JPMorgan Says the AI Selloff Is Easing. The GPU Depreciation Clock Says Otherwise.

On a Tuesday in mid-October, a JPMorgan cross-asset strategy team led by Nikolaos Panigirtzoglou circulated a note carrying one load-bearing claim: the AI stock selloff is easing. The evidence was threefold and entirely positional. Investors have largely unwound their large bets on semiconductor equities. Hedge funds and trend-followers have finished de-risking. And the bank's aggregate equity positioning gauge has fallen sharply from its mid-September peak.

Read it twice. The conclusion is about price. The evidence is about who still holds inventory. Nowhere in the note — at least in the version that reached the wires — is there a single number about the thing the selloff was allegedly about: financing.

That is the shape of this moment. Not a crash, not a recovery. A sideways tape in which the most consequential argument in technology capital allocation is being adjudicated by a positioning chart. Another rug pull? Or just another myth? Neither. Something quieter and more structural is happening: the AI trade has changed its question, and most of the market is still answering the old one.

The old question was whether AI had a killer app. The new question is who pays for it, and at what cost of capital. Those two questions do not share a chart.

Context matters here, because Panigirtzoglou's desk is not a fringe voice. JPMorgan's cross-asset strategy work is one of the few sell-side products that institutional allocators treat as a primary input rather than a courtesy read. When that desk says a selloff is fading, it moves risk budgets in Geneva, Zurich, and Singapore before the note finishes its first pass through the terminals. So the framing deserves scrutiny, not deference.

The backdrop is a capital expenditure cycle with no modern precedent in software. Hyperscalers have committed to hundreds of billions of dollars in data center and accelerator buildout, and the commitment curve has not flattened. What has changed is the funding mix underneath it. In 2023 and 2024, the debate was demand-side: were enterprises actually adopting this, or were we watching a demo economy with a procurement department? By late 2025, the debate had migrated to the liability side of the balance sheet. Bond issuance. Private credit. Structured vehicles. Vendor financing between suppliers and customers. The question stopped being 'does this work' and started being 'who is underwriting it.'

That migration is not a detail. It is the whole story. A capital cycle that moves from internal cash flow to external debt acquires a credit cycle's properties — longer transmission lags, quieter failure modes, and a much wider gap between the moment a problem exists and the moment it becomes visible in a price.

And this is precisely where the crypto-native reader should stop scrolling. The on-chain compute sector — decentralized GPU marketplaces, DePIN networks, tokenized clusters, restaking against hardware — is not a parallel universe to this argument. It is the highest-beta expression of it. When the marginal cost of debt rises, the marginal buyer of a tokenized compute allocation does not merely get cautious. It vanishes, because it was never buying compute. It was buying duration.

Start with the accounting, because the accounting is where this trade is most fragile and least examined. Hyperscalers depreciate servers and GPUs over five to six years. AI accelerators leap a full generation roughly every eighteen to twenty-four months. If the real economic life of the asset is shorter than the accounting life, then current earnings per share are systematically overstated — not fraudulently, but structurally, because depreciation is an estimate dressed as a fact.

The most honest depreciation test in this market is the second-hand GPU price, and no sell-side note on AI financing has cited it. That absence is not an oversight. It is a selection. Second-hand accelerator pricing is the one indicator that cannot be smoothed by accounting policy, and it is the one indicator that tells you whether the five-year schedule is a reasonable assumption or a fiction of convenience.

When I reverse-engineered the Solidity contracts behind the Zeppelin security library in 2017 — while ignoring, I should admit, three weeks of assigned bug-fixing tickets — the lesson that stuck was not about reentrancy. It was that you learn a system by reading its assumptions, never its outputs. The outputs are downstream of choices someone made and hoped you would not inspect. AI capex earnings are an output. Depreciation schedules are the assumption.

Now overlay the funding structure. The shift from operating cash flow and equity to debt, private credit, and structured special purpose vehicles imports something the AI equity narrative was never built to carry: a maturity wall. Investment-grade issuance from hyperscalers. Asset-backed structures against data center cash flows. Joint ventures with private capital where the risk sits off the sponsor's balance sheet. GPU-collateralized debt at compute startups. And the circular arrangements — supplier invests in customer, customer commits to purchase, commitment supports the valuation that supports the financing — that make revenue quality genuinely hard to audit from the outside.

The consequence is uncomfortable for anyone holding an opinion: because structured financing moves risk from the income statement to a joint venture, the question 'is the financing concern exaggerated' becomes technically un-falsifiable using public filings. An unfalsifiable claim should never be a position-sizing input. Yet that is exactly the epistemic status of the note's central conclusion.

There is also a semantic trap buried in the wire copy, and it is the single largest directional risk in the entire story. The phrase 'investors have largely unwound their large bets on semiconductor stocks' can mean two opposite things. If it means long exposure was closed, crowding has been released and the bullish reading holds. If it means shorts were covered, then bearish energy has been spent without converting to bullish positioning — a neutral-to-bearish configuration dressed in the same words. The paraphrase does not resolve the ambiguity, and the ambiguity determines the sign of the conclusion.

