The Six-Year Fiction: AI's Equity Pushback and the On-Chain Compute Trade

Last week I read the same depreciation footnote for the fourth time this year. Servers and networking equipment: six years. Nothing in the line moved. Everything around it did. In the same fortnight, two separate people described GPU-backed lending desks quietly widening collateral haircuts, and a crypto-native outlet — not a bank research desk, a crypto outlet — published a piece on investors pushing back against AI valuations while stock sales get harder. I trace the shadow before it casts. The shadow here is not a valuation multiple. It is an accounting assumption nobody on-chain has been asked to audit. The original report is thin by design. No company named, no discount quantified, no transaction size, no date. Four information points, three of them restatements of the headline. Read as a data source, it is close to empty. Read as a signal, it is not: the migration of an institutional anxiety into a crypto-native newsfeed marks the moment a professional concern becomes retail narrative. That transition has a price, and the price is usually paid by whoever holds the marginal asset. The piece establishes two directional facts — buyers are refusing to clear at offered valuations, sellers are finding fewer exits. Both are statements about pricing power, not fundamentals. The size of the gap, whether it is confined to second-tier names, whether it touches debt as well as equity: all missing. Crypto readers should care because the AI capital cycle has already been ported onto public ledgers in forms that inherit its weakest assumptions. Tokenized GPU fleets. Neocloud tokens. Compute-yield products paying a coupon out of rental income that has not yet been contracted. Real-world-asset wrappers around data-center leases. None of this is exotic; it sits in the same wallets as stablecoin yield and is marketed in the same language — real yield, productive collateral, cash-flow backed. The underlying arithmetic deserves stating plainly. AI capital expenditure runs in the hundreds of billions annually across the hyperscalers, while attributable revenue from models and applications remains an order of magnitude below that. The gap is not fraud. It is a bet, and every bet has a duration. The auditor's question is not whether the bet is right. It is who carries the loss if the duration is wrong, and whether that party has been told. The schedule is the thesis Consider one accelerator. Call it thirty thousand dollars, roughly the all-in cost of a high-end training part with its host and interconnect. Depreciated over six years, that asset consumes five thousand dollars of annual expense. Depreciated over three, ten thousand. Double the annual charge, same hardware, same revenue. Multiply by a fleet of a hundred thousand units and the swing is five hundred million dollars a year of reported earnings that exists or does not exist depending on a footnote. That is the arithmetic foundation of the entire valuation pushback, and it is almost never the number being argued about. Investors argue about multiples, about growth, about whether agents will finally monetize. The multiple is downstream. If the economic life of the hardware is closer to three or four years than six — and the people who run these fleets, rather than the people who finance them, generally say it is — then the earnings base everyone is applying a multiple to is overstated, by an amount that grows with every capex dollar deployed. I have watched this pattern from a different chair. In 2022 I spent three months reverse-engineering the UST mechanism, building a simulation that showed the incentive structure was fragile independent of market sentiment. The point was never that the market was wrong about Luna. It was that the fragility was structural — visible in the code, invisible in the narrative. Whether AI equities are expensive is a sentiment question. Whether the depreciation schedule is defensible is a structural question, and structural questions have answers. From equity to credit, and from credit to collateral When equity gets expensive, capital expenditure does not stop. It changes shape. The financing mix rotates from operating cash flow and equity issuance toward debt, project finance, joint ventures, and special purpose vehicles holding the asset off the sponsor's balance sheet. This rotation is well underway in AI infrastructure, and it changes the constraint set entirely. A capex plan funded by equity is constrained by venture risk appetite, which is sentiment. A plan funded by debt is constrained by loan-to-value ratios, collateral haircuts, and tenor, which are arithmetic. Arithmetic reacts to bad news non-linearly. This is where the crypto analogue turns uncomfortable rather than merely interesting. Tokenized compute products are, structurally, leveraged holders of depreciating assets funded with short-dated liabilities. The depositor's claim is redeemable on demand or at short notice. The asset behind it has a three-to-four-year productive life and a residual value nobody has independently marked. That is a maturity mismatch dressed as real yield, and I have been consistent about this shape of product: it performs beautifully in expansion and resolves first in contraction. The coupon is not a lie. It is priced off a rental rate and a residual value both set during the tightest supply conditions this market has ever seen. High-end GPU spot rental has already come off its scarcity peak while contract pricing holds. Supply keeps arriving. Power and interconnection, not silicon, are now the binding constraint on new capacity, which cuts both ways: scarcity supports rents, but delivery delays push revenue recognition rightward and lengthen the capital lock-up. Every quarter of delay is a