A Footnote, Not a Headline
In the final week of a quarter I will not date precisely, a mid-tier compute provider — the sort of company that owns racks rather than models — filed an amendment to a credit agreement that had already been amended twice. The headline figure was unremarkable: a nine-figure delayed-draw term loan, arranged by a private credit fund and syndicated to three insurers and one pension allocator. What held my attention was the collateral schedule. It did not list real estate, receivables, or patents in the ordinary sense. It listed accelerator clusters by generation, by interconnect topology, and by contracted utilization, with a residual value curve attached as an exhibit.
Then footnote fourteen. The useful-life assumption attached to those clusters was six years.
I have spent enough time inside reserve attestations to know that the interesting number is rarely the one on the cover. In 2022 I sat with two cryptographers auditing proof-of-reserves for three stablecoin issuers, and the discrepancy that eventually mattered was a fifty-million-dollar gap buried in the treatment of short-duration instruments reclassified between reporting dates. The failure was not fraud in the theatrical sense. It was a definitional choice, made once, repeated every month, and never revisited because revisiting it would have been expensive.
Footnote fourteen is a definitional choice. Six years for an accelerator is not absurd on its face. It is also not obvious. And when the same physical asset class carries a three-year useful life at one issuer and a six-year useful life at another, the market is not comparing two valuations. It is comparing two accounting philosophies and calling the difference a multiple.
This is the part of the artificial intelligence buildout that the crypto press reports as a growth story — firms expanding headcount, firms securing complex financing, valuations rising — and that credit desks read as something else entirely. Tracing the silent hemorrhage of algorithmic trust begins with a footnote, because that is where the hemorrhage always starts.
The Financing Toolbox Outgrew the Equity Market
Complex financing is not glamour. It is what happens when a single capital requirement exceeds the risk appetite of any single equity investor.
Consider the arithmetic. A frontier training cluster plus the land, the substation, the cooling loop, and the two-to-three-year construction runway now represents a capital outlay that a traditional growth-equity fund cannot underwrite alone without breaching concentration limits. The instrument has to be split. That splitting is what the trade press calls "complex financing," and the phrase hides a fairly mechanical set of tools.
Special purpose vehicles ring-fence data center and compute assets away from the parent balance sheet, which allows the parent to raise against the asset without consolidating the full liability in the way a direct borrowing would. Debt secured by accelerator clusters and by the contracted revenue those clusters are already committed to produce lets a company borrow against future invoices. Vendor financing lets a chip designer or a cloud operator extend credit or strategic equity to a customer who will then spend that capital on the vendor's own product — a structure that is circular on inspection and perfectly legal in practice. Compute-for-equity swaps let a provider take a stake in a customer in exchange for capacity rather than cash. Private credit funds, the large alternative managers with infrastructure debt strategies, have moved into this space as buyers of the senior tranches. And the term sheets themselves have grown hybrid: delayed-draw facilities that fund in stages as milestones are met, payment-in-kind toggles that let the borrower defer cash interest at the cost of accruing principal, preferred tranches with liquidation preferences stacked above common equity, and revenue-sharing agreements that convert a slice of future top-line into a current liability.
None of this is novel machinery. It is the standard toolkit of capital-intensive industries — shipping, aviation, telecom towers, LNG terminals. The reason it matters here is the historical rhyme. In 1999 and 2000, telecom equipment vendors financed their own customers so those customers could buy more equipment. The revenue was recognized, the receivables were real on paper, and the entire chain was circular in a way that only became visible when the marginal customer stopped being able to service the debt. The structure did not create demand. It accelerated the recognition of demand that was already priced in.
What is genuinely different this time is where the risk now sits. When a company funds expansion from equity, the risk is marked every day by people who can sell. When the same expansion is funded through private credit, insurance balance sheets, and structured SPVs, the risk is marked by an appraiser, on a schedule, against an assumption. The migration from equity to debt does not reduce the risk. It changes who is allowed to notice it, and how often. That is the entire mechanism, and it is unremarkable until it is not.
The ledger does not sleep, it only waits.
The Labor Data Contradicts the Sales Pitch
There is a second signal in the same story, and it is the one I find more diagnostically useful than the financing.
