Howard Marks on AI: The Most Consequential, Least Predictable Bet

In his November 2026 memo, Howard Marks poses 11 questions about investing in AI and concludes the stakes are huge while predictability is unusually low. The Oaktree Capital co-founder opens A Layman Thinks About AI, published November 11, 2026, by admitting he is not a tech person, then works through how AI works, who will profit, data centers, market concentration, the credit market, whether AI is a bubble, and how anyone should invest in something this hard to forecast. Along the way he asks Anthropic’s Claude to explain itself and to score the predictability of each corner of the AI industry.

TLDR

Marks argues that AI is too important for investors to ignore and too uncertain for anyone to forecast with confidence. An AI investor has to be at least somewhat right about the technology, demand for compute, competition, who captures the profits, financing and timing, and the odds of getting all of those right are slim. He warns that the U.S. economy, the S&P 500 and the investment grade bond market now lean heavily on one sector, with roughly $600 billion of AI-related debt issued this year. His answer is not to stay out but to act with calibrated humility: know which questions you can actually answer, size positions accordingly, and remember that price is what makes uncertainty acceptable.

Thoughts

The credit section is where this memo earns its keep, because debt is Oaktree’s home turf and the argument is one most AI commentary skips. Marks points out that hyperscalers which used to be debt-free, cash-rich and highly rated are now spending their cash hoards and borrowing heavily, that AI bonds already make up around 12% of the Bloomberg USD investment grade corporate index, and that these borrowers are rate-insensitive “takers” of capital because none of them can afford to slow down. His suggestion that this has started to crowd out other borrowers, even pushing up what the government pays, is offered as a question rather than a finding, and he lists inflation and deficits as separate drivers elsewhere, so it is fair to be cautious about how much of the rate rise to pin on data centers. Still, the structural points are hard to dismiss: lease out clauses tied to 99.9% uptime, chip collateral that may not outlive the loan, off-balance-sheet vehicles, and circular deals where a vendor finances the customer who buys its product. A lender who is buying a hyperscaler’s rating rather than the project’s own economics is the one most exposed if any of those assumptions crack.

The most quotable moment is Claude, Anthropic’s own model, rating the frontier labs a 2 out of 10 on predictability, the same band as early-stage venture capital, and calling the business “close to binary.” Marks enjoys the irony, but the deeper point is the mismatch he notes right after: the least predictable assets in the industry appear to be priced at a premium in private markets, and a large share of the S&P 500’s second-quarter earnings came from gains on stakes in those same companies. It is worth being clear about what the table is. It is a language model’s judgment calls, not a measurement, and Marks presents it that way. Its value is the ordering and the analogies, especially speculative “neocloud” data centers sitting next to telecom and fiber from 1998 to 2001, a comparison anyone who remembers the fiber glut should take seriously.

The concentration argument matters even to people who think they have no AI exposure. Marks notes that 42 AI-related companies produced roughly two-thirds of the S&P 500’s price gains since January 2024, according to J.P. Morgan, and that AI capital spending is a large part of GDP growth while other construction declines. An index investor is therefore an AI investor whether they chose to be or not. His Michael Milken comparison is sharp: the high yield market was built on diversification so that a few defaults would not sink a portfolio, and today’s stock and debt markets are leaning on a single outcome. That is the opposite of what diversification is supposed to deliver.

The data-center section punctures the comfortable idea that infrastructure is the safe, picks-and-shovels way into AI. Most of a data center’s cost is chips rather than the building, sites face local opposition and regulatory hurdles, and power, cooling, fiber and financing are all potential bottlenecks. Marks also catches a real paradox: the same moratoriums and constraints that make new projects harder should make existing, well-placed capacity more valuable. That makes the choice of which data center, which tenant and which lease terms more important than the decision to own data centers at all.

The ending reconciles the uncertainty with action through price. Marks says intelligent investing is largely about price, and that even the least predictable sectors can be good investments if they are cheap enough relative to intrinsic value, then closes with Warren Buffett’s 1979 line that “uncertainty actually is the friend of the buyer of long-term values.” The tension he leaves mostly implicit is that Buffett’s point depends on a gloomy consensus and low prices, while the frontier labs, by Marks’s own account, are being priced at a premium. His conclusion that the AI payoff question is “largely a matter of faith” is honest, and it suggests the practical answer for most investors is less about picking AI winners and more about knowing how much of their portfolio already depends on that faith.

