Every market has a confession, and most investors only hear it after the drawdown. This week, Booking Holdings' CFO said something that should have stopped the tape: even the biggest spenders in AI are guessing on their returns. Not "estimating." Not "modeling with wide confidence intervals." Guessing.
That single verb is more consequential than any earnings beat this quarter. It tells you that the firms holding the most complete first-party data — the ones who know exactly how many GPUs they bought and exactly how much revenue those GPUs generated — still cannot close the loop between capital expenditure and realized value. When the people with the best data in the world admit they are guessing, the rest of us are trading narrative, not numbers.
I started my career auditing Solidity, not price charts. In 2017 I traced the fee logic of a bonding-curve protocol and found an integer overflow that turned a routine calculation into a money printer. The bug was invisible until you followed the arithmetic to its edge. AI capex right now has the same structure: a transmission chain — capital to compute to capability to cash flow — that everyone assumes holds, but almost nobody can formally verify.
Booking Holdings is not a bystander, and that is what makes the quote load-bearing. Booking.com, Priceline, Agoda, KAYAK, OpenTable — the AI Trip Planner, the recommendation engines, the support automation. This is one of the largest application-layer buyers of AI on the planet. When an application-layer CFO questions the return, you are not hearing an analyst's opinion. You are hearing demand-side pushback against supply-side pricing.
The macro backdrop is a liquidity map drawn in a single color. Sequoia's David Cahn framed the "AI's $600 billion question" — the annual revenue the industry must generate to justify the infrastructure buildout. Goldman's Jim Covello published a skeptical ROI note. Four hyperscalers — Microsoft, Google, Amazon, Meta — ran capital expenditure above the $200 billion range in 2024, with 2025 guidance pushing toward the $300 billion magnitude. These are long-lead assets: GPUs, CoWoS advanced packaging, HBM memory, and increasingly electricity itself. You cannot order a data center and cancel it next quarter. Supply is rigid precisely where demand is speculative.
This is where the global liquidity story turns. Capital allocation is no longer responding to demand signals; it is responding to competitive fear. And fear, unlike demand, has no natural clearing price.
Here is the part the financial press keeps missing: the AI capex debate is a verification problem, and verification is exactly what blockchains were built to solve. I do not mean that tribally. I mean it structurally.
In my 2026 research on the AI-agent economy, I simulated 10,000 autonomous agents competing for finite compute. The finding that mattered was not the competition — it was the identity problem. Agents transacting autonomously need verifiable, non-transferable identity to resist sybil attacks, and they need zk-SNARKs to prove authenticity without exposing proprietary algorithms. The trust substrate for autonomous economic actors is cryptographic, not institutional. That conclusion quietly reframes everything downstream.
Now map it back to the hyperscaler problem. The reason a CFO is "guessing" is that the return on AI capex is unobservable in real time. Cloud revenue is disclosed, but attribution is fuzzy. Advertising efficiency gains are indirect. Productivity uplift at the application layer is diffuse. There is no settlement layer where the value created by a unit of compute can be traced to a cash flow. The system runs on accrual accounting and hope — and, more quietly, on depreciation policy. The assumed useful life of a GPU, three years or five, directly sets reported profit. That assumption is itself a guess wearing a spreadsheet.
There is a harder ceiling too. Electricity. Data center power draw and grid capacity are becoming a more binding constraint than silicon, and no amount of capital can compress a transmission line's construction schedule. Supply is rigid, demand is speculative, and the accounting that bridges them is soft.
In 2024, I calculated that the Bitcoin ETF structure introduced roughly a four-hour latency gap between on-chain liquidity and the traditional settlement layer — a predictable spread that returned 12% alpha in a single quarter. The lesson was never "crypto beats TradFi." The lesson was that when you can measure a temporal dislocation between two systems, you can price it. AI capex today has the dislocation, but no measurement rail. That is why it feels like faith.
The liquidity pool is a mirror, not a vault. AI capital expenditure reflects the confidence of the allocators, not the verified value of the assets. When you cannot mark the return, you mark the narrative — and narratives reprice faster than fundamentals ever do.
The second-order effect is already visible in the supply chain. If ROI anxiety becomes consensus, the transmission runs in three stages. Hyperscalers slow or re-optimize capex. Infrastructure suppliers — NVIDIA, TSMC, data center REITs, power providers — absorb the first shock, because their order books are the most levered to the capex assumption. Application-layer firms like Booking gain relative bargaining power, because falling AI prices cut their procurement cost. The chain runs from a seller's market to a buyer's market, and it runs fast.
The consensus reading of this story is "AI bubble." That is the lazy decoupling thesis, and it is wrong in an interesting way.
The blind spot is the assumption that capex and monetization move together. They do not. Crypto learned this in 2022, and I was one of the few analysts inside my firm who argued the FTX collapse was not a leverage story but a failure of recursive yield models — a single token de-peg cascading through lending protocols because the collateral and the liability were the same asset wearing different names. The AI capex complex has a structural echo: hyperscaler revenue partly funds the demand for hyperscaler compute. Circular, recursive, fragile.
But here is the inversion. If ROI is genuinely unverifiable, the winning assets are not the ones with the biggest capex. They are the ones that can prove monetization on-chain, in real time, to anyone who cares to check. Exit liquidity is just another person's thesis. AI capex feels like a top precisely because its exit liquidity is undefined — nobody can name who buys the compute when the growth story stalls.
And a decoupling is already underway that the mainstream frame ignores. Decentralized compute networks, verifiable inference, and on-chain settlement rails are not competing with hyperscalers on raw scale. They are competing on auditability. A CFO can guess on a black-box cloud bill. A CFO cannot guess on a settled on-chain payment. Regulation is the lagging indicator of chaos — and so, it turns out, is accounting. The algorithm optimizes for survival, not for you. The hyperscaler capex algorithm is optimizing to avoid losing the race, not to return your capital.
The signal is not that AI is over. The signal is that the market has begun to demand verification where it previously accepted belief. The infrastructure that survives this cycle will not be the one that spent the most. It will be the one that can prove what its compute actually earned — and that proof, increasingly, will be cryptographic rather than audited.
The open question for the next four quarters: when hyperscaler capex guidance finally decelerates, which assets reprice on fundamentals — and which were only ever priced on the guess?