Howard Marks asks Claude to grade AI predictability, and frontier labs score 2 out of 10

Howard Marks has a new memo out, and this time he brought a writing partner that runs on GPUs.

In “A Layman Thinks About AI,” the Oaktree Capital Management co-founder turned to Anthropic’s Claude for answers. He then asked it to score how predictable each slice of the AI trade is, on a scale of 1 to 10. The frontier labs, Anthropic included, got a 2.

The chatbot, in other words, rated its own maker as close to unknowable.

The scorecard

Marks is candid about his starting point. He says he has “absolutely no knowledge” of how AI works. So he outsourced the technical explanations to Claude and kept the investment judgment for himself.

The result is a predictability map of the whole AI stack. According to the memo, Claude put AI’s overall predictability at 3 out of 10.

The big-picture direction of the technology did better. That landed between 6 and 7 out of 10.

Speed of adoption, profitability and which companies end up winning all scored a 2 out of 10.

Here is how the rest of the stack shook out, per the memo:

Frontier AI labs, including OpenAI and Anthropic: 2 out of 10.

Chips: 4 out of 10.

Hyperscalers: 5 out of 10.

Data centers on long-term leases: between 5 and 6 out of 10.

The memo points to why the labs rank last. They emerged quickly, they need enormous amounts of capital, and their competitive advantages are not yet clear.

Money moving faster than certainty

He highlights projections of trillions of dollars in AI infrastructure spending. He also cites Goldman Sachs data showing that nearly $600 billion in AI-related debt has been issued globally in 2026.

Debt is the sharper part of the warning. Equity investors can absorb a bad guess and move on. Lenders who underwrite a 2-out-of-10 business are making a bet with a fixed upside and a less fixed downside.

Marks doesn’t stop at the cautionary note, though. He recalls a 1979 Warren Buffett observation that uncertainty can actually work in favor of the long-term investor.

How Marks got here

This isn’t Marks’ first pass at the subject. The memo references an AI-focused memo he published in December 2025. He also had earlier exchanges with Claude in early 2026.

What this means for investors

Leased data centers and hyperscalers sit in the middle of Claude’s scale. Their revenues are tied to contracts and existing businesses. Chips sit a notch lower, reflecting questions about how durable demand and pricing will be.

Frontier labs sit at the bottom. Their valuations depend on outcomes that, by this scorecard, are close to guesswork: who wins, how fast and whether the economics ever work.

The debt angle is the one to watch. Nearly $600 billion in AI-related borrowing this year, per Goldman Sachs, means a meaningful share of the buildout is financed by creditors rather than shareholders.