Then there is the missing denominator. The note reports that the positioning gauge fell 'significantly' from its mid-September peak. Fell to what? From the 99th percentile to the 85th, or back to the median? Composed of CFTC positioning, fund flow data, and volatility-target weights in what proportions? A decline of how many standard deviations? Without a percentile, 'limited downside' cannot be tested and cannot be refuted. In methodology terms, that is not a forecast. It is a mood with a Bloomberg terminal attached.

Code speaks, but culture listens. And what the culture of this market is listening to is not the percentile. It is the tone. A top-tier bank publicly reassuring clients that a financing concern is overstated performs a social function before it performs an analytical one: it gives institutional holders a reason to keep holding during a drawdown. That is a legitimate service. It is also not research in the falsifiable sense.

The conflict layer deserves its own paragraph, because it is structural rather than personal. JPMorgan announced a decade-long, roughly $1.5 trillion 'safety and resilience' investment initiative that reportedly includes substantial equity and debt financing arrangements directed at AI and frontier technology companies. If those figures hold against the official filings — and they should be verified rather than assumed — then the institution publishing 'AI financing concerns are exaggerated' is simultaneously an underwriter and a capital provider inside the very financing chain it is describing. That does not make the analysis wrong. It makes it a stakeholder signal that requires a discount, the same way you discount a founder's tweet about their own token.

The physical constraint has also moved, and this is where the crypto compute narrative and the hyperscaler narrative share a blind spot. The binding constraint is no longer 'do we have GPUs.' It is 'do we have power and an interconnection agreement.' Transformer lead times, grid queue duration, and capacity market clearing prices now determine delivery dates more than silicon allocation does. That stretches the gap between cash outflow and revenue recognition, which is precisely the gap that debt financing has to bridge. A longer bridge is a more expensive bridge, and the bridge is now the product.

The Cassandra complex is real. I spent the 2020 DeFi summer publishing threads about yield mechanisms that could not survive a rate change, and I was early enough to be dismissed and right enough to be unfollowed. So let me be precise about what the Cassandra complex actually predicts here: it is not that the bearish case wins. It is that the sell-side's first systematic denial of a risk tends to arrive while the risk is still unpriced, not after it has cleared. Denial is a timing signal about consensus, not a verdict about fundamentals.

Here is the contrarian turn, and it cuts against my own industry. Decentralized compute is routinely pitched as the resilient alternative to hyperscaler capex — the permissionless, demand-priced counterweight to a concentrated buildout. That framing is backwards. On-chain compute is the leveraged expression of the identical trade: it marks to the same second-hand GPU prices, it borrows against the same collateral with thinner liquidity, and its treasuries are denominated in tokens that correlate to the same risk appetite that funds the hyperscaler bond. When the credit cycle turns, decentralized compute is not the hedge. It is the amplifier, and it is first in line at the liquidation window.

Which brings us to the reflexivity loop that the note does not mention at all. Falling AI equities close the equity financing window for AI-native companies. Those companies have multi-year compute purchase commitments. When their cash position tightens, the commitments get renegotiated or defaulted, and cloud revenue guidance follows. That loop runs from the stock price to the revenue line, not the other way around — and in the tokenized compute market it runs faster, because the collateral is the token and the token is the equity.

NFTs aren't art; they're anthropology. Positioning charts aren't valuation; they're anthropology either. They tell you what a tribe believes about itself. The tribe currently believes the AI drawdown was mechanical. If the drawdown was instead a repricing of the discount rate applied to a decade of capex, then a flat positioning gauge is irrelevant, because DCF denominators do not care who sold.

The practical read is narrower than the headline suggests. Take the data point and discard the conclusion. A positioning gauge that peaked in mid-September and has since fallen is genuine information about near-term volatility — mechanical sellers are fewer, so the tape is thinner and more reflexive. That is useful for execution. It is worthless for valuation.

So watch credit, not equities. AI-related credit spreads. New-issue subscription ratios on data center paper. Private credit exposure to data center SPVs. Loan-to-value assumptions on GPU-collateralized facilities, and what a thirty percent markdown in second-hand accelerator prices would do to those covenants. And the most honest thermometer of all: the spread between spot and long-term compute rental pricing. If spot softens while positioning flattens, the financing concern was never exaggerated. It was delayed.

The sideways tape is not indecision. It is the interval between a question being asked and an answer being priced. The AI trade stopped asking whether the technology works and started asking who is holding the paper. JPMorgan answered the first question with data about the second.

So here is the one I keep returning to, the one that will separate the next twelve months from the last twenty-four: if the selloff really is fading, why is the most leveraged, most transparent, most mark-to-market version of this trade — the one trading on-chain in real time — still pricing a discount to the story everyone is now reassuring each other about?