Howard Marks: AI Is the Least Predictable Trade of a Generation

Howard Marks, co-founder of Oaktree Capital, has published a sweeping memo arguing that artificial intelligence is "the most consequential and least predictable development for investors in at least a generation" — and he enlisted Anthropic's Claude to help quantify just how unpredictable it is.

Claude's verdict, quoted verbatim in the memo: "My overall judgment: AI as a whole scores about a 3" on a 1-to-10 predictability scale. "That number hides a split. The direction is fairly predictable, maybe a six or seven: the technology will keep improving and spread through the economy. The things investors get paid for are much less predictable, around a 2: how fast it happens, how big the profits are, and who keeps them."

The memo, framed as a layman's deliberations rather than a technical analysis, arrives amid fresh market anxiety about AI valuations, leverage and the pace of compute demand. Its stakes are broad: Marks argues AI now constitutes an unusually powerful "fault line" running through the entire economy and capital markets, not just one sector.

Why does the probability math matter?

Marks opens with a compounding problem. Top-down investors must be right about the economy, the industry, the company and the market. At a 60% probability of getting each right — "which is probably generous" — the combined probability is 13%, assuming independent outcomes.

AI multiplies the variables. An investor must be at least somewhat right about the technology's evolution, compute demand, competitive structure, who benefits, financing availability and the timing of all of the above. "That's just something to think about as you read," he writes.

His grounding principles: investors can never know everything, the biggest mistakes come from people who think they do, and incomplete knowledge must not paralyze action. "It's the investor's job to bear risk in the pursuit of profit," Marks writes. "Given its vast potential, we can't decline to participate in AI just because the future is unclear."

What did Claude say about how AI works?

Marks asked Claude how AI does what it does, and quotes its answer at length: "We know what AI is made of and how it was built almost perfectly, but we know surprisingly little about how it actually does what it does." Claude described the systems as "grown rather than designed," with rules emerging "from the nudging, spread across billions of numbers that mean nothing individually."

On interpretability, Claude offered an analogy: "It's like having a complete record of every trade in the market and still not knowing why prices moved." It also flagged that a model's self-explanations may not match its internal computation. "The honest summary is that we've built a mind-like thing we can inspect completely and still can't fully explain," Claude said — a passage Marks notes with the aside that "here Claude talks like it's a person."

Who will profit — vendors or users?

Marks invokes Warren Buffett's tech-bubble observation that transformative technologies do not necessarily reward investors. If multiple AI models become equally smart, services may commoditize and sell on price, grinding down returns and shifting economic benefits from vendors to users — with implications for "the trillions of dollars being invested in AI infrastructure." Alternatively, some AI companies may become essential monopolies; he cites Nvidia (NVDA) and TSMC (TSM) as evidence of the value of strong competitive positions.

He also raises taxation: since AI will likely replace workers, and governments must replace lost payroll tax revenue, "what's to say AI won't be taxed like people? Many of today's top earners are taxed at combined rates of 50% or more. What tax bracket will AI be in?"

On competition, Marks notes that OpenAI and Anthropic were unknown to most people a few years ago, and asks whether today's leaders can likewise be displaced — by the hyperscalers Amazon (AMZN), Microsoft (MSFT), Alphabet (GOOGL), Meta (META) and Oracle (ORCL), or by companies that do not yet exist. Napster, Yahoo and Netscape did not survive the last technology transition. His proposed test for any AI bull: forecast a company's earnings per share five years out, and ask whether you would bet the actual result lands within plus-or-minus 50% of your number.

How exposed are the economy and markets?

The memo's most data-dense section catalogs concentration risks:

• The "Magnificent 7" out-earned the rest of the S&P 500 for 14 consecutive quarters through June 30.
• Per FactSet, Alphabet and Amazon booked a combined $151 billion of second-quarter gains on equity investments, mostly in Anthropic and SpaceX (which now includes xAI).
• Per J.P. Morgan, 42 AI-related companies generated roughly two-thirds of the S&P 500's price gains since January 2024.
• AI's contribution to GDP growth comes overwhelmingly through capital expenditures on data centers, which are rising while the rest of construction declines.

Marks draws a parallel to Michael Milken's sub-investment-grade debt revolution of 1977-78, which rested on incremental yield and diversification. Today's markets, he argues, rest heavily on a single factor: the success of AI.

What about the credit market?

Goldman Sachs data cited in the memo shows nearly $600 billion of AI-related debt issued globally so far this year. About $400 billion landed in the U.S. investment grade market — over one-fifth of total IG supply — lifting AI-related bonds to roughly 12% of the Bloomberg USD IG Corporate index and about 6% of the high yield market. Going forward, Marks estimates AI will require an average of $1 trillion of new capital per year, with hundreds of billions raised annually through debt.

He warns that hyperscalers, as "takers" of financing, could crowd out other borrowers and push up government borrowing costs. He flags risks in data-center leases with "out clauses" (such as a 99.9% reliability threshold), chip collateral that may not outlast the debt, off-balance-sheet special-purpose vehicles, and "circular transactions" in which money moves from A to B and back to A. Equity markets may also strain: traditional U.S. IPOs raised below $50 billion in eight of the last ten years, versus nearly $80 billion in 2020 and $140 billion in 2021.

How predictable is AI, and is it a bubble?

Claude's subsector scores, which Marks calls "judgment calls," range from frontier labs at 2 — "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" — to securely leased data centers and power providers, comparable to large industrials and regulated utilities. Marks finds it ironic that Claude, owned by Anthropic, rated frontier labs nearly unknowable, and notes the irony deepens: those least-predictable assets trade at a premium privately, and both OpenAI and Anthropic have filed for IPOs.

On the bubble question, Marks defines a bubble as "irrational exuberance leading to imprudent behavior, excessive investment, and too-high prices for participation." He holds to his February position: the technology is real and transformative, but whether spending, prices and debt yields are justified cannot yet be assessed. Recent worries — cheaper Chinese models, open-source competition, declining token prices, slowing compute growth, high leverage — "are nothing new, but a changed investment climate has rendered them sufficient to darken the mood."

His conclusion, after consulting knowledgeable people: answering whether AI will generate returns sufficient to justify trillions in investment "is largely a matter of faith." He quotes economist Stuart Chase: "For those who believe, no proof is necessary. For those who don't believe, no proof is possible."

The prescription is calibrated humility on position sizing and risk, not paralysis — because, as Buffett wrote in Forbes on August 6, 1979, and Marks gives the last word: "Uncertainty actually is the friend of the buyer of long-term values."