Over the past four quarters, the four largest North American cloud providers have guided toward combined capital expenditure in the range of three hundred to three hundred fifty billion dollars a year. The aggregate annualized revenue of the leading frontier model companies — the names that appear in every pitch deck and every conference keynote — is, on the most generous public accounting, in the low tens of billions. Ten to one. That ratio is the hardest number in the entire artificial intelligence debate, and it is the one number that never makes the headline.
The headline instead read: Howard Marks is not convinced the AI boom is a bubble.
Read that sentence twice. It is grammatically cautious and semantically reckless. 'Not convinced' sounds like doubt. But placed in front of 'is a bubble,' it functions as denial. The reader's mind files it under a category — a famous investor says the thing is not a bubble — and the qualifier evaporates by the second reading. This is not a lie. It is something more durable than a lie: a frame.
I do not trade AI equities. I trade the risk-appetite asset that correlates with them, week to week, at a coefficient high enough that I have stopped pretending the two books are separate. When a narrative like this one gets minted, it does not stay in the equity market. It leaks. It reaches the funding rates, the perpetual basis, the altcoin beta, the entire reflex chain that runs from a headline in a media feed to a liquidation cascade on an exchange at three in the morning. That is why I read the headline and not just the article. The headline is the thing that gets traded. The article is the thing that gets forgotten.
What the article actually contained matters, and it matters precisely because of how little it contained. Four information points. Three of them restatements of the same claim. No direct quotation from Marks. No publication date. No indication whether the original source was a memo, an interview, a podcast, or a paraphrase of a paraphrase. This is not an indictment of the writer. It is a statement about what the artifact can structurally bear. You cannot cross-validate four points. You cannot triangulate a position from a summary of a summary. What you can do is read the artifact as evidence of something else entirely — not evidence about AI, but evidence about where we are in the cycle that AI is currently hosting.
Let me be precise about the man, because the precision is the point.
Howard Marks built Oaktree on distressed credit. His professional instinct is not the price-to-earnings multiple; it is the recovery rate, the covenant, the seniority of the claim. When he writes about bubbles — and he has written about them for decades, through the 2000 unwind and the 2008 seizure and the 2021 froth — his framework has three legs, and none of them is 'the technology is fake.' The first leg is psychology: has the average participant stopped asking whether the price is justified? The second is price: has the current quotation already fully discounted the optimistic scenario, leaving no room for disappointment? The third is flow: have the incremental buyers been exhausted, so that the marginal seller sets the price?
Notice what that framework implies. It implies that the same technology is an opportunity at one price and a bubble at another. The verdict is not a property of the asset. It is a property of the entry point. Any headline that reports the verdict without the price has amputated the argument at the joint and presented the limb as a whole body.
The article's own body text apparently carried the missing qualifier — something to the effect that investors should remain wary of overvalued valuations. Put those two sentences next to each other. The headline says the boom is not a bubble. The body says the valuations may be overvalued. Operationally, those statements point in opposite directions. One says hold. The other says the price is wrong. A reader who stops at the headline has been handed a conclusion that the underlying source, in all likelihood, explicitly refused to give.
There is a date problem too, and it is not cosmetic. A bubble judgment is time-stamped the way a laboratory result is time-stamped. The same analyst looking at the same asset in December 2024 and in October 2025 could reach opposite conclusions without contradicting himself once, because the price moved and the earnings estimate moved and the financing cost moved. Strip the date and you strip the conclusion of its warrant. I have watched traders build positions on a two-year-old quote from a fund manager who had already reversed his view. The ledger remembers what the market forgets, but only if somebody writes it down.
Notice also what a statement like 'not convinced it is a bubble' cannot tell you: whether it means 'it will not collapse' or 'it has not yet reached the point of collapse.' Those two readings have opposite operational implications. The first says hold through anything. The second says you have time, and time is not the same as safety. Every bubble in history has been correctly identified as not-yet-popping by someone standing in the middle of it.
So let me write it down, in the terms that actually matter.
The first thing the headline obscures is that technological truth and investor return are independent variables. Railways were real. Radio was real. Commercial aviation was real. The internet was real. In every one of those cases the industry won and a large fraction of the capital that funded it lost. The pattern is not 'the technology failed.' The pattern is that capital arrived ahead of the revenue curve, front-loaded the buildout, and then waited — sometimes a decade — for demand to catch up to the steel that had already been poured. A transformative technology is not evidence against a bubble. Historically, it has been one of the bubble's preconditions, because transformation is what makes the overshoot feel reasonable while it is happening.
