The most expensive sentence in technology right now is not a valuation. It is a footnote.
Buried in the accounting policies of every hyperscaler is a line stating that servers and networking equipment are depreciated over roughly six years. Move that number to four and nothing about the business changes. No product ships differently. No model gets measurably smarter. No customer leaves. And yet, on a capital base now measured in the hundreds of billions annually, the arithmetic moves by tens of billions of dollars a year.
I went looking for that footnote because a crypto-native outlet published a piece whose substance was that investors are pushing back on sky-high AI valuations and that stock sales are getting harder. Four sentences of claim. Zero numbers. No named companies. No discount figures. No distinction between a failed IPO, a discounted follow-on, a tender at a markdown, or a convertible bridge.
That emptiness is the signal — not the claim, but the placement. A concern that lived on credit desks and in late-stage growth funds has crossed into a publication read mostly by people holding tokens. When a worry migrates from the people who price risk to the people who price narrative, you are no longer early in a cycle. You are in the middle of one.
I have a specific reason to care about the shape of this. In 2017, at nineteen, I spent three months manually auditing token distribution contracts instead of buying tokens. I found three logic flaws in a popular storage project's allocation mechanism and published them to an audience of roughly five thousand people. The lesson I took was not that I was clever. It was that value lives in what can be verified, and everything else is a story someone is telling you at a specific moment for a specific reason.
The AI capex boom is the largest story anyone is telling right now, and it is being financed almost entirely outside the visible ledger. Not failed transparency — structural invisibility. Data center joint ventures. Special purpose vehicles. Private credit facilities with negotiated terms. Off-balance-sheet capacity commitments. There is no public tape for any of it.
That is the part that should bother anyone who spent years arguing that open books beat closed ones. Open books, open ledgers, open hearts — the phrase is not decoration. It is an operating principle about where risk becomes priceable. When the financing of a three-hundred-billion-dollar annual buildout moves into instruments that never trade on an exchange, the market does not become more sophisticated. It becomes more fragile, because price discovery happens in rooms with three participants and no mark.
So when I read that stock sales are getting harder, my first instinct was not to ask whether AI is overvalued. My instinct was to ask which financing channel closed. The answer tells you whether this is a sentiment wobble or a change in the constraint itself.
Here is the arithmetic the headline skipped.
Hyperscaler capital expenditure tied to AI has crossed into the hundreds of billions annually — call it three hundred billion as an order of magnitude. Against that, the two most-discussed frontier labs are running annualized revenue in the low tens of billions, roughly thirteen billion and seven billion respectively on the most recent public estimates. Add every credible application layer, every enterprise seat, every API call, and you are still comfortably under a hundred billion of attributable AI revenue against three hundred billion of annual spend.
You do not need to believe AI is a bubble to find that gap uncomfortable. You only need to notice that closing it requires growth that compounds faster than anything in the history of enterprise software, sustained for years, without a demand pause.
The gap is not the interesting part. The financing structure is.
A capital cycle funded by equity is constrained by risk appetite. A capital cycle funded by debt is constrained by collateral, duration, and loan-to-value. Those are different physics. Equity can be talked into a story; debt needs the asset to survive a default. When a buildout of this size migrates from the first to the second — and it has, through corporate bond issuance, data center joint ventures, SPVs, and private credit — the binding constraint stops being whether investors like the future and starts being what a GPU is worth in three years.
Which brings us back to the footnote.
A three-hundred-billion-dollar annual capex program depreciated straight-line over six years produces roughly fifty billion in annual depreciation. Over four years, it produces seventy-five. Same assets. Same revenue. Twenty-five billion of pre-tax income disappears, and that is one year, before the growing base compounds.
The dispute is real. Hyperscalers use roughly six years. A growing number of investors and analysts argue the economic life of AI accelerators is closer to three or four, because the workloads change faster than the silicon wears out. Nobody is lying. But the difference between those two assumptions is, in a single stroke, larger than the entire annual revenue of most companies in the S&P 500.
This is where a valuation debate turns into an earnings event, and it is the single most under-watched lever in the entire story. Not the multiple. The useful-life assumption.
Now the reflexivity, which is where my own history makes me twitchy.
In 2020 I watched liquidity mining create a very specific shape on-chain: a protocol's token subsidized the liquidity that priced the token. It worked beautifully, and it worked right up until the subsidy stopped. In 2022 I watched collateral that was also the asset unwind in slow motion, each liquidation feeding the next price print. The mechanism is not mysterious. It is a loop where the same entity appears on both sides of the transaction.
The AI version of that loop is bigger and harder to see. A chip supplier invests in a model company. The model company commits to buy compute. A cloud provider buys the chips and hosts the model company. The cloud provider's forward revenue is underwritten by the model company's funding round, which was led by the chip supplier. Every leg is real. Every leg is also the same dollar, counted three times.
