The Depreciation Line Is the Only Audit That Matters: Reading Nscale's IPO

The Depreciation Line Is the Only Audit That Matters: Reading Nscale's IPO

The filing landed, and the headline said the numbers were stunning. There was no revenue figure in it. No net loss. No customer concentration ratio. No backlog definition. No debt maturity ladder. The S-1 that a crypto-native outlet framed as a revelation of "stunning economics" arrived with the accounting equivalent of a blank signature block — and the most consequential line in any AI infrastructure prospectus, the depreciation schedule, was never quoted. A reader unfamiliar with how these documents usually work would conclude the company has no financials. A reader who has audited them knows something different: the numbers exist, and the fact that they were omitted from the summary is itself the signal.

That asymmetry — a provocative adjective attached to an absent dataset — is where this analysis has to start. Because Nscale is not a software company, and its IPO is not a technology story. It is a story about electricity, grid interconnection rights, GPU residual value, and the quiet machinery of deferred depreciation. The crypto audience reading it, meanwhile, is being handed a narrative about redemption arcs and asset repricing without being told which line of the ledger carries the risk.

Context: what a neocloud actually is

Nscale belongs to a category the industry calls "neoclouds" — companies that buy GPUs at scale, house them in power-dense data centers, and rent the resulting compute to AI labs and hyperscalers under multi-year contracts. It is not a model developer. It does not publish weights, it does not fine-tune frontier architectures, and its competitive surface is not a developer ecosystem. Its product is bare-metal IaaS: raw accelerator hours, delivered on time, priced against scarcity.

The structural mechanics of this business are well established by now. Revenue is recognized under take-or-pay agreements — contracts in which the customer commits to pay for capacity whether or not they consume it. Capex is enormous and front-loaded. Financing is typically debt, secured against the GPUs themselves. And GAAP net income is almost always negative, because the three largest line items — GPU depreciation, interest expense, and stock-based compensation — sit below a gross margin that looks spectacular.

There is a second context that matters more to this particular publication's readership. Nscale sits at the intersection of two capital pools that have spent eighteen months converging: AI data center development and the Bitcoin mining industry. A generation of mining operators — IREN, TeraWulf, Cipher, and a long tail of smaller sites — discovered that the industrial real estate they occupied, with substations already energized and interconnection agreements already signed, could be repurposed into AI hosting capacity at valuations several multiples higher than their mining operations commanded. This has become the most reliable capital-arbitrage trade in the sector, and it is the reason a company like Nscale appears on a crypto outlet's front page at all.

So the question is not whether Nscale is real. It is whether the way it is being described — stunning — refers to margins, to losses, or to both at once. Those are not the same claim, and only one of them survives contact with a depreciation schedule.

Core: the financialization of time-to-power

Strip away the branding and Nscale's business model is a spread trade. It borrows against assets at one rate and earns under contracts at a higher rate. The asset is not the GPU in isolation. The asset is the combination of an energized facility and a signed long-term agreement, pledged together as collateral. This is closer to a data center REIT or an independent power producer than to a technology company. The distinction is not academic. REITs and IPPs are valued on cash flow and replacement cost. Software companies are valued on growth and narrative. The market, when it looks at AI infrastructure, routinely applies the second framework to a company that only works under the first.

This is where my own history with collateral structures becomes relevant. Between 2017 and 2020, I spent years reverse-engineering governance systems and stress-testing lending protocols — the 2x2 DAO deconstruction, the Aave v2 liquidation modeling. The lesson those years burned into me is that the collateral is never the asset the borrower thinks it is; it is the asset the liquidator can actually sell in the worst hour of the worst week. A GPU is not a Treasury bill. It has a depreciation curve, a fab generation cadence behind it, and a secondary market that thins to nothing the moment supply normalizes.

Three mechanics carry the entire structure. The first is power. In the United States and the United Kingdom, the queue to interconnect a new data center to the transmission grid routinely runs past five years, and in some regions past 2030. A brownfield site — a retired industrial facility or a mining operation — that already holds a high-voltage interconnection can be delivering compute in twelve to twenty-four months. That gap, called time-to-power, is Nscale's only durable moat. It is real. It is also finite, and it decays the moment new generation capacity and grid upgrades come online. Logic holds until the ledger bleeds — and this ledger bleeds against a clock.

The second mechanic is the anchor customer. Public reporting indicates Nscale signed a multi-billion dollar, multi-year compute agreement with Microsoft, tied to a UK campus running NVIDIA's GB200 systems, and participates as a compute partner in the OpenAI Stargate Norway initiative alongside Aker. These are not marketing footnotes. They convert customer concentration from a pure liability into what looks like a credit-quality advantage: investment-grade and quasi-investment-grade counterparties make GPU-backed debt cheaper to underwrite. The story writes itself — stable offtake, blue-chip tenants, a financing cost advantage over unanchored peers.

But this is also the third mechanic, and the one the optimistic reading skips. Customer concentration is not a risk that gets retired by signing a good customer. It gets deferred. CoreWeave's 2024 revenue — roughly $1.9 billion — leaned on a single customer, Microsoft, for around sixty-two percent of it. The company only diversified its client base in 2025 by adding a roughly $11.9 billion OpenAI agreement. Concentration is the endogenous property of this model, not an execution failure. When one contract is the difference between solvency and restructuring, the credit quality of that one contract is the entire credit quality of the enterprise.