The second thing it obscures is the arithmetic. Roughly three hundred to three hundred fifty billion dollars of annual capex against a frontier-lab revenue base in the low tens of billions. Even after you credit the cloud providers with the AI-attributable portion of their existing compute revenue — and you should, because some of that capex is genuinely monetizing — the capital-intensity ratio sits in territory that has no recent precedent outside of state infrastructure programs. The bull case requires the revenue line to accelerate meaningfully, not next year, but across the next several years, and to keep accelerating after the easy enterprise seats are filled. That is possible. It is not absurd. But it is an assumption, and the current price is a bet that the assumption is not merely possible but inevitable.
The third thing it obscures is the discount rate. An AI cash-flow story is an ultra-long-duration asset. Most of the present value sits far out on the curve, in years where the terminal assumptions live. Long-duration assets are exquisitely sensitive to the rate used to discount them, and we are not in the regime that made the 2010s so forgiving. The narrative and the rate environment are pulling in opposite directions, and the narrative has been winning because flows have been winning. Flows are not a valuation argument. Liquidity is a mirror, not a floor. It reflects the price back at you and it never once promises to hold it.
The fourth thing it obscures is where Marks would actually look. This is the part that frustrates me most about the coverage, because it is the part that requires knowing the man's day job. Marks is a credit investor. His sensitivity to bubble formation is tuned not to the equity multiple but to the debt stack. And the AI buildout has developed a debt stack: special-purpose vehicles for data center construction, securitized GPU receivables, private credit funds underwriting the intermediate layer, vendor financing arrangements in which the chip supplier invests in the customer who then buys more chips. That last structure has a name in every other market — circular financing — and it has a history. When the same balance sheet appears on both sides of a transaction, the revenue is real but the risk is not distributed. It is concentrated and hidden, and it is priced as if it were distributed.
If the AI capex cycle turns, I do not expect the first crack to appear in the equity market. I expect it in the credit market — in the spread on a data center asset-backed security, in the rollover terms of a construction loan, in the discount at which GPU-backed paper trades when the collateral is a depreciating semiconductor rather than a building. Equity markets can stay irrational longer than shorts can stay solvent. Credit markets cannot. Credit has a maturity date. When the refinancing window closes, the argument ends whether or not anyone has been persuaded.
The fifth thing it obscures is depreciation. This is the most under-discussed number in the entire sector and it does not appear in a single retail-facing article I have read this year. Cloud providers depreciate AI servers over roughly five to six years. If the economic life of that hardware is closer to two or three — because each generation of accelerator resets the performance baseline and the resale value of the previous generation collapses — then reported operating income is systematically overstated, not through fraud, but through an assumption. Assumption differences are legal. They are also worth tens of percentage points of earnings. A change in a depreciation schedule is one of the quietest disclosures a company can make and one of the loudest signals it can send. Silence in the code screams louder than volume.
The sixth thing it obscures is concentration. The largest technology companies now represent roughly a third of the index by market capitalization. That is a different order of magnitude from the single-sector concentration of 2000. It means an AI repricing is not a sector event. It is an index event, and an index event transmits through passive vehicles, through pension allocations, through the wealth effect that turns a drawdown in a brokerage account into a contraction in consumer spending. The transmission channel is mechanical and it does not require anyone to be wrong about technology.
The seventh thing it obscures is the game theory of the capex itself. Every hyperscaler knows the return on this spending is unproven. None of them can stop. Stopping means ceding the next workload migration to a competitor, and the next workload migration is a one-way door. So individual rationality produces collective overshoot, and the overshoot cannot be unwound by a rational, coordinated decision, because there is no coordinator. This is why I do not believe in a soft landing for the capex cycle. Soft landings require a brake, and the only brakes available here are external: the physical limit of power, the cost of financing, or demand that fails to materialize in a way too obvious to narrate away.
Power deserves its own sentence. Interconnection queues, transformer lead times, the signing cadence of nuclear power purchase agreements — these are the physical ceiling on how much compute can be built, regardless of how much capital wants to build it. That ceiling works in both directions. It is the reason the bubble cannot inflate indefinitely, and it is also the reason compute remains scarce and the incumbent position remains defensible. Both things are true at once, and anyone who tells you only one of them is selling something.
I have watched this exact shape before, in my own market, and it did not end the way the models said it would. During the 2020 DeFi summer I ran a personal book of about a hundred fifty thousand dollars across liquidity pools while the people around me chased four-digit APYs. I moved sixty percent into stablecoin pairs because the yield was boring and the mechanism was legible. That decision looked stupid for eight months and then it looked like the only decision anyone had made. The lesson was not 'avoid risk.' The lesson was that when a return is quoted without a mechanism, the mechanism is you. I have applied that filter to every AI revenue projection I have read since, and the filter has not failed yet.
