Howard Marks on AI: 11 Questions Before You Invest

Howard Marks's new memo on AI contains no forecast. He opens by saying he has "absolutely no knowledge" of the subject, then spends eleven sections asking questions he can't answer. That is exactly why it's worth reading.

The number that carries the memo is 13%. Give yourself a 60% chance of being right on each of four things (the economy, the industry, the company, the market) and the odds of getting all four right are about 13%. AI investing needs more links than that to hold, and almost none is knowable today.

This post turns "A Layman Thinks About AI," the Oaktree Capital memo dated 11 November 2026, into something usable: the arithmetic of "how many things must go right," the concentration and credit numbers behind the AI boom, and the checklist Marks says to run before you buy. In short:

• Your odds are lower than your confidence suggests
• 42 AI-related companies drove about two-thirds of recent S&P 500 gains
• AI bonds are already about 12% of the US investment-grade index
• The most exciting part of the stack is the least predictable
• Price, not conviction, is your protection

The Memo's Real Product: Odds, Not Answers

Marks, Oaktree's co-chairman, starts from a simple idea. Investing means positioning for future events, so you need both a forecast and an honest estimate of the odds it's right. For AI, what has to go right runs from the technology and compute demand to competition, who benefits, financing and timing. Table 1 extends his four-factor math to AI.

Table 1: How fast the odds fall as links are added

Source: Marks's 60% × 4 example (Oaktree memo, 11 Nov 2026); the other cells are the author's arithmetic and assume independent outcomes, as Marks does. The memo prints "(0.64)" beside the 13% figure; 0.6 to the fourth power is 0.1296, so 13% is right.

His conclusion isn't to wait for clarity. It is to accept that you can't know everything and that some risk has to be borne. The rest of the memo is about which risks, and at what price.

The Technology Nobody Fully Understands

Marks went to Anthropic's Claude for the technical explanations. Models learn by predicting the next word and adjusting billions of internal settings, trillions of times. Nobody writes the rules. Interpretability research has caught a model choosing a poem's rhyming word before writing the line, and describing a textbook method for arithmetic while doing something else internally.

His open questions are the ones that decide the investment case: alignment, honesty, creativity, whether AI can do second-level thinking (his term for thinking differently from the crowd), whether one model or many specialized ones win, and what "AGI" even means: average person, 90th percentile, or the smartest human.

Who Profits? Not Necessarily the Inventors

Marks reminds readers of Warren Buffett's point that world-changing technology doesn't automatically reward its investors. If models become equally smart, price competition could hand the benefit to users. If a few become essential, they can charge a lot. Nvidia and TSMC show the second case is real.

One of his lighter lines carries weight: if AI replaces workers, governments may tax it like workers. Many top earners pay combined rates of 50% or more. "What tax bracket will AI be in?"

On competition, OpenAI and Anthropic were barely known a few years ago. The previous era's winners (Microsoft, Amazon, Apple, Google) were not obvious early, and Napster, Yahoo and Netscape faded. His question: what is the probability that today's leaders are still leaders in five to ten years?

The data-center "layup" that isn't

"Picks and shovels" sounds safe: if you can't pick the winner, own the infrastructure. Marks pushes back. The building shell is a small share of the cost. Chips (GPUs) make up most of it, and the obsolescence risk generally sits with the user. Siting, power, cooling, fiber, local opposition and financing all constrain supply. Recent moratoriums on construction could make existing centers more valuable, but overbuilding remains a risk, and winners and losers may diverge more than usual.

AI Is a Fault Line Through Your Portfolio

Marks calls AI a small part of the US economy but an outsized part of its growth, largely through data-center spending while other construction declines. The market numbers he cites make the same point.

Table 2: How concentrated the AI trade already is

Source: Oaktree memo, "A Layman Thinks About AI" (11 Nov 2026), figures as cited there. I have not re-derived them. The under-a-tenth comparison is the author's, and the memo does not say whether all 42 sit in the index.

Marks contrasts this with Michael Milken's high-yield thesis of the late 1970s, which rested on diversification across many bonds. Today's market, he suggests, leans heavily on one factor.

What AI Means for Credit

This is where the memo overlaps with our recent look at why rates are rising. Marks cites Goldman Sachs: nearly $600 billion of AI-related debt issued globally so far this year, about $400 billion of it in the US investment-grade market. Table 3 shows how big that is.

