The Hidden Balance-Sheet Risk in AI Buildouts
The AI buildout's risk is not the size of the spending but how it is represented: depreciation racing obsolescence, obligations living in footnotes, revenue that finances its own customers.
Investors scrutinizing the AI buildout overwhelmingly ask one question — how much is being spent — and debate whether the number is justified. That is the visible argument, and it is largely beside the point for security analysis. Spending is not a balance-sheet risk; spending is a decision. The risk lives in how the spending is represented: which assumptions convert cash out the door into reported earnings, which obligations sit outside the debt line, and whose credit ultimately stands behind the revenue the buildout claims to generate. Two companies can spend identically and carry entirely different risks, because one has represented the spending conservatively and the other has borrowed flattery from its own accounting.
Here is the framing we use: the balance sheet remembers what the income statement is allowed to forget. Accrual accounting grants management a set of discretions — asset lives, capitalization choices, commitment disclosures, structure design — each of which can shift economic weight from the present income statement into the future balance sheet. None of these discretions is improper; all of them are directional. In a capital cycle this large, small per-dollar flattery compounds into sector-wide earnings that are systematically ahead of sector-wide cash economics, and the reconciliation always arrives. The analytical task is not to allege wrongdoing; it is to measure the gap and know which side of it each holding sits on.
The depreciation gap: accounting life versus competitive life
The largest single discretion is the useful life assigned to capitalized compute. When a company buys accelerators, the cash leaves immediately but the expense arrives on a schedule management chooses; stretch the assumed life and each year carries less depreciation, reporting higher margins on identical economics. The temptation is structural, because the sums are now large enough that a modest change in assumed life moves reported earnings materially — and because the change is defensible in the filing even when it is aggressive in the economics. The correct life for a compute asset is not how long it functions but how long it earns: the fleet is competing against successive hardware generations that improve performance per watt, and in a facility where power is the binding constraint, an older accelerator is displaced not when it breaks but when its electricity is worth more feeding a newer part. That is the depreciation gap: the distance between the schedule in the filing and the earning-power curve of the asset — and every year of gap is earnings reported now that must be un-reported later, through impairments, accelerated depreciation, or quiet margin erosion as the fleet ages.
The gap has a tell, and it is behavioral rather than numerical: watch for institutions arguing both sides of the asset's life. A company that extends depreciation schedules — asserting compute earns for many years — while simultaneously pitching its own customers on the necessity of constant refresh to stay competitive is making two incompatible claims about the same object, and only one of them flows through its income statement. Inconsistency between the depreciation narrative and the sales narrative is among the most reliable early signals that the accounting life has detached from the competitive one.
The obligations that never reach the debt line
The second discretion is structural. A buildout of this scale is increasingly financed through instruments that are economically debt-like while sitting outside the line item labeled debt: take-or-pay capacity commitments that obligate payment whether or not the compute is used; multi-year purchase obligations disclosed in footnotes rather than carried as liabilities; joint ventures and special-purpose vehicles that hold the assets and the borrowings off the operator's consolidated balance sheet, with the operator's lease or capacity contract serving as the structure's real credit support. Each structure has legitimate uses; each also shares one property — it makes the visible balance sheet lighter than the economic one. A firm whose stated leverage is modest but whose footnotes carry decade-long unconditional payment obligations has leverage; it has simply relocated it to where ratio screens do not look. The fixed-charge burden, not the debt line, is the object to underwrite.
The third discretion is the most cyclical and historically the most dangerous: circularity — arrangements in which a supplier's revenue depends on capital the supplier itself provides, directly or through the same closed loop of counterparties. A chip vendor investing in a compute provider whose purchases become the vendor's revenue; a platform committing capacity payments to a startup that spends them on the platform's own services; an ecosystem in which the same pool of capital circulates through several income statements, being recognized as revenue at each stop. Within the loop, every entity's growth is genuine on its own filing and conditional in aggregate: the system's revenue rests on the weakest balance sheet in the circle plus whatever external capital keeps the circle funded. Circular revenue is not fake revenue — the invoices are real — but its quality is that of the loop's marginal financier, and quality is precisely what the income statement does not disclose.
