A Layman Thinks About AI

The memo is organized as eleven questions. Marks says up front he is “not a tech person,” and he leans on Claude twice: once to explain how models work, and once to score the predictability of each part of the AI industry.

Key claims

• More variables means lower odds of being right. Marks reprises his top-down arithmetic: with a 60% chance of getting each of economy, industry, company and market right, the chance of getting all four is 13% (0.6^4). An AI investor has to be “at least somewhat right” about even more: the technology’s evolution, demand for compute, competitive structure, who benefits, the availability of financing, and timing.
• Who profits is open. Citing Warren Buffett on the internet, transformative technologies “don’t necessarily reward investors.” If several models become equally capable, model access could become an undifferentiated commodity and the benefits go to users, not vendors. The counter-case is monopoly-like labs; for now Nvidia and TSMC are “racking up the big profits.” He also raises an under-discussed question: if AI replaces taxpaying workers, will governments tax AI “like people”?
• Leadership is fragile. OpenAI and Anthropic were unknown to most people a few years ago; if they can rise that fast, they can be challenged by the hyperscalers (Amazon, Microsoft, Alphabet, Meta, Oracle) or by companies that don’t exist yet. Microsoft, Amazon, Apple and Google survived the last 20-25 years; Napster, Yahoo and Netscape did not.
• A concrete forecasting test. Name a five-year EPS figure for an AI company and ask whether you’d bet the actual result lands within ±50% (guess $10, outcome between $5 and $15). Marks says the conglomerates of the 1970s and e-commerce in 1999-2000 “seem simple and straightforward compared to AI.”
• Data centers are not a layup. The Picks and Shovels logic is real, but the building “shell” is a small part of a data center’s cost; GPUs are most of it, with obsolescence risk usually borne by the tenant. Land, permitting, local opposition, fiber, power, cooling and financing are all constraints. Those same bottlenecks (including a wave of local data-center moratoriums the week before) make existing capacity more valuable.
• The market has an AI “fault line.” The Magnificent 7 out-grew the rest of the S&P 500 on earnings for 14 consecutive quarters through June 30, 2026. Per FactSet, Alphabet and Amazon booked $151 billion of Q2 gains on equity stakes, mostly Anthropic and SpaceX (now including xAI). Per J.P. Morgan, 42 AI-related companies produced about two-thirds of the S&P 500’s price gains since January 2024. Contrast Michael Milken’s high-yield thesis, which depended on diversification; today markets lean on a single factor.
• Credit is where the strain shows. Per Goldman Sachs, AI-related debt issuance is nearly $600 billion globally in 2026 so far, about $400 billion in US investment grade (over one-fifth of IG supply), taking AI to about 12% of the Bloomberg USD IG Corporate index and about 6% of high yield. AI likely needs about $1 trillion of new capital a year. Rate-insensitive hyperscalers could crowd out other borrowers, including the government. Traditional US IPOs raised under $50 billion in eight of the last ten years (about $80 billion in 2020, $140 billion in 2021), so debt and equity appetites may compete. He flags data-center lease “out clauses” (e.g. termination if uptime falls below 99.9%), chip collateral that may not outlast the debt, off-balance-sheet SPVs, and “circular transactions” where vendors finance their own buyers.
• Claude’s predictability scores. Asked to rate AI on a 1 (unknowable) to 10 (fully predictable) scale, Claude gave AI overall about a 3: direction a 6 or 7, but the things investors are paid for (speed, size of profits, who keeps them) around a 2. Frontier labs scored 2, comparable to early-stage VC. Marks notes the least predictable assets carry the highest private-market premiums, and that OpenAI and Anthropic have both filed for IPOs.
• Is it a bubble? “Largely a matter of faith.” His definition: “irrational exuberance leading to imprudent behavior, excessive investment, and too-high prices for participation.” Nothing has changed since his February memo: the technology is real, but until we know how AI will be used and who will profit, the capex and the prices can’t be judged. The resolution: calibrated humility on strategy, sizing and risk, plus a willingness to act, because “intelligent investing is largely about price.” He closes with Buffett in Forbes (August 6, 1979): “Uncertainty actually is the friend of the buyer of long-term values.”

Claude’s predictability table

Reproduced from the memo (Marks presents these as Claude’s “judgment calls”):

AI sector	Predictability (1-10)	Comparable industry	Its score
Regulated power utilities serving AI load	6	Regulated utilities	6-7
Data centers leased long-term to big tech	5-6	Large industrials	5-6
Hyperscalers	5	N/A
Chips	4	Airlines, shipping	3-4
Speculative or “neocloud” data centers	2-3	Telecom/fiber ‘98-‘01	2-3
Frontier labs (OpenAI, Anthropic, xAI, etc.)	2	Early-stage VC	1-2

Notable quotes

“Investing consists of just one thing: positioning capital to benefit from future events, and most investors cope by relying on forecasts.”

“Here’s the test I propose nowadays: 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?”

“It’s incontrovertible that disproportionate reliance on a single sector makes the economy and stock market more vulnerable than they would be if the factors underlying them were broad-based.”

“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, on frontier labs, quoted by Marks (“Ironic, isn’t it, given that Claude belongs to Anthropic?”)

“Whereas in real life things fluctuate between pretty good and not so hot, in the minds of investors they go from flawless to hopeless.”

“The key boils down to this rather complex one: will AI provide sufficient value to users to warrant sales of compute and prices for compute high enough to result in adequate returns on the trillions of dollars being invested in it (and for which companies)?”

“Two kinds of investors lose a lot of money: those who know nothing and those who know everything.” — an old saying Marks applies to 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, Forbes, August 6, 1979, which Marks uses as his close

How it connects

• Howard Marks and Oaktree — the author and the memo series; earlier memos are collected in Oaktree — Memos from Our Chairman, 1990 – July 2011 (Howard Marks) and queued as Memo to Oaktree Clients. See also The Illusion of Knowledge (Howard Marks memo), whose theme (“the biggest mistakes are made when people think they” know everything) runs through this one.
• Mastering the Market Cycle — Marks cites it directly for the rule that a forecast needs a probability attached; The Most Important Thing is the home of the “second-level thinking” he asks whether AI can do. Market Cycles and Bubbles for the “flawless to hopeless” swing.
• Dr. Tokens or: How I Learned to Stop Worrying and Love the AI Bubble — Kyle’s own essay on the AI bubble question, which already cites Marks; this memo is the 2026 update of that debate.
• AI Capex, Data Centers, Picks and Shovels — the memo’s data-center section is a skeptical read of the picks-and-shovels trade.
• Commoditization — the central “who profits” question: equally smart models selling on price.
• Forecasting — the ±50% EPS test and the 0.6^4 = 13% arithmetic are compact forecasting-humility tools.
• Concentration and Risk — the “fault line” argument; Michael Milken’s diversification thesis as the counterexample.
• Private Credit — the debt-market section (IG concentration, SPVs, lease out clauses, chip collateral) bears directly on AI infrastructure lending.
• Value Investing and Margin of Safety — Marks’s closing: low predictability can be acceptable if the price is low enough relative to intrinsic value.
• Texas Teachers’ CIO Warns AI Capex Boom Echoes Past Busts and Treasury Has an Internal Report Warning About the Dangers of an AI Bubble — other institutional voices on the same AI-capex risk.
• Anthropic, OpenAI, Nvidia, TSMC, Claude — the companies and model at the center of the memo; Claude scores frontier labs (its own maker included) at 2 of 10, the same as early-stage venture.