Now to the line nobody quoted. Depreciation. In an AI infrastructure S-1, the accountant's choice of useful life for a GPU is the single most manipulable assumption in the document. If a management team depreciates accelerators over six years while their economic productive life — the period during which they can command premium rental rates — is closer to three, then current-period profit is systematically overstated and the eventual impairment is a cliff, not a slope. This is not a hypothetical. It is the recurring debate across every neocloud income statement ever published, and it is settled not by the contract but by the pace of NVIDIA's release cadence. Every new architecture — Blackwell, then Rubin — pushes the older generation down the rental price curve. The GPUs do not stop working. They stop being worth what the balance sheet says they are worth.

Here the crypto parallel is exact and worth stating plainly. In 2020, when I stress-tested Aave v2's liquidation incentives across more than five hundred simulated scenarios, the recurring failure mode was never the smart contract logic. It was the oracle. The contract behaved perfectly against a price that had already stopped being true. The same failure mode governs GPU-collateralized debt. The debt instrument behaves correctly against a residual value assumption that no oracle can validate in real time, because there is no real-time market for used GB200 racks. The algorithm saw the crash, not the pain. The depreciation schedule is the oracle, and it is set by the same people whose bonuses depend on it being generous.

Which brings the whole structure into focus. Nscale is, at bottom, an exercise in securities-ing the gap between a signed contract and a delivered asset. The contracts are long-dated. The demand for the compute is real today. But the revenue recognition is linear across the contract term while the cash and the depreciation front-load differently, and the backlog — the number that would tell us how much of the "stunning revenue" is already contracted versus hoped for — was left undefined in the coverage. Trust is a variable, not a constant, and here the variable is the useful-life assumption.

There is one more loop worth naming, because it appears in the structure of several of these companies and almost never in the marketing. NVIDIA is simultaneously the supplier, an ecosystem partner, and — by multiple public accounts — an investor in neoclouds. That creates a chain: the vendor invests in the customer; the customer uses the capital to buy the vendor's product; the product is pledged as collateral for debt that is repaid from contracts the vendor's own demand helps create. When demand is genuinely scarce and rising, this loop is harmless and even efficient. When demand softens, every link contracts at once, and the correlation that made the structure look diversified becomes the mechanism of its fragility. This is not fraud. It is a cycle, and cycles have a direction.

Contrarian: the blind spot is the customer, not the competitor

The consensus reading of the neocloud sector frames competition as intrasectoral — CoreWeave versus Nebius versus Crusoe versus Nscale, all fighting for the same gigawatts and the same hyperscaler contracts. That framing is comfortable because it lets everyone be a winner in a growing category. It is also wrong about where the risk lives.

The real threat to every neocloud is not the neocloud next door. It is the customer. Microsoft, Amazon, Google, and Meta are all building their own AI data centers, at capital expenditure levels no independent operator can match. The existence of the neocloud category is not a permanent feature of the market; it is a function of a specific window — the period during which hyperscaler self-built capacity could not keep pace with model-training and inference demand. Nscale's own anchor customer is also, in a very literal sense, its largest potential competitor. When self-built capacity catches demand, or when AI capex growth decelerates from its current pace to anything resembling normal, the first thing that gets cut is outsourced compute. Owned capacity is a sunk cost that keeps running. Rented capacity is a line item that gets cancelled.

This is where the crypto readership should pay attention for reasons that have nothing to do with Nscale's P&L. The brownfield-conversion trade — mine site to AI campus — has become the valuation template for an entire industry. If that template holds, it pulls capital and operator attention away from Bitcoin mining and toward AI hosting. This is a second-order effect on network hashrate that almost no one is pricing. Mining economics already compress margins toward the cost of power; if the marginal miner can earn multiples more by hosting GPUs than by hashing, rational capital leaves. The hashrate does not collapse overnight — the transition is slow and lumpy — but the direction is set by a spread that currently favors AI by a wide margin. The Bitcoin network's security budget, already argued over, is downstream of a decision made in a boardroom about whether to install ASICs or accelerators. Decentralization is a promise, not a guarantee, and the promise here is being quietly repriced by the same forces pushing Nscale toward a public listing.

There is a subtler blind spot. If Nscale genuinely carried strategic value — power assets, anchor contracts, a UK and Norwegian footprint aligned with sovereign AI policy — a hyperscaler or NVIDIA itself could simply buy it outright, likely at a price below what a post-IPO market capitalization would demand. The choice to list instead of sell suggests one of two things: either the owners believe the public market will pay a premium that no strategic buyer would, or the asset structure — affiliated-party debt, joint-venture layering, a distressed-debt heritage — is inconvenient to integrate through acquisition. Both readings are informative, and neither appears in the flattering coverage.

Takeaway: what to watch, and what it would mean

I am not forecasting that Nscale fails. I am forecasting that the resolution of its valuation will be determined by three disclosures that have not yet surfaced: the depreciation schedule and its useful-life assumptions; the definition and composition of backlog, and specifically how much of it depends on capacity that has not yet been energized; and the terms of the affiliated-party financing that traces back to its earliest backers. If the useful life is conservative — four years or less — and customer concentration is genuinely diversified, and free cash flow trends toward positive, then the structural-fragility thesis collapses and should be discarded without ceremony. If none of those hold, then the "stunning economics" referred to a margin that exists only because the depreciation clock has not yet started ringing.

The broader signal is what matters for anyone holding exposure to this cycle. When companies that own no models, publish no software differentiation, and monetize primarily the timing gap between grid interconnection and demand can reach public markets at high multiples, the market is pricing a future in which compute scarcity persists indefinitely. It will not. We coded the escape, but forgot the exit. The exit, when it comes, will be written in an accounting footnote that very few people read on the way up — and quoted by everyone on the way down. Track the depreciation line. It is the only audit that matters, and it is the one the headline declined to mention.