Goldman Sachs published a note warning that the AI investment boom will not last forever. The market read it as sentiment. It is a maturity schedule.
The arithmetic the headline buried: AI capital expenditure is incurred up front, in cash, on a fixed timeline. The revenue meant to justify it is back-loaded, conditional, and denominated in a price nobody has negotiated yet. Between those two facts sits a duration gap. Widen it by two quarters and the financing structure carrying it stops being an investment and becomes a liability.
I have watched this shape before. In 2022 I modelled the UST peg as a subsidy schedule — daily burn against daily inflow — and published the decoupling three weeks before it happened. The input that mattered was never narrative. It was the ratio.
Context
Goldman's position, as reported, is not a rejection of AI. It is a downgrade of timing. Economic impact is framed as real but less immediate and less transformative than consensus. That is a duration call wearing a technology costume.
The capex side is publicly verifiable. The four largest hyperscalers have each guided to sustained infrastructure spending across multiple fiscal years, and depreciation on those fleets is now visible in their income statements.
Crypto absorbed the same thesis through a different funding channel. Where hyperscalers issue debt and equity, the AI-token sector issues dilution. DePIN compute networks, agent frameworks, and tokenized-GPU protocols raised capital against a single promise: decentralized supply would undercut centralized cloud pricing. In the current bear market, most of that sector trades far below its cycle highs, and the LP cohorts that underwrote it have largely exited.
Two funding channels, one narrative. That is why a bank note about capex is a crypto story.
The depreciation clock
The most under-modelled line item in the AI trade is not revenue. It is useful life.
A GPU's economic obsolescence runs on a different clock than its accounting depreciation. Hyperscalers have extended useful-life assumptions on server and accelerator fleets, which lowers the annual depreciation charge and raises reported operating income. If the physical clock is shorter than the accounting clock, the earnings are not fraudulent. They are early. The correction arrives in a guidance revision, and it arrives all at once.
Watch guidance, not launch events. Capex guidance revisions are the transmission mechanism from a duration problem to a price problem.
Capacity as a derivative
Decentralized compute carries a structural flaw no whitepaper resolves. Node operators are paid in tokens. Their cost base — electricity, bandwidth, hardware — is denominated in fiat.
Halve the token and operator revenue halves while the cost base holds. Below a threshold, rational operators exit. Capacity does not decay gradually; it steps down as the marginal operator shuts off. Advertised network capacity is therefore not a fixed resource. It is a levered function of its own token price. Reflexive, and collateralized by nothing.
Protocol integrity is binary; trust is a variable. When the token price is an input to the supply curve, headline capacity is a marketing number, not an engineering one.
What my 2025 audit found
I benchmarked ten projects claiming decentralized validation or decentralized compute. Eight resolved to centralized cloud infrastructure. Traceroutes terminated inside the address ranges of three major providers. The nodes were containers.
Read that against Goldman. The bull case for AI-crypto convergence rests on decentralized supply absorbing demand when centralized capex discipline tightens. That supply does not exist at scale. The sector is not a hedge against the capex cycle. It is a high-beta expression of it — same narrative, worse financing.
The oracle nobody stress-tested
The agent narrative has a dependency chain that fails quietly. An on-chain agent executing financial actions depends on off-chain inference. That endpoint is an oracle feed. It has latency, a rate limit, a pinned model version, and a custody key.
None of those properties are enforced on-chain. A 429 response during a volatility window is not a degraded experience. It is an unhedged position. I ran that scenario against three agent frameworks in 2024. Two had no fallback path. The third failed over to a second key held by the same operator.
Code is law, but logic is the jury. Here the logic depends on a feed with no SLA and no slashing condition.
What the bulls got right
Two things, and neither is small.
Inference unit costs have fallen faster than any comparable compute cost curve in commercial history. That is genuine deflationary pressure, and it eventually flows to application margins. Second, overbuild is not destruction. Dark fiber laid in 1999 was written down to nothing, then repriced into a decade of cloud revenue. Excess accelerator capacity gets repriced the same way. The assets outlive the thesis.
So the bearish read is wrong about the asset and right about the price. And Goldman's note is not a measurement — it is a distribution of expectations from an institution with positions, clients, and a book to manage. Treat it as a data point, not an oracle.
The blind spot both camps share: neither models the emission leg. These networks pay for supply with dilution, and dilution is a cost that appears on nobody's income statement.
Takeaway
Volatility is the tax on uncertainty. The question is not whether AI reshapes the economy. It is who holds the duration, and at what price they booked it.
Ask for three documents. The accelerator depreciation schedule. The node operator's fiat cost base. The inference endpoint's rate limit and failover contract. If a project cannot produce all three, its capacity is a claim, not an asset. Recovery is not a phase; it is a reconstruction. Start with the ledger.