NextFin News - Howard Marks, the Oaktree Capital Management co-founder whose warnings ahead of the dot-com crash and the global financial crisis made him one of investing's most watched contrarians, is declining to call artificial intelligence a bubble - and he says that uncertainty is exactly what investors should sit with. In his latest memo and a series of interviews this month, Marks argues that optimism has dominated markets since late 2022, that enthusiasm around AI is undeniable, but that too many unknowns remain over profitability, valuations, and the technology's impact on jobs to say whether the exuberance is rational or not.
The tension behind his caution is stark. The S&P 500 is trading around 7,650, within a few percent of its all-time closing high of 7,798.99 set in August, and up roughly 12% from the end of 2025. Nvidia - the pick-and-shovel supplier of the AI buildout - carries a market value of about $5.4 trillion, larger than the combined capitalization of five S&P 500 sectors. Behind those prices sits a capital-spending wave unlike anything in living memory: the five largest cloud and AI infrastructure companies have committed to more than $600 billion of spending in 2026, a roughly 36% jump from a year ago. Marks's refusal to resolve the bubble question is not an evasion. It is a deliberate statement about what can and cannot be known when a new technology has no history to constrain imagination.
What Marks Actually Said - and What He Did Not Say
The memo, prompted by questions from clients in Asia and the Middle East, opens with Marks's standard disclaimers: he is not active in the stock market, merely watching it as "the best barometer of investor psychology," and he is "no techie." From there he makes a distinction that much of the commentary has flattened. "I've concluded there are two different but interrelated bubble possibilities to think about: one in the behavior of companies within the industry, and the other in how investors are behaving with regard to the industry," he writes. He then declines to judge the first and focuses on the second - whether the financial world is pricing AI with excessive optimism.
Market bubbles aren't caused directly by technological or financial developments. Rather, they result from the application of excessive optimism to those developments.
That sentence does the heavy lifting. Marks is not arguing that AI is a fraud or that the technology will disappoint. His point is narrower and harder to refute: a real revolution can still produce a financial mania if investors pay prices that assume a limitless future. "Futures that are perceived to be limitless can justify valuations that go well beyond past norms - leading to asset prices that aren't justified on the basis of predictable earning power," he writes. The mechanism is not the technology; it is the removal of history as a disciplining force. Because there is no track record for frontier AI, nothing restrains what investors imagine it can earn.
He has lived through this before. Marks notes that two of his best calls came in 2000, on tech and internet stocks, and in 2005-07, on subprime credit - and that in neither case did he possess expertise in the underlying subject. What he observed was behavior, not technology. The value of those calls, he says, "consisted mostly of describing the folly in that behavior, not in insisting that it had brought on a bubble." Applied to AI, that means the actionable question is not "is this a bubble?" but "how should I behave given that I cannot know?"
His framing echoes a point he made in an interview in March, when he said he does not believe AI has human intuition - the "hairs on the back of your neck" reaction a reader gets from a risky prospectus. That is a limitation of the technology. The bubble question is a limitation of the investor.
The Money Behind the Mania - and Where It Turns Risky
If the behavior Marks is watching has a dollar figure, it is the hyperscaler capital budget. Amazon, Alphabet, Microsoft, Meta, and Oracle have collectively committed to more than $600 billion of spending in 2026, with estimates running as high as $725 billion. Roughly three-quarters of that, by one credit research estimate, goes directly to AI infrastructure - GPUs, servers, networking gear, and data centers. The spending is real, it is contracted, and it is already showing up in supplier revenue: cloud growth at the big providers is running from 24% to 48% year over year.
But the composition of the financing is shifting, and that is where Marks's warning sharpens. "To date, much of the investment in AI and the supporting infrastructure has consisted of equity capital derived from operating cash flow," he writes. "But now, companies are committing amounts that require debt financing, and for some of those companies, the investments and leverage have to be described as aggressive." Equity funded from cash flow absorbs mistakes silently. Debt does not. When the capital stack flips from retained earnings to borrowed money, the tolerance for a slow payoff collapses, and the reckoning moves from a future earnings disappointment to a present-day refinancing event.