Table 3: The scale of AI borrowing

Source: Oaktree memo; debt issuance cited to Goldman Sachs, index weights and capital needs as stated by Marks. The 5% and 14% ratios are the author's arithmetic. Oaktree's Wayne Dahl cited AI capex above $5 trillion by 2030, which works out to a similar $1 trillion a year.

Marks flags that rate-insensitive hyperscalers may push up government borrowing costs, the same competition-for-capital effect behind higher long-end yields. Hyperscalers, once cash-rich and nearly debt-free, are now borrowing and spending cash, which could change their credit ratings. He also lists the structural risks lenders should watch:

• Lease "out clauses," for example if a center misses 99.9% reliability
• Chip collateral that may not outlast the debt
• "Circular" deals among chipmakers, hyperscalers, labs and builders
• Off-balance-sheet special-purpose vehicles

The Predictability Ladder

Marks asked Claude to score how predictable AI is, from 1 (unknowable) to 10 (fully predictable). Overall AI scored about 3. The direction of the technology scored 6 or 7. What investors get paid for scored about 2. You can be right about the technology and wrong about the investment. Table 4 breaks it down by segment.

Table 4: Predictability by segment

Source: Oaktree memo; scores are Claude's own "judgment calls" as reported by Marks, not Oaktree's. Sizing column is the author's gloss.

Is AI a Bubble? Marks's Three Tests

Marks lists a nasty backdrop: inflation, federal deficits and debt, higher interest rates, and the seeming intractability of the war in Iran. Then he asks whether AI is a bubble and refuses a yes or no. He wants three separate questions answered: Is the technology overhyped and likely to disappoint? Is spending excessive and fated to produce inadequate returns? Are investors taking too much risk and paying too much?

His stance: AI is not an illusion, but spending and prices can't be judged until usage and profit pools are clearer. Whether AI delivers enough value to justify the returns demanded on trillions of dollars is, he writes, largely a matter of faith.

The Five-Question Checklist

Marks closes with questions he says apply to every investment. I've paraphrased them and added where AI makes each one hard.

Table 5: The five-question checklist

Source: Oaktree memo, Question Eleven, paraphrased; third column is the author's.

Marks adds that high uncertainty can be acceptable if the price is low enough. The problem is that AI makes the "price is low enough" calculation unusually hard.

What I'd Watch

• Price per unit of AI capability. Marks notes it keeps falling fast. If revenue growth stops outrunning price declines, the economics change.
• Hyperscaler credit ratings. Cash-rich borrowers turned heavy spenders can be downgraded.
• Chip useful life versus debt tenor. A loan outliving its collateral is the quiet risk.
• AI's share of the IG index. About 12% now. A bigger share means more of your "safe" bond fund is the same bet.
• Construction limits. Moratoriums raise the value of existing sites and lower the odds of an oversupply glut.

What If I'm Wrong?

• The scores come from a model. Marks used Anthropic's Claude for the technical explanations and the predictability ladder, and Anthropic itself is one of the companies being scored (a 2, among the lowest). Treat the scores as a thinking aid, not data.
• The independence assumption cuts both ways. Table 1 multiplies probabilities as if links are independent. If they're positively linked, the whole chain succeeds more often than the table says, and also fails together.
• "No forecast" has a cost. If you must allocate capital, humility isn't a position. Marks's answer is price and sizing, and that works only if you can judge price.
• Figures are as Marks cites them. The Goldman, FactSet and J.P. Morgan numbers come from the memo. I haven't re-derived them.
• This is a fresh memo. It is dated 11 November 2026 and the data moves daily.

Closing Thoughts

Marks gives Warren Buffett the last word, from a 1979 Forbes piece: you pay a very high price in the stock market for a cheery consensus, and uncertainty is the friend of the buyer of long-term values.

The corollary is the one to hold onto. Uncertainty only helps if the price already reflects it. An investor who pays a confident price for an unknowable outcome has bought the opposite of Buffett's advice.

So run the test on your own AI exposure, direct or through an index fund. What would you bet your five-year earnings guess lands within 50% of?

This article is educational and is not personalized financial advice. Investing involves risk of loss, and past performance does not predict future results. It summarizes and comments on a publicly available Oaktree Capital memo; views attributed to Howard Marks are his, figures are as he cites them, and the analysis beyond that is the author's own. Data as of 12 November 2026.