The Neocloud Balance Sheet: Who Owes What to Whom

A neocloud buys hardware, powers it, contracts it out and collects rent. The revenue looks recurring, which invites comparison with software, but the cost structure does not resemble software in any respect.

The correct comparison set is equipment leasing. An aircraft lessor buys an asset with a long delivery lead time, finances it with debt secured against lease contracts, earns a spread over the cost of that debt, and lives or dies on two things: the creditworthiness of its lessees and the residual value of the asset when the lease ends. Every one of those sentences describes a neocloud with the word aircraft replaced.

The differences run in both directions. Accelerators have a far shorter competitive life than airframes, which is unfavourable. They also have far shorter delivery lead times and can be redeployed between customers with far less friction, which is favourable. The financing structures used are similar enough that credit analysts who covered transport leasing have adapted to the sector faster than technology analysts have.

What follows from this framing is that valuing a neocloud on revenue growth is a category error. The business is a spread business. The relevant questions are what the debt costs, how long the asset earns, who is paying the rent and what happens when the contract ends.

Four inputs determine whether a given neocloud is a durable business or a financing structure waiting for a refinancing window, and all four are disclosed.

The first is the cost and tenor of the debt. Contract backed hardware debt in this sector has typically been written with amortisation over five to six years, which is a reasonable schedule for a facility and an aggressive one for the equipment inside it.

The second is the useful economic life assigned to the accelerators. This is an accounting estimate rather than a physical fact, and it drives depreciation, reported earnings and any covenant tested against earnings. A one year extension to assumed life materially improves every reported metric without changing a single cash flow.

The third is the proportion of capacity under contract. Contracted capacity is a receivable. Uncontracted capacity is a hope, priced at whatever the spot market for compute will bear at the time, and that market is considerably thinner than the headline demand narrative implies.

The fourth is customer concentration, which is discussed separately below because it deserves to be.

The simulator embedded here allows these to be varied directly. Setting leverage, contract coverage and demand change produces a debt service coverage figure and a break order across the participants in the chain. It is a scenario model with stated assumptions rather than a forecast, and its purpose is to make the sensitivity visible rather than to predict any particular firm's outcome.

Almost every disagreement about neocloud valuation reduces to a disagreement about how long an accelerator earns its original rental rate.

The optimistic case is well supported. Older generation hardware does not stop working when a new generation ships. It moves down the stack into inference serving, fine tuning, academic research and smaller enterprise deployments, all of which are large and growing markets. Utilisation of previous generation capacity has generally remained high. On this view a six year depreciation schedule is realistic and possibly conservative.

The cautious case is equally well supported. What matters financially is not whether the hardware runs but what rate it commands. Each new generation delivers a substantial improvement in performance per watt and per dollar, and the rental price of the previous generation adjusts to reflect that within quarters rather than years. A machine still running at seventy per cent of its original rate is not a machine that supports debt written against one hundred per cent of it.

Both positions can be true at once, and typically are. The asset remains useful for six years and earns its original economics for three. Financing written against the first fact and repaid out of the second is where the strain appears, and it appears as a refinancing conversation rather than as a default, which is why it tends to surface quietly.

The credit quality of a neocloud is not primarily a function of technology, utilisation or hardware pricing. It is a function of who signs the contracts.

A lease book underwritten by a small number of very large tenants is a counterparty credit exposure with data centre characteristics attached. If one anchor tenant renegotiates terms, defers a deployment or declines to renew, the revenue impact is immediate and substantially unreplaceable in the near term, because there is no deep spot market able to absorb that volume of capacity at contract prices on short notice.

This is why the identity of the tenants matters more than their number. A book anchored by companies generating tens of billions in annual operating cash flow is genuinely good collateral, and lenders are correct to treat it that way. A book anchored by companies whose ability to pay depends on their next funding round is a different instrument wearing the same label, and the distinction is not always drawn in the covenant package.

It is also the reason the single most consequential event in this sector would not be a default. It would be a successful renegotiation. Once one anchor tenant demonstrates that a take or pay commitment can be reopened on commercial terms, every lender holding similar contracts as collateral has to reconsider what that collateral is worth, and they will do so simultaneously.

The companion piece on circular financing sets out the general structure. The neocloud layer is where it becomes a credit question rather than an accounting one.

The chain runs as follows. A chip vendor may hold an equity position in a model laboratory. The laboratory contracts for compute capacity from a neocloud. The neocloud borrows against that contract to buy accelerators from the chip vendor. A lender advances the money.

Every link in that chain is a legitimate, disclosed commercial transaction. The aggregate structure nonetheless has a property none of the individual links has: a lender at the end of the chain has extended credit against a contract whose ultimate payer is partly funded by the company at the start of it. That lender may have performed thorough diligence on the neocloud and none at all on the funding position of the laboratory two steps removed, because that is not conventionally where a secured lender looks.

This is not an accusation of concealment. Each participant discloses its own position accurately. It is an observation about aggregation, and aggregation risk is a recurring theme in every credit cycle for the simple reason that no single participant has an obligation to describe the whole.

If this sector experienced stress, the sequence would be reasonably predictable, and knowing the sequence is more useful than estimating the probability.

The first observable would not be a default. It would be a change in financing terms: wider spreads on new contract backed debt, shorter tenors, higher equity contribution required, tighter covenants. Credit markets reprice before anything visible happens in operations, and this is the earliest reliable signal available to an outside observer.

The second would be a change in depreciation policy or an impairment charge. Any lessor shortening its assumed useful life is telling the market something about residual values that its rental rates have already told it privately.

The third would be a contract renegotiation, disclosed or leaked. As noted above this is the genuine contagion channel, because it revalues collateral across every comparable book at once rather than affecting one borrower.

The fourth would be consolidation. Weaker operators would be acquired for their power contracts and interconnection rights rather than for their hardware, because in a constrained grid the scarce asset is the megawatt rather than the machine. That outcome would be orderly and would not constitute a crisis, though it would be reported as one.

It is worth being explicit about the limits of this analysis, because the sector attracts both promotional and apocalyptic coverage and neither is useful.

Nothing here suggests that compute demand is fake. Paying enterprise usage is real, it is large, and it is growing. Nothing here suggests that neoclouds are unsound businesses. Several are well capitalised, sensibly contracted and diversified across tenants. Nothing here predicts default by any particular firm, and the analysis deliberately avoids naming one, because the structural point does not depend on any individual case.

What the analysis does establish is that this layer carries the financing risk of the entire build out on a shorter asset life than its debt schedule assumes, and that its credit quality is determined by a small number of tenant relationships. Those are facts about the structure rather than judgements about any company, and they are the reason this layer rather than the chip vendors is where an attentive observer should be reading disclosures.

Four disclosures carry most of the information. The tenant concentration note, which states how much revenue depends on how few customers. The depreciation policy and any change to it. The maturity schedule of contract backed debt against the expiry schedule of the contracts securing it, because a mismatch between those two schedules is the single most informative comparison available. And the terms achieved on the most recent financing, which is the credit market's live opinion, updated more frequently and more honestly than any analyst rating.

All four are published. None requires special access. Together they answer the question this piece began with, which is who owes what to whom, and what has to remain true for it to be repaid.