One key risk to consider is the possibility that the boom in data center construction will result in a glut. Some data centers may be rendered uneconomic, and some owners may go bankrupt.
Marks is not standing aside entirely. Oaktree has made a few data-center investments, and its parent, Brookfield, is raising a $10 billion fund for AI infrastructure backed by its own capital and equity commitments from sovereign wealth funds and Nvidia, with "prudent" debt layered on top. That is the tell: the investor who refuses to call the market is still willing to own the infrastructure - just on a balance sheet he controls, with a partner whose chips sit at the center of the buildout. It is a bet on the railroad, not on which trains will run on it.
The broader financing architecture points the same way. The chip supplier at the center of the buildout has partnered with six major asset managers to establish compute-infrastructure financing platforms intended to mobilize more than $500 billion of third-party capital. Leverage is not a side feature of this cycle; it is the design.
Cyclical Wave on a Structural Shift - Keeping the Two Separate
The central analytical question is whether today's AI uncertainty is cyclical - a mean-reverting surge of optimism that will fade - or structural - a regime change that will not correct on its own. The answer is both, and conflating them produces the wrong conclusion.
The structural leg is the technology itself. AI is not a fad, and Marks has said as much: the power, speed, and autonomy of the systems are real, and the productivity implications are not going to reverse. A structural shift does not need to be fully monetized to be real; the railroad transformed the economy long before every line laid in the 1860s earned a return. On this dimension, waiting for certainty before participating is its own form of risk.
The cyclical leg is the financing and valuation wave wrapped around the technology. This is the part that mean-reverts. Three historical comparisons make the pattern visible. In the 1840s railway mania, historians of the episode note that more new rail lines were authorized in the peak years than Britain would build in the following century; most of the promoters went bankrupt, yet rail transport became foundational. In the late-1990s fiber-optic buildout, companies laid enough dark fiber to carry traffic many times over for a decade; the builders were wiped out, but the glut of cheap bandwidth became the substrate of the broadband internet. In 2005-07, subprime mortgages were packaged into securities rated on the assumption that national home prices could not fall together; they did, and the structure collapsed even though housing itself did not disappear.
The recurring mechanism is the same in each case: a genuine, durable innovation attracts capital on terms that assume the innovation's success is both certain and imminent. The innovators and their customers may prosper over time. The financiers who paid for capacity ahead of demand often do not. That is the distinction Marks is drawing when he separates company behavior from investor behavior. AI the technology can be structural while AI the trade is cyclical - and the cyclical leg is the one currently setting asset prices.
The Second-Order Question the Market Is Not Asking
The first-order read of Marks's caution is obvious: be careful, valuations are rich, do not chase. The market has priced that. What the market has not fully priced is the second-order consequence of the financing shift he identifies - from equity to debt - and what it does to the timing of the reckoning.
When AI investment was funded from operating cash flow, the feedback loop was slow. A data center built ahead of demand sat half-empty; the parent company absorbed the idle capacity; earnings growth slowed; the stock re-rated over quarters. Now, with debt financing and third-party capital pools - including the more than $500 billion in compute-infrastructure financing partnerships anchored by the chip supplier and major asset managers - the feedback loop tightens. A project that cannot service its debt does not wait for a multi-year earnings cycle to be judged. It defaults, the lender forecloses, and the asset is repriced immediately.
This creates an asymmetry that equity investors are underweighting. The upside from AI adoption is diffuse and accrues over years; the downside from a financing break is concentrated and arrives on a covenant date. In other words, the distribution of outcomes is becoming more skewed: the technology can be right and the capital structure can still blow up. That is the lesson of the fiber-optic glut - the world did use the bandwidth eventually, but the bondholders who funded the build did not live to collect the payoff.