Key Takeaways

Marks calls AI the most consequential and least predictable development for investors in at least a generation, and says uncertainty is not a reason to abstain.
Compounding forecasts is brutal: a 60% chance of being right on each of four independent calls gives only about a 13% chance of getting all four right, and AI involves even more variables.
Quoting Claude, he notes that AI systems are “grown rather than designed,” and that even researchers can explain only small slices of how they reach answers.
If several models become equally capable, AI could become a commodity sold on price, shifting the economic benefit to users rather than vendors.
Marks raises a question few are asking: if AI replaces taxpaying workers, governments may tax AI to replace that revenue.
His forecasting test: name an AI company’s earnings per share five years out, then ask whether you would bet the real number lands within 50% either way.
Data centers are not a safe “bricks and mortar” play, because chips make up most of the cost and power, sites, fiber and financing can all become bottlenecks.
According to J.P. Morgan, 42 AI-related companies generated around two-thirds of the S&P 500’s price gains since January 2024.
Goldman Sachs counts nearly $600 billion of AI-related debt issued globally this year, around $400 billion of it in U.S. investment grade, more than a fifth of total IG supply.
AI may need about $1 trillion of new capital a year, which could compete with other borrowers and strain the equity market, where U.S. IPOs raised under $50 billion in eight of the last ten years.
Lenders should watch lease out clauses, chip collateral that may wear out before the debt, off-balance-sheet vehicles, and circular vendor financing.
Claude scored AI overall about a 3 out of 10 on predictability, with the direction of the technology around 6 or 7 and the things investors get paid for around 2.
In that ranking, regulated utilities serving AI load were the most predictable segment and frontier labs such as OpenAI, Anthropic and xAI the least.
Marks defines a bubble as irrational exuberance leading to imprudent behavior, excessive investment and too-high prices, and says whether AI qualifies is still largely a matter of faith.

Detailed Summary

Why a Self-Described Layman Weighs In

Marks opens by admitting he has no special technical knowledge, but argues AI is too big for any investor to skip thinking about. Investing, he says, is positioning capital to benefit from future events, so every AI decision rests on forecasts and on the odds those forecasts are right. An AI investor has to be at least somewhat right about the technology, demand for compute, competition, the beneficiaries, financing and timing. He reminds readers that the biggest mistakes come from thinking you know everything, yet incomplete knowledge cannot become an excuse for doing nothing.

Nobody Fully Knows How AI Works

For the first question, Marks turns to Claude, which explains that we know how models are built and trained but little about how they actually produce answers. Training nudges billions of internal settings until rules emerge that nobody wrote, and interpretability research can explain only small slices, such as a model choosing a rhyme word before writing the line. Claude adds that a model’s explanation of its own reasoning may not match what happened inside. Marks asks whether anyone can reliably answer the investment questions that follow if few people understand the machine itself.

What AI Will Be Able to Do

AI has moved from pattern recognition and extrapolation, which already changed fields like coding, to something that can fairly be called reasoning. Marks lists the open questions: whether AI can reflect human values, stay honest and admit what it does not know, offer skepticism instead of sycophancy, and handle creativity and what he calls second-level thinking. He also asks whether one model will do everything and what artificial general intelligence should even mean, average person, 90th percentile, or the smartest person alive. The answers will decide how much work gets handed to AI.

Who Profits and Who Leads

Citing Warren Buffett’s point about the internet, Marks notes that transformative technologies do not always reward investors. AI could become a commodity that grinds down returns, or a few companies could become essential monopolies, as Nvidia and TSMC look today. He wonders how AI will be taxed if it replaces workers. On competition, he notes that OpenAI and Anthropic were unknown to most people a few years ago, and that the leaders of 20 to 25 years ago included winners like Microsoft and Amazon and casualties like Napster, Yahoo and Netscape.

