The Missing Clause
A crypto news outlet published a headline this cycle stating that Howard Marks is not convinced the AI boom is a bubble. I read it three times. Not because it surprised me, but because the summary sitting directly beneath it said something materially different โ that investors should still be cautious about stretched valuations.
A verdict and a caveat are not two phrasings of one thought. They are two different trades. One says stay long; the other says re-check what you paid. In a market where the marginal buyer is a passive index allocation and the marginal seller is a treasury desk, that distinction is worth more than most of the research notes that crossed my desk this quarter.
I am not writing this to scold an aggregator. Headlines are optimized for clicks because clicks are the revenue model, and everyone in this industry prices that in. What interests me is narrower and more useful: that this specific compression happened, on this specific topic, at this specific venue, at this specific point in the cycle. Four variables, one reading.
Here is the arithmetic the headline left out and the piece never touched. In the last set of guidance I reconciled โ mid-2025 vintage, and it needs re-verification against the most recent 10-Qs โ the four largest North American hyperscalers were pointing toward roughly $300 to $350 billion in combined annual capital expenditure. The combined annualized revenue of the leading frontier model developers was in the low tens of billions. Those two numbers do not live in the same order of magnitude, and the gap is not closing at the speed the current valuations require.
That is the argument. Everything else is commentary.
Context: What Marks Actually Sells
Howard Marks built Oaktree around distressed credit. That single biographical fact should reframe how you read any headline carrying his name into a market like this one, because credit investors do not think in multiples. They think in covenants, maturities, recovery rates, and the identity of the party on the other side of the trade when the music stops.
His framework on bubbles is well documented across three decades of memos, and it is not what the headline implied. His test has never been 'did the asset go up a lot' or 'is the narrative hot.' His test is a conjunction: psychology overheated, price already discounting an optimistic scenario in full or beyond it, and the pool of incremental buyers exhausted. Every one of those conditions is price-dependent. The same technology at half the multiple is an opportunity; at twice the multiple it is a trap. Strip the price clause out and you have not summarized his view โ you have deleted the load-bearing wall.
For calibration: Marks published 'bubble.com' in January 2000, roughly ten weeks before the Nasdaq topped. He was not early that time; he was almost exactly on time. But he has also said, repeatedly and in public, that being early and being wrong are indistinguishable for the eighteen months in which a career gets destroyed. That is a credit investor's instinct speaking โ the acknowledgement that correct analysis delivered at the wrong funding tenor is still a loss.
The venue matters as much as the quote. The chain here runs from a value investor's conditional remark, through an aggregator, into a crypto audience. That is a two-hop transmission with an incentive at each hop. The aggregator needs a clean binary. The crypto audience needs reassurance that the risk-asset complex is intact, because their holdings and AI equities have spent two years moving on the same liquidity impulse. Neither hop requires the price clause to survive. Both hops benefit if it does not.
And what was actually in the piece? Four information points, three of them restatements of the same claim, zero direct quotations, no publication date, no indication of whether the source was a memo, an interview, or a podcast appearance. As an analyst, that is a null input for content. As a narrative hunter, it is a positive signal for positioning โ and the distinction between those two readings is the whole job.
Core: Three Numbers Nobody Puts on the Same Slide
The capex-to-revenue ratio is the hardest data in this debate and the least quoted, because it is boring and it does not flatter anyone. Let me put the three figures side by side the way I would in an internal memo.
First, the spend. Roughly $300 to $350 billion of annual hyperscaler capex, with a meaningful share of that attributable to AI infrastructure โ accelerators, networking, cooling, and the power contracts behind them. Second, the revenue. The leading model developers combined are running annualized figures in the low tens of billions, and a non-trivial portion of that is enterprise contract value booked but not yet consumed, which behaves differently from recognized revenue under stress. Third, the ratio. Even crediting the cloud businesses with their AI-adjacent incremental revenue โ and the disclosure here is vague enough that nobody outside the companies can cleanly separate it โ the capex-to-AI-revenue ratio sits at a level with no modern precedent outside of state infrastructure programs.
That is not a claim that the technology is fake. It is a claim that the financing of the technology has run ahead of the monetization of the technology by a margin that requires either extraordinary revenue acceleration or a repricing. There is no third exit.
Now the number that genuinely is not being discussed: depreciation.
In 2022 and 2023, several hyperscalers extended the assumed useful lives of servers and networking equipment, and the aggregate earnings benefit ran into the billions. That was a defensible accounting judgment for general-purpose compute. It is a far less obvious judgment for accelerators, where the generation cadence runs closer to eighteen to twenty-four months and where the secondary market for prior-generation parts tells you something the spreadsheet does not.
