The Guessing Game: AI's Capex Faith, Booking's CFO, and the On-Chain Compute Reckoning
Hook
There is a certain silence that follows a CFO who says the quiet part out loud. It happened this quarter on a Booking Holdings call, and the phrase moved through my feeds faster than any mainnet upgrade I have covered all year: even AI's biggest spenders are guessing on returns.
Not analysts. Not the short-sellers who have been wrong three cycles running. The chief financial officer of the largest online travel company on earth โ a company that buys artificial intelligence the way the rest of us buy electricity.
I have been the analyst who reads forty whitepapers over a weekend and surfaces on Monday hollowed out. I know the texture of a narrative that has outrun its evidence. What struck me about that sentence was not the doubt. It was the honesty. The most dangerous thing in any capital cycle is not a bad investment โ it is an investment nobody can measure. So I did what I do when a macro signal crosses into my beat. I pulled the numbers and started counting the ways a guess can be dressed as a forecast.
Context
Booking Holdings is not a speculative AI play. It is an AI consumer. Booking.com, Priceline, Agoda, KAYAK, OpenTable โ the Trip Planner, the customer-service agents, the ranking and recommendation layers โ all of it runs on inference capacity rented from the same hyperscalers that spent somewhere near two hundred billion dollars on capital expenditure last year. When the buyer of a product publicly doubts its return profile, the seller should listen.
The number that haunts this conversation has a name. In 2024, Sequoia's David Cahn framed what he called AI's six-hundred-billion-dollar question: the industry must generate roughly six hundred billion dollars in annual revenue to justify the infrastructure being poured into it. Goldman Sachs' Jim Covello followed with a blunter version โ that the technology may not be worth its cost, at least not on the timeline the market is pricing.
Those were analyst voices. They are easy to dismiss in a bull tape. Booking's CFO is something else entirely: a demand-side witness with a balance sheet. His doubt is not philosophical. It is procurement.
This is where my beat intersects. Because the crypto industry has spent four years building its own answer to the same question โ decentralized physical infrastructure networks, DePIN, the promise that idle GPUs in Seoul and storage arrays in Reykjavik and stranded wind capacity in Texas could be stitched into a compute market that undercuts the hyperscalers. The pitch was always that the cloud is overpriced and underutilized. The Booking quote suggests the opposite problem may be arriving first: the cloud is overbuilt, and its buyers are no longer sure what they are paying for.
I have watched this movie before, from a different seat. In 2017 I read forty-plus whitepapers at the peak of the ICO mania and wrote a series arguing that most projects lacked viable roadmaps. It gained fifty thousand views in a week and cost me several friendships in the industry. In 2020 I spent three months interviewing yield farmers about the psychological toll of infinite returns. By 2021 I had retreated to a cabin in Benguet to process how hollow the NFT boom felt, and I came back writing about the crisis of digital ownership. Each cycle taught the same lesson: the story arrives before the receipts, and the receipts are always late. We burned out trying to own the future. The AI trade is now running that script at a scale crypto can only envy โ and, I suspect, imitate.
Core
Let me start with the accounting, because that is where the guessing actually lives.
A GPU is a depreciating asset, and the depreciation schedule is a choice. Three years or five or six. That choice flows directly into reported profit. If a hyperscaler assumes a five-year useful life for a chip that is economically obsolete in three, it has quietly borrowed earnings from the future. The depreciation schedule is the single most under-discussed guess in the entire AI trade. When a CFO says returns are unknowable, part of what they mean is that the cost side is a convention and the revenue side is a hope.
Now stack the supply side. Advanced packaging capacity โ CoWoS โ is a bottleneck measured in quarters, not weeks. High-bandwidth memory is allocated a year out. Then there is the constraint the equity research notes rarely model: electricity. A frontier data center is a power plant with a server farm attached. Grid interconnection queues in Northern Virginia, Ireland and Singapore run into years. Export controls on advanced accelerators further distort where capacity lands, which is why the same policy conversation in Washington reshapes procurement decisions in Hong Kong and Singapore months later. These are long-lead, rigid commitments. Once you sign, you cannot un-sign quickly. That rigidity is what turns a soft demand signal into a hard problem.
Which brings me to the on-chain compute markets, and to a piece of my own work I keep returning to. In the spring of 2025 I led a small team of three experts through what became our report on decentralized AI compute โ the thesis being that training and inference would migrate toward permissionless markets the way liquidity migrated toward automated market makers. We interviewed operators, audited token emissions, modeled cost curves against reserved-instance pricing. What we found was more interesting than the pitch deck.
The decentralized compute market is not competing on price. It is competing on the pricing of uncertainty. Akash, Render, io.net, Bittensor โ each of them, in a different way, is an attempt to let a spot market discover the value of compute in real time rather than bury it inside a five-year depreciation assumption. That is genuinely novel. It is also genuinely fragile.
Here is the mechanism. A hyperscaler amortizes a GPU over years and sells a reserved instance. A DePIN network pays a supplier in tokens, and the token price becomes the subsidy that makes the economics work. When the token is up, compute is cheap and suppliers are happy. When the token is down โ as it is in this bear market โ the subsidy evaporates, and the "cheap compute" was never cheap at all. It was financed by the same speculative faith the Booking CFO is now questioning, just relocated onto a blockchain.