quarter of depreciation booked against an asset that is not yet earning. What proof of reserves cannot prove Here is the gap I find genuinely under-discussed. The on-chain compute sector has adopted proof of reserves as its credibility primitive. It proves the GPUs exist. It does not prove what they are worth, how long they will remain worth it, or whether the lease revenue backing the coupon is contracted or projected. Existence is the easy claim. Useful life is the hard one, and no merkle root attests it. In 2020 I wrote a Python harness that fired ten thousand simulated arbitrage attacks at Curve's stableswap invariant to test whether the model held under adversarial flow. It held because the invariant was public and could be attacked from every angle. That is what made the design trustworthy — not the team's confidence, but the fact that anyone could reconstruct the failure conditions. There is no equivalent for a GPU fleet. The residual curve is private, the rental book is private, the depreciation policy is a management election. When the only public artifact is a hash of a machine serial number, the market is pricing a story and calling it a balance sheet. My 2025 work on agent security sharpened this. We built a verification layer that forced human approval before an autonomous agent could commit a high-value on-chain action, because hallucinated parameters were reaching execution. The lesson generalizes: autonomous capital allocation without a checkpoint is not efficiency, it is an unmonitored attack surface. A treasury agent that can roll over a compute-backed facility, or post additional collateral, or approve a counterparty, is exactly that surface. Most designs I have reviewed do not log the decision boundary, let alone gate it. The reflexivity loop, and the fragmented market that hides it There is a second-order structure the equity story only hints at. Chip suppliers invest in model companies. Model companies commit to cloud providers. Cloud providers buy from chip suppliers. The loop is real and not sinister; it is how strategic capital works. But loops amplify in both directions. A valuation markdown reduces financing capacity, which reduces purchase commitments, which reduces supplier revenue expectations, which reduces the supplier's equity value, which marks down the investment it holds in the model company. The crypto version is shorter and faster: issuer funds market maker, market maker provides liquidity to the lending pool, the pool accepts the issuer's token as collateral. Same topology, fewer circuit breakers, no quarterly cycle to slow it down. I model three outcomes — layered clearing, a credit event, and a reacceleration — and only the middle one is a break rather than a repricing. It is also the least priced. And the market that would need to price all of this is not one market. Compute liquidity is scattered across a dozen chains, each with its own wrapped representation, its own oracle, its own shallow book. I have argued for years that additional interoperability does not solve fragmentation; it institutionalizes it. Every new venue adds a quote without adding depth. A residual value curve requires concentrated, continuous trading to form. Split one asset across twelve ledgers and you get twelve noisy prints and no curve at all — which means the collateral behind these loans is marked by whoever is most optimistic, in the venue where nobody is watching. Finding the pulse in the static, in this sector, means reading the marks rather than the marketing. The conflation, and the question nobody asks The report's central flaw is a conflation worth naming precisely, because it is the flaw most readers will inherit. Overvaluation is a fundamentals judgment. Illiquidity is a plumbing judgment. If AI-adjacent stock sales are failing because rates moved and risk appetite tightened, the thesis is untouched — price corrections happen while fundamentals improve, as they did in the late nineties before the thesis was actually tested. If sales are failing because buyers did the depreciation arithmetic and did not like the answer, that is a different event. The piece does not distinguish them, and without a discount figure, a transaction size, or a date, neither can anyone reading it. The blind spot that follows is subtle. Everyone audits the model. Nobody audits the schedule. Every review I have been part of — the 2017 Crowdsale audit where an integer overflow in the distribution logic would have drained a treasury, the Art Blocks entropy review where a block-hash dependency made a seed guessable — began from the assumption that the interesting failure lives in the assumption layer, not the implementation layer. On-chain compute has the same shape. The contracts are usually fine. The economic life of the collateral is where the loss is hiding, and it hides in plain sight because it is an input, not a bug. Vulnerability is just a question unasked, and the unasked question here is simple: who marks the residual value, and against what standard? I am not calling a top. The structural read is layered, not systemic: well-capitalized players with strategic or sovereign backing keep spending through a drawdown, while the marginal levered compute holder and the AI-adjacent token with no cash flow get repriced hard. The signal to watch is not a multiple. It is a footnote. Watch whether any major operator shortens its server depreciation schedule; that single line, if it moves, reprices the complex. Watch pre-leasing rates on capacity under construction, GPU spot rental prints, and haircuts at the lending desks. Logic blooms where silence meets code, and right now the silence is in the schedule. When the first compute-backed pool posts a mark-to-model writedown, will its depositors have known which clock was ticking?