In an upcycle, that structure looks like strategic alignment and gets a premium. In a downcycle it inverts: valuation falls, funding tightens, orders get cut, the supplier's guidance drops, the supplier's equity falls, and the book value of its investments falls with it. The amplifier works in both directions with roughly equal gain.
I have argued for two years that the data availability layer is the most over-financed idea in rollup land — that the overwhelming majority of rollups do not produce enough data to justify dedicated DA capacity, and that the market built cathedrals for a congregation that never arrived. The AI data center buildout is the same error with better lighting. Dedicated capacity is a bet on pre-leasing, and pre-leasing is the number almost nobody quotes. Watch it. A pre-leasing miss is a harder signal than multiple compression, because it is a signed contract that did not get signed.
And while I am on the subject of things that should not exist: the trade I keep seeing pitched in crypto circles, wrapping GPU leases into yield-bearing tokens, is a Rolls-Royce hauling cargo. It insults the vehicle and it does not carry much. You are taking a depreciating asset with renegotiated cash flows and no secondary market, wrapping it in a settlement layer built for trust-minimized finality, and calling the result yield. The chain adds nothing except an audit trail for a number that was negotiated in private.
Speaking of which, the pricing of those private credit facilities deserves scrutiny. There is no reference curve for compute-backed lending. The rate is negotiated, which means it is an administered price wearing the costume of a market price. I have made the same criticism of lending protocols for years: the kinked interest rate curve everyone calls market-driven is a set of governance parameters chosen by a vote and then narrated as supply and demand. Both are fictions. The difference is that the on-chain version at least publishes its parameters.
Here is where I part company with the consensus reading.
Everyone is watching the valuation multiple. The multiple is the last thing to move and the least informative. By the time a headline says investors are pushing back, the repricing has already happened somewhere else — in tender discounts, in liquidation preferences stacked on top of the last round, in ratchet provisions that quietly restore an earlier price. The absence of a down round is not evidence of price stability. It is evidence that the reset was engineered to be invisible.
That lag runs six to twelve months. It is also the entire information advantage available to anyone reading this.
The second blind spot is closer to home. The AI valuation pushback is a crypto risk event before it is an equity risk event. AI-adjacent tokens — the compute networks, the agent coins, the decentralized training projects — are the highest-beta, lowest-cash-flow expression of a narrative now being repriced at the source. They have none of the strategic capital protection that shelters a hyperscaler and none of the revenue that anchors an application company. When the marginal dollar leaves the story, it leaves the marginal asset first and hardest.
Literacy in the blockchain age is power, and the literacy being tested right now is the ability to tell a fundamental question from a liquidity question. Whether AI is overvalued and whether AI stock is hard to sell are not the same sentence. The first is a judgment about cash flows. The second is a judgment about who has dollars this quarter. A piece that blends them cannot support the conclusion people will draw from it, and the people who draw it will be holding the most fragile version of the trade.
What actually happens next is less dramatic than either the bulls or the bears want. It looks like layered clearing. Players with strategic capital behind them — cloud providers with balance sheets, labs with sovereign and corporate backers — keep spending and treat a higher cost of capital as an inconvenience. Players dependent on market-clearing equity get squeezed into vertical niches or sold. The leveraged compute middle layer, the neoclouds and third-party data centers whose business is borrowing against depreciating hardware to resell it, faces the worst of both: falling spot rental rates, rising collateral requirements, and no strategic parent to absorb a bad quarter.
That is not a systemic turn. It is a sorting. And it produces the one thing crypto people should actually be excited about: a merger window. Cash-rich buyers acquiring model and agent assets well below 2024 marks. Chaos is just creativity waiting for structure, and the structure here is consolidation.
There is a second-order consequence worth naming. When equity financing tightens, more teams switch from training their own foundation models to fine-tuning open ones. That is not an ideological conversion. It is a budget line. Open source is becoming the financially rational choice for everyone who cannot raise at 2024 prices, and the beneficiaries are the unglamorous layers: inference optimization, private deployment, evaluation, compliance tooling. Building bridges where others build walls is usually good business, and this is one of those moments.
So what do I actually watch, if not the multiple?
Three signals, in order of value. The depreciation language in the next round of quarterly filings — if a major operator shortens useful life, the entire sector's earnings base resets and no demand narrative survives it. Actual discount levels on any AI-linked offering that does come to market, and whether an offering is withdrawn entirely. And GPU spot rental rates alongside secondary residual values, because those are the only public prices for the collateral sitting underneath the debt.
The audit is not the end, but the beginning. We are not watching a bubble deflate. We are watching a market discover, slowly and in private, what the assets are actually worth — and then discover that the answer was never going to be found on a public ledger, because the financing left it.
If the most important number in this cycle is a footnote nobody reads, what does that say about how much of the global economy has quietly moved off the books?
