytes whisper truth — and the silence here is louder than the $6.2 billion headline. Watch one line. When the S-1 lands, read the depreciation schedule and the related-party notes before you read a single revenue figure. Security is the shape of freedom, and in a capital structure, the depreciation assumption is the security. Everything else — the sovereign-AI story, the Microsoft anchor, the Norwegian hydro — is downstream of how long the hardware is assumed to last. The compute cycle will answer the rest. I will be listening to the version of this that survives the audit.