Forecasting AI Company Earnings

Valuing a company means projecting units, prices, costs, reinvestment and returns, and AI adds unusual unknowns: what the product will do, whether it is priced as a commodity, how long chips last, how fast data centers must be depreciated, and how long the spending race continues. Marks says he never faced uncertainty this large analyzing conglomerates in the 1970s or the internet in 1999 and 2000. His test is to forecast an AI company’s earnings per share five years out and ask whether you would bet on landing within 50% either way.

Data Centers Are Not a Layup

Data centers, or AI factories, look like the cautious picks-and-shovels way to invest, since any winner will need compute and compute is scarce today. But the building shell is a small part of the cost, with GPUs making up most of it, and builders face site acquisition, local resistance, fiber connectivity, power, cooling, construction and financing constraints. Those same constraints, including recently passed moratoriums, could make existing capacity more valuable. Marks expects the gap between winners and losers to be wider than usual for infrastructure.

An Economy Leaning on One Sector

AI is a small part of the U.S. economy but an outsized source of growth, mostly through data-center construction. The Magnificent 7 grew earnings faster than the rest of the S&P 500 for 14 straight quarters, and FactSet reported that Alphabet and Amazon booked $151 billion of second-quarter gains on equity stakes, mostly in Anthropic and SpaceX, which now includes xAI. Marks calls AI exposure a fault line running through portfolios and contrasts it with Michael Milken’s high yield thesis, which depended on diversification.

A New Era for AI Lenders

AI-related borrowing has reached nearly $600 billion this year, about 12% of the Bloomberg USD investment grade corporate index and around 6% of U.S. high yield. Marks asks whether rate-insensitive hyperscalers are crowding out other borrowers, whether the equity market can supply hundreds of billions more each year, and whether the hyperscalers’ ratings will hold as they add leverage. He flags lease out clauses, chip collateral, circular transactions and off-balance-sheet vehicles, and concludes lenders must be exceptionally vigilant.

Rating AI’s Predictability

Drawing on his book Mastering the Market Cycle, Marks says a forecast needs a view on its own probability of being right. Claude rated AI as a whole about 3 out of 10, then scored subsectors: regulated power utilities serving AI load 6, long-term leased data centers 5 to 6, hyperscalers 5, chips 4, speculative neocloud data centers 2 to 3, and frontier labs 2. Marks also flags geopolitics, from TSMC’s role in Taiwan to U.S. export limits pushing China toward cheap competing models.

Is It a Bubble, and How to Invest

Recent worries about inflation, deficits, rates, the war in Iran, cheaper and open-source models, falling token prices and heavy leverage have darkened the mood. Marks says the technology is real, but whether the spending and the prices are justified depends on whether AI creates enough value to earn adequate returns on trillions of dollars, which for now is largely a matter of faith. He urges investors to ask which questions they are equipped to answer, how many things must go right, what they lose if wrong, and how much optimism is already in the price, and he gives Buffett the last word.

Notable Quotes

AI is the biggest development of our time and could be one of the most impactful things the world has ever seen, so investors don’t have the option to abstain from thinking about it.
Howard Marks, on why a non-tech investor has to engage with AI

Tell me what you think the e.p.s. for a given AI company will be in five years. Will you bet that the eventual answer will be within +/- 50% of your figure?
Howard Marks, proposing a test for AI earnings forecasts

But today, our stock and debt markets are heavily reliant on a single factor: the success of AI.
Howard Marks, contrasting today’s markets with Michael Milken’s diversification thesis

These businesses are close to binary. Capital needs are huge, prices per unit of capability keep falling fast, and it’s unclear whether a lasting moat exists.
Claude, as quoted by Marks, on frontier AI labs

Whereas in real life things fluctuate between pretty good and not so hot, in the minds of investors they go from flawless to hopeless.
Howard Marks, on the recent shift in mood around AI

The future is never clear; you pay a very high price in the stock market for a cheery consensus. Uncertainty actually is the friend of the buyer of long-term values.
Warren Buffett in Forbes, August 6, 1979, quoted by Marks to close the memo

Read the full memo here.