The term sheets moving through AI infrastructure in early 2025 assume a six-year economic life for GPUs that many hardware engineers quietly peg at two to three. That gap โ between the number on the collateral schedule and the number in the physical world โ is where the next leverage unwind is being written. And I have seen this movie before, except the last time it was collateralized by something even less durable than silicon: a governance token with a 20% yield and a Discord full of believers.
When Jefferies' Chris Wood โ the strategist behind the Greed & Fear note โ warned that the US AI boom could end in "massive capital destruction," the crypto press filed it under the usual frame: bubble, rotation, opportunity. That framing is lazy. Wood did not say AI fails. He said the financing structure fails. Those are different claims, and only one of them has a pre-written script.
Chris Wood is not a neutral observer. As Jefferies' global head of equity strategy, he has carried a consistent macro tilt for years โ bullish on China, skeptical of US valuation extremes, allergic to concentration risk. His "capital destruction" language sits squarely inside that prior. So the headline is a strategy note, not a research finding. Treat it accordingly. But a strategy note can still point at a real mechanical fault line, and this one does.
Here is the shift that matters. Through 2023, AI's buildout ran on hyperscaler free cash flow โ Microsoft, Google, Meta, Amazon spending money they had already earned. Comfortable. Self-funded. Boring in the best way. Then, through 2024 and into 2025, the marginal dollar of AI capex migrated to debt and structured finance: leveraged neoclouds, data-center SPVs, private credit, and the giants themselves issuing sizeable corporate bonds. The capital did not change. Its owner changed โ and the new owner charges interest.
I remember the same structural pivot inside DeFi, just faster. In 2021, yield farming looked self-sustaining while the collateral appreciated. The moment the collateral stopped appreciating, the loop inverted and the liquidation engine did the rest. AI infrastructure is running the same architecture on a slower clock.
The AI capex cycle is not a technology story right now. It is a credit story wearing a technology costume.
Start with the collateral. Neoclouds and data-center operators increasingly finance GPU purchases against the GPUs themselves โ a structure that works only if the asset holds residual value. If a GPU's effective economic life is three years and the loan amortizes over six, you have a duration mismatch sitting inside a depreciation schedule. That is not a rounding error. When the collateral's true half-life is shorter than the loan's, the borrower is underwater before the first payment resets. The depreciation debate โ six years versus two or three โ is the accounting version of a liquidity crunch waiting to be recognized.
Consider the demand assumption next. Every leveraged compute asset prices in a rent that assumes demand keeps compounding. DeepSeek reset that assumption in January 2025, when a low-cost model demonstrated near-frontier reasoning at a fraction of the expected inference cost โ a shock that vaporized hundreds of billions in market value in a single session. Efficiency is a gift to the application layer and a curse to the infrastructure layer. Cheaper inference lowers the unit value of compute precisely as the debt against that compute comes due. The mechanism is identical to what restaking revealed in 2024: security guarantees are only as strong as the economic incentives underneath them, and incentives reprice faster than contracts do.
And the revenue gap compounds it. Hyperscaler AI capex runs into the hundreds of billions annually; aggregate AI revenue, however fast it grows, is still an order of magnitude behind. Nobody in the bullish camp wants to say the quiet part: the ratio of annual AI revenue to annual AI capex is the only number that decides whether "capital destruction" has teeth. It is also the number the strategy note never provides. An unfalsifiable warning is not analysis. It is positioning.
The neocloud model, stripped of its branding, is a spread trade with silicon in the middle. Borrow against GPUs, rent them out, service the debt from rental income, refinance the residual. It works spectacularly while rents rise and rates stay low. It fails spectacularly when either leg breaks โ and both legs are exposed to the same variable, which is the market's belief in perpetual compute demand. There is a circular-financing smell here that any DeFi native recognizes instantly: chipmakers investing in the customers who commit to buying their chips, who then lease capacity back to the cloud providers who funded the original purchase. When the same dollar is counted three times, the demand signal is no longer a signal. It is a rumor with a term sheet.
Layer the electricity bottleneck on top. Data-center power demand is outpacing grid expansion by a wide margin, which pulls generation and transmission projects into the same leveraged structure. Now the collateral is not just silicon โ it is megawatts, substations, and long-dated power purchase agreements whose economics assume utilization rates that efficiency gains keep eroding. The stack is getting taller and the base is getting narrower.
Now map this onto the crypto AI complex โ DePIN compute networks, tokenized GPU marketplaces, AI-agent protocols โ and the picture sharpens. These assets inherit the same leverage thesis with none of the cash-flow buffer. A hyperscaler can absorb a mispriced depreciation schedule against operating income. An on-chain compute protocol financing hardware against a token that trades on sentiment cannot. When I modeled autonomous agent behavior for the 2026 cycle, the volatility was not driven by intelligence โ it was driven by the fact that ten thousand bots sharing one data feed amplify each other's FOMO with zero latency. The transmission chain runs: demand expectation, then rents, then GPU residual value, then leveraged borrower solvency. Watch the fourth link. That is where the pain localizes.
And note what the warning conveniently omits. The "destruction" risk does not sit with Microsoft. It sits with the leveraged middle โ the neoclouds, the data-center REITs, the private-credit vehicles, and, yes, the tokenized-compute protocols that borrowed against hardware they never fully owned. Tracing the alpha through the noise of consensus means ignoring the headline's target and finding where the leverage actually lives.
Here is the angle the crypto commentariat is missing, and it is the one worth your attention: the "AI capital destruction, therefore money rotates into crypto" narrative is itself the consensus trade, which means it is already priced. The reflexive story goes like this โ AI money dies, capital flees to digital assets, crypto's AI tokens are the cheap call option. Elegant. Wrong, or at least incomplete. Crypto's AI sector is not a hedge against the AI capital cycle; it is a higher-beta expression of it. If GPU-backed lending unwinds, the tokenized-compute protocols unwind first and hardest, because their collateral is the same silicon with a thinner margin for error. Every rug pull has a pre-written script, and this one borrows the AI narrative as its cover story. The genuine contrarian position is unglamorous. The survivors will not be the projects with the best AI story โ they will be the ones whose collateral, revenue, and token emissions are structurally aligned. Decentralization is a spectrum, not a switch, and so is leverage. The protocol that prices GPU residual value honestly, amortizes against a realistic useful life, and does not pay emissions to fake demand will look boring in a bull market and solvent in a bust.
Watch three signals over the next two quarters: the depreciation disclosures in upcoming hyperscaler earnings calls, credit spreads on neocloud paper, and the secondary market for used H100s. Those will tell you whether the collateral is holding before the collateral tells everyone else. Innovation hides in the edges of the norm โ and right now the edge is not the model. It is the residual value. The code does not lie. The depreciation schedule might.