The illustration: how vendor financing ended last time
The telecom equipment era remains the canonical history lesson, and it deserves precise telling because its mechanics — not its magnitude — are what generalize. In the late 1990s, equipment vendors competing for carrier business began financing their own customers: extending credit so that carriers, many of them thinly capitalized challengers, could buy equipment the vendors then recognized as revenue. On every individual filing the practice looked defensible — receivables were assets, the sales were real, the customers were growing. In aggregate, the industry had quietly become its own largest source of demand: revenue growth was being manufactured by balance-sheet extension, and the quality of that revenue was the credit quality of the weakest carriers in the loop. When external funding for the carriers stopped, the loop unwound in sequence — customer defaults, receivable write-offs, then the revelation that a meaningful share of the boom's reported revenue had been, economically, the vendors' own capital making a round trip.
The generalizable lesson is not that vendor financing is inherently toxic; it is that revenue quality degrades invisibly while the loop is funded and reveals itself only when the loop breaks. Every reported number was technically accurate until the quarter it wasn't. An investor who tracked only income statements had no warning; an investor who tracked the loop — who financed whom, and whose external capital kept the circle turning — saw the fragility years early. The AI ecosystem's circular arrangements are structurally milder and the core participants incomparably better capitalized, which changes the odds but not the method: map the loop, find its marginal financier, and grade every dollar of intra-loop revenue at that financier's credit, not at the reporter's.
The strongest case against this framework
The serious objection is that this framework mistakes conservatism for accuracy. Longer depreciation schedules, the objection runs, may simply be honest: compute fleets demonstrably earn for years beyond the training frontier, cascading into inference and internal workloads, so matching depreciation to a longer earning life is better accounting, not flattery. Off-balance-sheet structures are how every large infrastructure class has always been financed — project finance is a technology, not a trick. And the circularity charge proves too much: in any dense ecosystem, major players are one another's customers and investors; drawing the loop widely enough makes all commerce circular.
Each point is partially right, and the framework sharpens rather than collapses under them. On depreciation: second lives are real, but the test is whether the residual fleet earns revenue commensurate with its carrying value — an accelerator cascaded to low-value workloads at a fraction of its original productivity justifies a steeply front-loaded earning curve, not a flat one, and depreciation should follow the earning-power curve, not the possession curve. On structures: project finance is indeed a mature technology, and the analysis it deserves is the mature one — recompute leverage with footnote obligations included, and ask whether the structure allocates risk or merely conceals it; a structure that survives that recomputation was never the target. On circularity: the test is directional dependence, not mutual commerce — the question is whether the loop's revenue would survive the withdrawal of its weakest member's external funding. Ecosystems pass that test; daisy chains do not. This is how Sterling's Embedded Intelligence reads AI-exposed filings: not hunting for villains, but repricing three specific discretions — asset lives against competitive obsolescence, stated debt against total fixed obligations, and reported revenue against the credit of the loop it travels through.
How to apply this framework
- Track free cash flow against net income through the buildout. A widening spread is the depreciation gap and capitalization discretion at work — earnings arriving ahead of cash economics. The spread itself is not damning; a spread that widens while management extends asset lives is.
- Recompute leverage from the footnotes. Add take-or-pay commitments, unconditional purchase obligations, and the debt of SPVs whose real credit support is the company's own contracts. Underwrite the total fixed-charge burden, not the line labeled debt — screens read the label; downturns read the obligations.
- Listen for the two-sided asset-life argument. Depreciation schedules lengthening while the sales narrative preaches constant refresh is the behavioral tell that accounting life has detached from competitive life. One of the two claims will eventually flow through the income statement — historically the harsher one.
- Map the loop and grade revenue at the marginal financier's credit. For any AI-exposed name, trace whether its customers' purchasing power originates inside the same ecosystem — supplier investment, platform commitments, vendor credit. Intra-loop revenue is worth the credit of the loop's weakest funded member, and the moment to know that is before external funding conditions decide to demonstrate it.