AI companies are expanding headcount. Substantially. This is not a rounding error or a footnote in a hiring report — it is the structural counterweight to the entire public thesis that these firms exist to eliminate labor. The sales pitch says substitution. The payroll says something else.
When I pull apart where the hiring is actually landing, the pattern is consistent and unglamorous. The growth is not concentrated in frontier research. It is concentrated in go-to-market functions and customer success, in solutions engineering, in data annotation and quality evaluation, in safety and alignment review, in inference infrastructure operations, and — the part that almost never makes it into a strategy deck — in the physical trades: electricians, HVAC specialists, construction crews, substation engineers, and the logistics people who move extremely heavy equipment into places where the grid connection does not yet exist.
This tells you something specific about where value capture has migrated. When model capability converges — when the gap between the best and the fourth-best system on the standard benchmarks collapses into single-digit percentages — the differentiating variable stops being the model. It becomes the ability to insert that model into a customer's existing workflow, to manage the political resistance inside that customer's organization, to guarantee a service level, and to absorb the integration cost. That is delivery work. Delivery work is labor.
The corollary is uncomfortable for people selling the substitution narrative. A meaningful share of the AI value chain is still labor-intensive, and the labor is not the kind that automates itself. Annotation has shifted from general-purpose tagging toward domain-expert evaluation in law, medicine, finance, and low-resource languages. Preference data for reinforcement learning from human feedback requires humans whose judgment is defensible. Synthetic data reduces the demand for cheap labeling and increases the demand for expensive verification, because somebody has to confirm that the generated corpus is not quietly drifting.
There is also a less charitable reading of the hiring data, and in a late cycle it deserves airtime. Some of the expansion exists to demonstrate momentum to investors who are underwriting a growth curve. Narrative hiring is cheap to announce and expensive to unwind. It shows up first as a slide in a deck and last as a line item in a restructuring plan.
And here is the part that should worry anyone holding the equity: headcount is a contractual outflow, while compute revenue is a variable inflow. Sales capacity and delivery teams are hired against a revenue forecast. If that forecast slips by two quarters, the cost does not slip with it. That is negative operating leverage wearing a growth costume.
Anatomy of Compute-Backed Credit
Let me be precise about the collateral, because the details are where the discipline lives.
A loan secured by accelerator clusters rests on three assumptions stacked together. First, that the hardware retains enough residual value over the debt's life to satisfy the coverage ratio at any point where the lender might need to seize it. Second, that the counterparties holding the contracted utilization contracts do not default or renegotiate. Third, that the spot market rental rate for equivalent compute does not fall below the level implied by the contracts currently in the collateral pool.
Each assumption is defensible individually. The problem is their correlation. These three assumptions fail together, and they fail in the same direction, for the same reason. A deterioration in end-customer demand pressure reduces utilization, which pushes rental rates down, which pressures the negotiated contracts on renewal, which forces a write-down of the residual value curve — and by that point the collateral coverage ratio that justified the original loan-to-value is a historical document. There is no diversification here. It is one bet expressed three ways.
The second structural feature is contract quality. Utilization that is contracted with an investment-grade hyperscaler for four years is a fundamentally different asset than utilization that is merchant — sold on the open market, month to month, priced at whatever the spot rate clears. Lenders know this and price it accordingly, but the disclosure at the borrower level frequently blends the two into a single "utilization" percentage. It is the modern equivalent of the reserve report I dissected in 2022: an aggregate figure that is technically accurate and materially misleading.
The third feature is counterparty concentration. If a substantial fraction of the contracted revenue behind a collateral pool flows from two or three customers, then the loan is not secured by hardware. It is secured by the continued solvency and continued strategic commitment of a small number of buyers — and strategic commitment is the first thing that gets reassessed in a budget cycle.
My work on stablecoin reserve transparency taught me a habit I have not been able to break: when a balance sheet presents an aggregate, find the definition underneath it. Reserve reports and compute-collateral schedules fail the same way. Not through invented assets, but through convenient definitions applied consistently enough that nobody re-tests them until the underlying conditions change.
Depreciation Is the Real Monetary Policy of AI
Footnote fourteen deserves its own section because the useful-life assumption is the single most powerful discretionary lever in AI infrastructure accounting, and it is almost never discussed in the same breath as valuation.
