A headline crossed my feed this week: "Howard Marks is not convinced the AI boom is a bubble." It did not appear in a value letter. It appeared on a crypto news site. That is the first thing worth pricing.
The second is the payload itself: four usable information points, three of them restating the same claim. No direct quote. No date. No venue — memo, podcast, or interview, unspecified. A paraphrase of a paraphrase, wrapped in a title that reads like permission.
I have spent thirteen years watching how a thesis gets compressed on its way to retail. In 2017 I audited fifteen ICO whitepapers during the Ethereum mania and found one whose market cap ran 300% past any defensible utility value. The tell was never the technology. It was the funding structure, and how many marginal buyers were left. So when a distressed-credit legend gets flattened into "not a bubble" on a risk-asset site, I stop reading the sentence and start reading the plumbing.
Marks is not a technologist and has never claimed to be. He is a credit investor who built Oaktree on the observation that the same asset is an opportunity at one price and a trap at another. His bubble test, applied consistently across three decades of memos, has three parts: psychology running hot, price already discounting the optimistic case in full, and the marginal buyer exhausted. Nothing in that framework says a technology is fake. Everything in it says the answer depends on the entry price.
Strip the price and you have not summarized Marks. You have inverted him. The body of the piece reportedly still carries the caveat that investors should be wary of stretched valuations — the opposite operational instruction from the headline. That gap is not a factual error, so no fact-checker will ever catch it. It is a framing error, and framing errors propagate faster than corrections do.
Now the numbers the piece never touches. North American hyperscaler capex guidance has been running in the $300–350 billion range annually, and I would flag every figure in this paragraph as needing verification before anyone sizes a position on it. Against that, the annualized revenue of the leading model developers sits in the low tens of billions, combined. Even crediting the AI increment inside cloud segments, the ratio of capital deployed to revenue recognized has no close historical precedent outside of state infrastructure programs.
Concentration matters more than the ratio. The largest technology names now represent roughly a third of S&P 500 capitalization — a figure I would also verify. That converts an AI repricing from a sector event into an index event, transmitted through passive flows, pension allocations, and household wealth effects. This is the part of the debate that has nothing to do with whether the models are good.
Here is where my own work colors the read. In 2020 I ran a backtest on Aave v2 yield strategies for a Nordic fintech and found impermanent loss quietly erasing 40% of headline APY in volatile pairs. In 2022 I watched TerraUSD break and mapped the depeg against dollar index spikes. Both taught the same lesson from different angles: the durability of a yield depends on the structure that funds it, not the narrative that advertises it. AI capex is a yield story told in reverse — capital deployed today against revenue that must arrive later, and the structure funding it is where I would look first.
Start with accounting. Cloud providers depreciate AI servers over roughly five to six years. The economic life of that silicon, given the cadence of accelerator releases, is plausibly two to three. This is not fraud; it is an assumption. But assumptions of that size move earnings by tens of percentage points, and a depreciation schedule is the cheapest place in a capital cycle to hide an optimism you would never say out loud. No discussion of the bubble question I have read this year touches it. That silence is itself informative.
The funding side is where it gets uglier. This cycle differs from 1999 in one structural respect: a meaningful share of the build-out runs through off-balance-sheet vehicles, joint ventures, private credit, and vendor financing — chip suppliers investing in customers who then buy more chips. Behind every transaction is a map of human greed, and circular financing is that map drawn in a circle. It makes the risk opaque to equity holders and legible to creditors. Which means the first crack almost certainly appears in credit spreads on data-center paper and GPU-backed receivables, not in a forward P/E multiple.
The plumbing has physical limits too, and they cut both ways. Grid interconnection queues, transformer lead times, and nuclear power purchase agreements cap how fast capex can compound — a brake on the bubble and a support for compute scarcity at the same time. Both readings can be true. They need to be verified separately. The brake is not symmetric with the support, though: queue data is public and PPA data is not.
One more parallel from my 2024 work on the spot bitcoin ETFs. I correlated IBIT inflows with Federal Reserve balance sheet expansion and concluded the products were less a new asset class than a liquidity conduit. The same lens applies here. AI capex is a long-duration cash flow story financed at the short end, and long-duration assets are brutally sensitive to discount rates. A bubble debate that never mentions the rate path is not a debate about valuation. It is a debate about belief.
Which is why the crypto venue is not incidental. AI equities and crypto majors have traded as the same risk-appetite asset in different clothing for three years now. "AI is not a bubble" is not neutral information to that audience. It is permission. In a bear market, permission is the most expensive thing you can hand a leveraged audience.
Now the part I will not let the bulls or the bears have.
"AI is real" and "AI equity holders make money" are independent propositions, and history is unambiguous about their coexistence. Railways were transformative and railway investors were wiped out. Radio arrived and radio stocks did not survive. The internet of 1999 was real, and capital deployed near the peak of that build-out produced a decade of negative real returns. If Marks believes in AI's transformative potential, that belief is not evidence against a bubble. It is the definitional precondition of one: real narrative, capital front-run.
But the inversion cuts the other way too. If "AI is a bubble" becomes consensus, capital withdraws before the technology finishes deploying — and the narrative genuinely slows the thing it correctly identified. Being right about a bubble early can destroy the asset you were right about. That reflexivity is harder to trade than any valuation multiple.
There is also an interest map to read. The loudest voices insisting there is no bubble tend to sit inside the capex chain — sell-side coverage, compute suppliers, cloud incumbents, and outlets whose audiences hold correlated risk. And when traditional value investors speak publicly on AI, note the timing. The pivot was not a retreat, but a recalibration — and public statements from that camp usually arrive when they must explain the miss, not when they are building the position.
So here is where I stand for the next four quarters, and it is not a prediction.
Track four numbers, none of which is a stock price: GPU rental spot rates on H100 and H200, where a sustained decline is the earliest honest demand signal; spreads on data-center ABS and private credit; the cadence of power purchase agreements against announced capex; and any downward revision to server depreciation schedules, which is the cleanest tell on earnings quality.
Yields are not gifts; they are risks wearing suits. The AI trade has spent three years being sold as a gift. We do not predict the wave; we engineer the vessel. The question for 2026 is not whether the technology is real. It is who is holding the depreciation schedule when the marginal buyer stops answering.