The lesson from 2021 holds. When the asset is the incentive, the incentive is the asset's price. DePIN compute networks are not immune to the AI ROI problem. They are its most leveraged expression.
So let me separate the layers, the way I did when the headline crossed my desk.
Layer one is training. This is where capex concentrates and where the guessing is loudest. A frontier training run costs hundreds of millions to over a billion dollars, and the return is a capability curve whose commercial value is unproven at scale. Decentralized training is technically possible and economically marginal โ bandwidth and synchronization costs eat the savings on all but the most embarrassingly parallel workloads. I would not bet a treasury on it in this cycle.
Layer two is inference. This is where real demand lives, and where on-chain markets have a structural argument. Inference is parallel, latency-tolerant across many workloads, and price-sensitive. A network that routes a batch job to the cheapest available GPU and verifies the work cryptographically is offering something the hyperscalers structurally cannot: no lock-in, no reserved-instance markup, no five-year amortization hiding inside the price. The bear case for the hyperscalers is the bull case for verifiable inference โ but only for the networks that can actually prove the work was done.
And that proof is the whole game. Verification is the moat. A decentralized compute network without a credible verification layer is a marketplace for someone else's hardware with extra steps and worse support. I have audited enough of these to know that most of them, when you read the technical documentation closely, are simply trusting the supplier. That is not decentralization. That is a cloud provider with a token.
Bittensor deserves a specific note here, because it is the most honest experiment in the space and the most misunderstood. Its subnets are markets for model quality, not raw compute, and their emissions reward measurable improvement in a task. That is a genuine attempt to price output rather than input. But the measurement itself is the vulnerability: whoever defines the metric owns the reward, and a metric can be gamed far more cheaply than a model can be improved. It is a beautiful mechanism and a permanent attack surface.
There is a parallel I keep drawing, and it comes from my Layer 2 work. After Dencun, blob space was cheap โ absurdly cheap โ and the entire rollup ecosystem exhaled. Fees collapsed. Throughput soared. I wrote, in a piece that annoyed several people, that this reprieve was temporary by design: blob demand would saturate within two years, and rollup gas would double again. The mechanism is simple. Subsidized capacity always fills. The cheapness was never a steady state; it was an opening.
The same physics governs compute. Every efficiency gain in inference โ quantization, speculative decoding, better kernels โ lowers cost per token and therefore increases the number of tokens demanded. Jevons paradox, applied to silicon. The return the Booking CFO cannot find is not missing because AI is worthless. It is missing because the cost curve is falling faster than the pricing model can adapt, and the value is leaking to whoever captures the demand. In a market where unit cost collapses, the winner is not the cheapest producer. It is the one who owns the demand relationship.
Which is exactly why the CFO's doubt should terrify the sell side and comfort the buy side. Booking owns a demand relationship. It is not guessing whether AI is useful. It is guessing how much of the surplus it can capture before the vendors reprice. That is a very different โ and much better โ guess to be making.
Contrarian
Now the part I have been circling, and it will not flatter my own industry.
The crypto response to AI capex anxiety has been a kind of triumphant schadenfreude: see, the cloud is a bubble, decentralized compute is the answer. That is a comforting story and it is mostly wrong.
The blind spot is this. Decentralized compute networks are making the same guess as the hyperscalers, but with a shorter horizon and a weaker balance sheet. When Microsoft spends two hundred billion on capex it is buying optionality with cash flow from Office and Azure and Windows. When a DePIN protocol emits tokens to subsidize GPU supply, it is buying optionality with the future price of its own currency. One of those instruments can survive a decade of uncertainty. The other cannot survive a year of it.
The Booking quote is being read as a verdict on AI. I read it as a verdict on financing structures. The hyperscaler can afford to guess, because its guess is amortized. The token network cannot, because its guess is marked to market every block.
There is a second blind spot, one that touches how we cover this at all. The headline โ even the biggest spenders are guessing โ is being weaponized to confirm a pre-existing belief: that AI is a bubble, that it all unwinds, that the skeptics were right. But that is a single sentence, lifted from a context almost none of us have seen, in a venue with no native stake in AI. We are building a grand narrative from one data point, again. We burned out trying to own the future in 2017, when I read forty whitepapers and found mostly vapor. The reflex to declare the end is the same reflex that declared the beginning. Both are guesses, and both are cheap.
The genuinely contrarian position โ the one that costs something to hold โ is that the CFO's honesty is bullish. A market that prices uncertainty has stopped lying to itself. De-rating risk is real, but de-rating is not death; it is the price of clarity. And clarity is the only environment in which verifiable compute can actually win, because verification only matters when someone is finally checking.
Takeaway
So here is the question I am holding into the next earnings season, and it is not whether AI pays off. It is who is allowed to keep guessing, and for how long.
If the answer is only those with cash flow, then the hyperscalers absorb the pain, the DePIN networks consolidate or die, and the survivors are the ones who solved verification rather than the ones who solved marketing. If the answer is everyone, the subsidy continues and we learn nothing. And if the answer is no one, we finally find out what compute is actually worth โ which is the only number that ever mattered. The chart will not tell you which. The procurement officers will.