Here is the mechanism, stripped of accounting jargon. An accelerator cluster is capitalized as an asset. Each period, a portion of that asset is expensed as depreciation. Depreciation is a non-cash charge, so it is added back to arrive at the operating cash flow figure that lenders and equity analysts most often quote. The size of that add-back is determined entirely by how many years management asserts the asset will be economically useful.
Stretch the assumed life from three years to six, and you halve the annual depreciation charge. Reported operating margins improve. Reported return on invested capital improves. The add-back that supports the debt service coverage ratio gets larger. And none of this requires a single additional dollar of revenue.
The selection is not random. Companies with heavy debt service and thin coverage have a structural incentive to extend useful lives. Companies sitting on cash and reporting to public markets under intense scrutiny have less room to do so, and often choose shorter lives precisely because it is defensible. Which means the two populations are not comparable on a like-for-like basis, and the valuation multiples the market applies to them are comparing apples to a different fruit with the same label.
When I was building the liquidity-trap model in 2020 — four hundred hours backtesting early Ethereum pools against Treasury bill yields — the whole exercise was about stripping token emissions out of headline yield so I could see the real number. Substituting one inflated component for another is the oldest trick in yield analysis. Depreciation policy is exactly that, moved from the revenue line to the expense line. If you are valuing an AI infrastructure company on EBITDA, you are valuing it on a number that management selected.
The monitoring signal is dispersion. When useful-life assumptions across a sector begin to diverge — three years here, five there, six somewhere else — the divergence itself is the signal. It means the accounting is being used as a competitive instrument rather than a description of reality, and the sector-level aggregate is no longer meaningful.
Tenor Mismatch and the Thin Secondary Market
The quietest structural flaw in this whole edifice is the mismatch between how long the debt lasts and how long the asset lasts.
Compute-backed facilities are commonly structured with five-to-ten-year tenors, because that is what institutional credit investors want and what the amortization schedule requires to keep annual debt service manageable. The economic life of the collateral, under most reasonable assumptions, is three to four years before it becomes a second-tier asset in relative performance terms, and perhaps five to six before it becomes genuinely uneconomic to operate. That gap is not a rounding issue. It is a refinancing obligation disguised as a maturity schedule.
In aviation, this problem is solved by a remarketing industry that has spent fifty years building the documentation, the appraisal standards, the buyer networks, and the legal infrastructure to move aircraft between operators. A narrowbody aircraft has a global market, published valuation benchmarks, and a predictable depreciation curve anchored to maintenance cycles and fleet economics.
Accelerators have none of that. There is no mature secondary market with published clearing prices and standardized condition assessments. The buyer universe is small, geographically constrained, and in several jurisdictions legally restricted from purchasing the hardware at all. Export controls introduce a value cliff that has nothing to do with the asset's performance and everything to do with policy. Obsolescence in this asset class is not linear — it arrives in steps, each tied to a product generation launch, and the step function is driven by a schedule that the collateral holder does not control and cannot observe in advance.
So the lender's position is: a five-to-ten-year claim against an asset whose value steps down on someone else's product roadmap, in a market with no depth, subject to a policy regime that can remove buyers overnight.
That is not a reason the loans are unsound. It is a reason the risk premium should be wider than it currently is, and a reason the refinancing wall matters more than the coverage ratio. When I designed the AI-agent economy model in 2026 — ten thousand autonomous agents executing micro-transactions, a modeled two million dollars of daily volume — the hardest part was never the game theory. It was the settlement layer's assumptions about finality and collateral. The same problem, scaled down.
Where Crypto Actually Enters
I have spent enough years watching real-world-asset narratives to be immune to most of them. So let me be specific about the channels where blockchain infrastructure is genuinely load-bearing in the AI economy, and where it is decoration.
Settlement is real and almost entirely unromantic. Cross-border compute invoices, GPU leasing arrangements, and annotation payouts to distributed workforces are increasingly cleared in stablecoins. The reason is not ideology. It is that a labeling contractor network operating across a dozen jurisdictions can be paid in a dollar-denominated instrument in minutes rather than in correspondent banking timeframes, without maintaining twelve local banking relationships. The volumes here are substantial and the motivation is operational. This is the version of the RWA thesis that actually works, and it works precisely because nobody is trying to be clever.