I have watched the infrastructure version of it too. Rollups have spent three years advertising sub-cent transaction fees, and a meaningful portion of that subsidy came from blob space priced at a level the protocol was never designed to sustain. When that space saturates — and on the current demand curve it saturates within two years — the fee line goes back up, and a whole cohort of applications that only existed because the marginal cost was approximately zero discovers that it has a marginal cost. Cheap capital and cheap blockspace are the same drug. They make a business model look viable for exactly as long as the subsidy lasts, and then they make it look like a fraud, even though nothing about the underlying product changed.
Liquidity fragmentation is the other version of this. Every cycle, a new cohort of products is launched to solve the fragmentation problem, and every cycle the fragmentation turns out to be a feature of a market that has more venues than it has depth. The solution is not a new aggregator. The solution is that the market has not yet earned the right to be deep. The narrative exists because the product needs a problem, not because the problem needs a product.
The same pattern holds in the consensus layer. After the fourth halving, miner revenue compressed and hash power continued to concentrate, to the point where a handful of pools control the majority of the work. The network is still secure in the formal sense. The decentralization claim is thinner than the marketing suggests. A consensus that is concentrated is not consensus; it is coordination with extra steps. I raise this here because the AI buildout has an identical structural feature — the compute layer is concentrating into a very small number of firms, and the concentration is being driven by exactly the economics that make the buildout necessary. The industry that promises to decentralize intelligence is centralizing the substrate on which intelligence runs. The algorithm does not care about your conviction.
Now the part that the article will never say, and the reason I am writing this at all.
When a value investor of Marks's stature is quoted as saying a boom is not a bubble, the quote does enormous work in the distribution layer. It becomes the line a broker repeats to a hesitant client. It becomes the reply-guy sentence under a chart of a parabolic stock. It becomes, in aggregate, a small but measurable reduction in the perceived probability of loss — and the perceived probability of loss is the only variable that governs how much leverage enters a system. FOMO is the tax on unexamined desire, and the tax rate is set by headlines like this one.
There is a reflexivity trap underneath the whole debate, and almost nobody discusses it. If the AI-is-a-bubble consensus becomes strong enough, capital withdraws ahead of the revenue curve, the buildout slows, the applications that would have justified the capex never get built, and the bears turn out to be right for entirely the wrong reason. A narrative can be self-fulfilling in both directions. This is not an argument for optimism. It is an argument for humility about what any of us actually knows.
And there is a distributional question worth asking about the venue. This story surfaced in crypto media, not in a technology publication. Why? Because the audience overlap is total. The same risk appetite that buys a frontier model stock buys the token with the AI branding, and both books feel the same weather. 'AI is not a bubble' carries emotional value to that audience, because the unstated syllogism is: and neither is my book. I am not accusing anyone of bad faith. I am pointing out that the second-order transmission of a claim is never neutral, even when the first-order source is.
Which brings me to the thing I actually do with a piece like this. I do not use it. I index it. I put it in the column that tracks where the debate has migrated — from the specialists to the generalists, from the growth investors to the value investors, from the technology press to the financial press to the crypto press. Historically, that migration is not an early-cycle phenomenon. It is a middle-to-late-cycle phenomenon, because it means the argument has been settled by people who have stopped checking the assumptions and started checking the temperature of the room.
Here is what I am watching instead, and what I would tell anyone building a position on either side of this.
Watch the spot rental price of H100 and H200 compute. A sustained decline in rental rates is the earliest observable signal that demand has softened relative to supply, and it front-runs every earnings call by at least a quarter.
Watch the quarterly ratio of capex guidance to AI-attributable revenue run-rate. If that ratio stops widening, the market can tolerate almost anything. If it widens while the revenue line decelerates, the repricing is mechanical.
Watch data center private credit spreads and GPU-backed securitization issuance. Credit finds the fracture before equity does, because credit has a maturity date and equity does not.
Watch for a change in depreciation schedules. When a company extends the assumed useful life of its servers, it is telling you something about the return on those servers that no investor deck will ever state directly.
And watch the primary source. Go find the Marks memo or the Marks interview in full. Read the conditional clauses. Count how many of them the headline preserved. Between the block and the breath, truth resides — and in this case the breath is a price, a date, and a sentence about valuation that did not fit in the title.
The question I keep coming back to is not whether artificial intelligence will change the world. It will. The question is who pays for the change, at what entry point, and whether the people paying were told the price before or after they were told the verdict.
The ledger remembers what the market forgets. Right now the market has forgotten the price. Someone should write it down before the headline does.