There is also a cross-asset transmission channel. If data-center owners begin to default, the losses do not stay in private hands. They transmit to the lenders, to the credit funds that hold the loans, and to the valuation marks on private portfolios that mark only quarterly. Public equities reprice in an afternoon; private credit reprices whenever the next valuation committee meets. That lag is not a cushion - it is a source of tail risk, because by the time the loss is visible, the capital that could have absorbed it has already been committed elsewhere.
The Strongest Counter-Thesis - and What Would Prove Marks Wrong
The most serious challenge to Marks's caution comes from the valuation side, not the technology side. Nvidia, at a market value of roughly $5.4 trillion, trades at about 20 times forward earnings - approximately half the multiple it carried three years ago, before AI, and near the cheapest valuation it has posted since late 2018. The argument is straightforward: if earnings grow into the multiple, today's prices are not a bubble; they are a fair discount rate applied to a real earnings stream. Add the fact that the hyperscalers funding the buildout are profitable, generating large cash flows, and that their cloud businesses are growing 24% to 48% annually, and the dot-com comparison starts to fray. The 1999 market was pricing companies with no earnings. Today's market is pricing companies with record earnings.
This counter-thesis is backed by a mainstream view. "The AI earnings-driven tech boom continues," one portfolio chief said of the record-setting market in August. "It's an earnings boom, not a bubble." It attacks Marks's argument at its foundation: if prices are justified by predictable earning power, then the "excessive optimism" mechanism never engages, and the bubble framework does not apply.
The falsifying signal is quantifiable. Marks's thesis holds if the ratio of committed AI capital to realized AI revenue fails to converge over the next four to six quarters. Specifically: if hyperscaler capital expenditure remains above $600 billion annually while AI-attributable revenue across the same companies grows below 30% year over year, and if data-center utilization metrics show capacity being built ahead of contracted demand, then the financing is running ahead of the economics and the cyclical leg of the trade is in bubble territory. Conversely, if AI revenue grows at or above the pace of spending for four consecutive quarters, the earnings-boom thesis wins and Marks's caution will have been too early.
What Comes Next - Split by Time Horizon
Short term (sentiment and liquidity): The market is positioned for continued optimism, with the S&P 500 near records and AI-linked names carrying outsized weight. The near-term risk is a liquidity or financing event - a delayed project, a refinancing miss, a credit mark-down - that forces a rapid de-risking. The trigger to watch is any sign that debt-funded AI projects are being repriced or delayed.
Medium term (fundamentals): The base case is that AI revenue continues to grow strongly but not fast enough to validate the full capital stack immediately. In that scenario, returns compress in the infrastructure layer while the chip suppliers and the hyperscalers with pricing power continue to earn. The exposed are the highly levered data-center developers and the lenders that funded them; the beneficiaries are the asset owners who can wait out the cycle and the companies that capture the productivity gains without owning the infrastructure.
Long term (structural): The technology itself is likely to deliver transformative productivity, and the economy will absorb the capacity - just as it absorbed the railroads and the fiber-optic networks. The long-term winners will be the companies that use AI to raise margins, not necessarily the ones that built the compute. Marks's own posture captures this: own the infrastructure on prudent terms, stay selective in the equity market, and do not confuse a correct view of the technology with a correct view of the price.
Scenarios: The base case is a cyclical drawdown in AI infrastructure valuations within a structural uptrend for the technology. The upside case is that AI revenue inflects above spending growth, validating the capital stack and extending the rally. The downside case is a credit event in data-center financing that forces a broader de-risking across private and public markets.
The market's question is whether AI is a bubble. Marks's answer - that the question itself cannot yet be answered - may be the more useful one. In a mania, certainty is the commodity everyone is selling; the investor who can tolerate not knowing is the one least likely to pay for it.
Explore more exclusive insights at nextfin.ai.