On-chain collateral registries and tokenized credit are real but narrower than advertised. There is genuine value in a shared, timestamped, append-only record of SPV reporting, collateral schedules, and amortization events — the kind of record that lets a senior tranche holder verify that the schedule they were promised is the schedule being executed. The catch is structural: the institutions doing this do not want a public chain. They want a permissioned ledger with whitelisted participants, a regulated transfer agent, and a compliance wrapper. The public chain is used as a database with better audit properties, not as a trustless settlement layer. That is a legitimate improvement in accounting transparency. It is not a disintermediation. The intermediary remains; it is now auditable.
Decentralized compute markets matter as a price signal, not as a capacity competitor. A distributed GPU marketplace's real contribution is that it publishes a continuously observable rental rate for a given accelerator class. Against a backdrop where most compute pricing is negotiated privately and disclosed only in aggregate, a live spot reference — even one covering a trivial share of global capacity — is a monitoring instrument. I care far less about whether these networks capture meaningful share than about whether they give me a daily print on where marginal rental rates are heading. In an opaque market, the value of a small transparent venue is informational, not economic.
And then there is the speculative layer, which in a bear market deserves a hard look. Token launches wrapping AI agent narratives, agent-to-agent payment tokens, and compute-themed assets with no verifiable revenue have been the dominant speculative expression of this cycle. This is where the bear market rules apply with no exceptions: survival outranks gain. If a protocol's revenue depends on the continuation of a narrative rather than the delivery of a service, its drawdown is not a buying opportunity. It is a countdown.
One Marginal Buyer
Everything above describes microeconomics. The macro overlay is where the timing lives.
In 2025 I ran an eighteen-month regression linking spot Bitcoin ETF inflows to changes in global broad money supply, and the result that held up under repeated specification was a fourteen-day lag between liquidity expansion and price appreciation, once I controlled for regulatory hedging behavior around the flow data itself. The coefficient was not enormous. The consistency was what mattered.
The lesson I took from refining that model through six iterations is that market participants systematically overestimate the number of marginal buyers. There is a pool of risk capital that moves between asset classes based on relative expected return and available liquidity. When central bank balance sheets expand and the reverse repo facility drains into the system, that capital needs a home and it does not care about narrative boundaries. When liquidity contracts, that capital leaves the highest-duration, highest-beta expressions first. It does not distinguish between a neocloud's equity, an AI infrastructure bond fund, and a large-cap digital asset. It treats them as the same trade with different tickers.
This is the relevant frame for the AI financing story. The same allocators funding compute-backed private credit are the allocators with digital asset exposure. The same liquidity conditions that permit a nine-figure delayed-draw facility to price at a tight spread are the conditions that permit a broad crypto bid. Liquidity is a ghost; solvency is the body. The ghost moves first and moves faster, and it moves both markets simultaneously.
Which means that the crypto market's exposure to the AI capital cycle is not a narrative correlation. It is a funding correlation, and it runs through the same balance sheets.
The Recoupling Nobody Priced
The consensus view, repeated across every research desk I read last quarter, is that AI and crypto have decoupled. The argument is intuitive: different use cases, different investor bases, different regulatory treatment, and a visible divergence in performance during several stretches of the past eighteen months.
I think the consensus is reading the middle of the distribution and calling it the distribution.
The decoupling is real in the body of the return series and false in the tail. Two assets funded by the same marginal pool of risk capital will appear uncorrelated across ordinary weeks, because idiosyncratic flows dominate at that frequency. They will converge violently in a liquidity event, because the correlation that matters is not return correlation but funding correlation, and funding correlation is only observable when funding is scarce.
Here is the specific transmission path I keep coming back to. AI infrastructure credit is now held by insurers, pension allocators, and private credit funds. Those same institutions hold digital assets through a mixture of direct exposure and fund vehicles. If compute rental rates decline and collateral coverage ratios deteriorate, the affected holders do not sell their worst-performing position. They sell their most liquid one. In a risk-off episode, that is frequently the digital asset book, because it clears in hours rather than in weeks. The crypto drawdown arrives first, looks crypto-native, and is in fact a transmission of a credit event that originated somewhere else entirely.