Run the sensitivity yourself. Take an operator deploying $50 billion a year into accelerators. At a six-year assumed life, the annual depreciation charge is about $8.3 billion. At four years, it is $12.5 billion. That is a $4.2 billion pre-tax swing from a single assumption, with zero change in cash flow, zero change in demand, and zero change in the technology. On a company earning $100 billion pre-tax, that is a four-point margin hit manufactured entirely inside the accounting policy note. Extend the same logic across the sector and you are talking about tens of billions of annualized pre-tax earnings that exist because of a useful-life assumption rather than a customer.
This is not fraud. It is assumption divergence, and assumption divergence is where the mispricing lives, because it is invisible to anyone reading a headline about whether something is or is not a bubble. My audit work on protocol treasuries taught me the same lesson in a different domain: the number that kills you is never the number on the dashboard. It is the assumption underneath it that nobody thought to challenge because it had been constant for three years.
Third layer: how the spending is financed, which is where this cycle diverges hardest from 1999.
A material share of AI infrastructure buildout is now funded off balance sheet or through structures that blur the line between customer, supplier, and creditor. Special purpose vehicles holding data center assets with private capital partners taking majority equity. GPU-backed securitizations and asset-backed facilities collateralized against accelerator fleets whose residual values depend entirely on a rental market that has been repricing downward. Vendor financing, where the chip supplier invests in the customer that then commits to buying the chips. Structured joint ventures for single campuses running into the tens of billions, as reported, with the equity partner's return tied to lease payments from an operator whose own revenue is a function of demand that has not yet arrived.
In 2000, the leverage was visible. It sat in telecom bonds with names you could look up, and when it broke, it broke in public. This cycle, the leverage sits in private credit vehicles, in ABS tranches, and in lease structures that never touch a public filing at the granularity required to price them. Private credit has grown into a roughly $1.7 trillion asset class and data center exposure has become one of its favored allocations, precisely because the yields are attractive and the marks are infrequent.
The implication is structural: the first place this cycle cracks is unlikely to be an equity index. It is more likely to be a credit spread that nobody outside the deal room is watching.
Fourth layer: concentration, which converts a sector problem into a systemic one.
The top ten names in the S&P 500 now represent roughly a third of the index by weight. There is no modern precedent for that, including 2000, when the technology sector was large but the single-name concentration was nothing like this. This matters mechanically, not rhetorically. A repricing of AI infrastructure is not a sector rotation that a portfolio manager can underweight into. It is an index event that transmits through passive flows, pension allocations, and household wealth effects into consumption. The Fed's reaction function in that scenario is not a thought experiment; it is the whole ballgame.
And then the layer I actually trade: how all of this reaches crypto.
Crypto's own history supplies the cleanest control group available, and it is one most AI analysts have never examined. I have spent years documenting a network that was technically real, genuinely decentralized, and economically disappointing for the better part of a decade โ routing failure rates that never fell far enough, channel management complexity that pushed every rational user back to custodial rails, liquidity provisioning that only worked at sizes nobody needed. The technology shipped. The economics did not. The narrative survived anyway, because narratives about infrastructure have a longer half-life than the infrastructure's own unit economics.
That is the control group for AI. Not because the technologies are comparable in scale, but because the failure mode is identical: a real technology whose hard problems were priced as already solved, held aloft for years by the fact that belief is cheaper to maintain than a P&L.
In crypto, that failure mode now has a live ticker. The compute-adjacent token complex โ decentralized GPU marketplaces, inference networks, agent-adjacent assets โ is functionally a levered, uncollateralized claim on accelerator rental economics. These tokens have no capital structure, no covenants, no credit facility to bridge a demand gap, and no ability to survive a drawdown longer than their treasury runway. When the underlying rental rates move, these repricing events are violent and immediate, because there is nothing between the narrative and the bid.
I have watched the governance layer of several of these networks from the inside. The pitch is community-directed allocation of compute subsidies โ token holders deciding which clusters get funded. The reality is turnout under five percent on the votes that matter, with three or four wallets determining outcomes, and the largest of those wallets frequently sitting on the same side of the table as the foundation that authored the proposal. That is not a governance failure unique to compute. It is the default state of on-chain governance, and it means the sector's ability to make a disciplined capital allocation decision under stress is roughly zero.
Contrarian: The Question Is Malformed, and the Answer Is Somewhere Else
Everyone is asking whether AI is a bubble. That is the wrong question, and the fact that it is being asked this way across both technology and crypto media tells you more about positioning than about fundamentals.
The tradable question is not yes or no. It is: which instrument reprices first, and in what order?
My answer, and I will put it plainly because this is the only part of the piece that is actually actionable: credit reprices before equity, and the accelerator rental market reprices before credit. Equity indices are the last place the information shows up, because equity has unlimited duration and a passive bid that absorbs bad news for quarters at a time. Credit has a maturity date. It cannot wait.
Here is the second contrarian claim, and it is the one that should worry crypto holders specifically: AI is not crypto's tail risk. AI is crypto's competitor.
From 2017 through 2021, crypto was the highest-Sharpe speculative narrative available to a generalist allocator. That monopoly ended in 2023. The marginal narrative dollar since then has gone into compute, and it went there for good reasons โ real revenue, real customers, real cash flows, and a story that a pension consultant can defend to a board. Crypto's share of that marginal dollar fell, and the reflexive attempts to reclaim it have mostly taken the form of rebranding: every project with a GPU somewhere in its architecture became an AI project, and every AI project with a token became a decentralized intelligence network.
So the consensus trade โ 'when AI pops, money rotates into crypto' โ is not supported by anything I can measure. In a genuine VaR shock, capital does not rotate. It goes to cash, and everything with duration gets sold in the same hour. Bitcoin's correlation to the Nasdaq during drawdowns has been repeatedly demonstrated since 2022. The correlation is not a fundamental property of either asset; it is an artifact of the fact that both are now funded by the same liquidity regime and liquidated by the same risk desks. Rotation is a story people tell in calm markets. Liquidation is what happens in the other ones.
The third claim is the one that keeps me up: if the bubble consensus forms too early, the damage is real, not just financial. Capital withdrawal slows deployment, deployment slowdown delays the cost curve decline, and the delayed cost curve decline is what actually determines whether the unit economics ever work. That is reflexivity in its purest form โ the expectation of failure manufacturing the failure. It is a worse outcome than a repricing, because a repricing transfers wealth while a delayed technology cycle destroys output. Almost nobody models it, because almost everyone models prices.
And the fourth: the assets most exposed are not the AI companies. They are the traditional assets that put on the costume. Utilities that re-rated on data center load growth. Transformer and switchgear manufacturers trading at software multiples. Data center REITs. Industrial gas suppliers. Every one of them now carries an AI premium that their pre-AI valuation framework never supported, and every one of them will be sold by generalist managers who never intended to own an AI bet in the first place. In crypto, the same dynamic applies to a smaller and more vicious universe: the tokens that rebranded in 2024 and have no product, no revenue, and no treasury.
The mispricing is not in the asset. It is in the sentence. And sentences are cheaper to correct than balance sheets โ which is exactly why the correction, when it comes, will start with the thing everyone is currently quoting instead of the thing everyone should be reading.
Takeaway: What I Am Actually Watching
I have run a version of this playbook before. In 2022 I shorted algorithmic stablecoins through options while the industry was still writing threads about innovation, and the edge was not superior modeling โ it was reading the mechanism instead of the narrative. The mechanism said the peg could not hold at that collateral ratio. The narrative said the mechanism was elegant. The mechanism filed first.
So here is the dashboard, ordered by signal speed rather than by importance, because in a bear market the timing of the signal is what determines whether you survive to use it.
Accelerator rental spot rates are the fastest tell. Downward moves in the spot market for current-generation parts precede demand recognition in filings by at least a quarter, and they are observable in public marketplaces every day. If spot pricing breaks below the level at which GPU-backed facilities can service their debt, the credit conversation starts within weeks, not quarters.
The capex-to-run-rate ratio, computed quarterly against guidance, is the second. Not the absolute number โ the derivative. A flattening capex guide against an accelerating revenue run-rate is the bull case. A rising capex guide against a decelerating run-rate is the event.
Depreciation policy changes are the third, and the most underrated. Any downward revision to assumed useful life is a confession about profit quality, delivered quietly in a footnote, and it will be the single most informative disclosure of this entire cycle. Watch for it before you watch for anything else.
Private credit and ABS spreads on data center and GPU collateral are the fourth. They are the only place where the leverage is actually visible, and they will widen before equities acknowledge anything.
The article that started this piece will not appear anywhere on that list, and that is precisely its value. It is not evidence about AI. It is evidence about where we are in the conversation โ and the conversation has now traveled from technology media to crypto media, from growth investors to value investors, from people who build the thing to people who trade the ticker of the thing. Historically, that journey is not an early-cycle phenomenon.
Which leaves one question worth sitting with. If the technology is real, the capital is committed, the depreciation is a choice, and the leverage is invisible โ then the only variable that actually decides the outcome is time. Not whether AI matters. How long the financing can wait for the revenue to arrive, and who is holding the paper when it stops waiting.