That is the recoupling. It does not show up as correlation. It shows up as sequence.
There is a secondary version of this worth flagging. If AI structured credit does experience a correction, the most likely outcome is not a clean repricing. It is a delay — extensions, covenant waivers, amended amortization schedules, PIK elections, and quiet amendments. Which brings back the footnote I opened with. The amendment count on that credit agreement was already three. Each amendment bought time. Each bought time at a price, paid in accrued principal and weakened protections.
Code is law, but humans write the loopholes. Financial covenants are code. Amendments are the loopholes. And the loopholes are always easier to write at the top of a cycle than at the bottom, because at the top there is still something to negotiate over.
Complex Financing Is a Symptom, Not a Skill
I want to close the analytical loop on the specific claim that complex financing pushes valuations higher, because that framing quietly reverses cause and effect.
Structures are not invented because they create value. They are invented because a simpler structure could not clear at the price the seller wanted. When equity alone cannot fill a capital requirement at an acceptable valuation, the solution is to slice the claim: senior tranches with modest returns and strong protections, mezzanine tranches with higher returns and weaker ones, and common equity at the bottom absorbing whatever remains after the preferences are satisfied. Each slice finds a buyer. The aggregate headline valuation looks higher than any single equity round could have achieved.
That headline is not the value of the common equity. It is the value of the whole capital structure, expressed as though it belonged to the residual claimant. Liquidation preferences, ratchets, participation rights, and PIK accrual all sit ahead of common. A rising headline valuation and a falling common-equity value are not contradictory outcomes. They are the same transaction described from two different positions in the stack.
The second convention worth interrogating is the word complex itself. Complexity in a capital structure is neither sophistication nor innovation in any meaningful sense. It is what happens when the required capital exceeds what the market will fund simply. In capital-intensive industries with long-lived, fungible, well-understood assets, complexity is a mature market functioning normally. In an industry where the underlying asset steps down in value on someone else's product roadmap and has no remarketing infrastructure, complexity is a way of distributing an exposure that nobody wants to hold whole.
Which is not automatically a bad thing. Distribution is how large risks get funded. But it does mean that the appearance of a sophisticated financing market should not be read as evidence of underlying soundness. The sophistication is in the paper. The soundness is in the terminal return on invested capital of the end customer, and that number is currently the least observable variable in the entire system.
What to Watch, and Where the Cycle Sits
I do not think this ends in a single dramatic event. I think it ends in a long sequence of amendments, waivers, and quiet reclassifications, followed by a recognition that the useful lives were wrong, arriving roughly two years after the useful lives were wrong.
The instruments I would watch, in rough order of lead time, are these. Secondary accelerator transaction prices, where they exist at all, because they are the only unbiased read on the residual value curve that underpins every collateral schedule in the sector. Published spot rental rates from any venue transparent enough to publish them, because the direction matters more than the level. The dispersion of useful-life assumptions across issuers, because divergence is the tell. The share of a borrower's revenue that originates from entities that are also investors in that borrower, because circularity is the mechanism that makes the whole chain look healthier than it is. And the refinancing calendar, because a tenor that outlives its collateral does not create a maturity problem at origination. It creates one at the wall.
The bitcoin ETF flow regime remains the fastest-moving input I track, with its fourteen-day transmission lag into price, and it will continue to function as the canary for the funding correlation described above. If flows turn structurally negative while AI infrastructure credit spreads are widening, the sequence has begun, and no amount of narrative separation will hold two books funded by the same balance sheet apart.
We are, I think, in the part of the cycle where the structure works extremely well and the assumptions have not yet been tested. That is not the dangerous part. The dangerous part is the month afterward, when everyone discovers that the assumptions were never the same, that the definitions were chosen rather than derived, and that the loans were always backed by something that had no market except the one that was about to close.
Designing the cage to see how the bird flies is a legitimate research method — until you need the bird to survive outside it.
The question is not whether AI infrastructure is overbuilt. The question is who is holding the residual claim when the useful life turns out to have been three years all along, and whether they know it yet.