Howard Marks: AI Predictability Scores Only 2 Out of 10, Sounding a Warning on the Eve of Anthropic's IPO
nashnova research
Oaktree Capital co-founder Howard Marks released a memo just before Anthropic's IPO pricing, using Anthropic's own Claude to score AI predictability — the result was 2 out of 10 — questioning what underpins investor confidence in a trillion-dollar-plus listing.
What happens when you ask AI to grade itself?
Marks asked Claude to rate predictability across AI sub-sectors. Claude gave frontier AI labs — including Anthropic itself — a score of 2 out of 10.
This means → even AI considers its own industry "almost impossible to call." Marks uses this score to challenge every valuation model built on certainty assumptions.
Marks openly admits he has "no understanding of how AI works," but argues that is precisely the point: Claude describes its own mechanism as "no one wrote rules for its answers — they emerged from billions of individually meaningless numbers."
In plain terms = you have the complete record of every trade in the market, yet still don't know why the price moved. That is roughly AI's level of self-understanding.
Six variables stacked — what are the real odds?
Marks lists six core uncertainties in AI investing: the technology itself, compute demand, competitive landscape, who benefits, financing availability, and the timing of all of the above.
He runs a simple probability exercise: assume each variable is independent with a 60% win rate. Stack all six, and the overall probability drops to roughly 5%.
This means → even if you are "moderately bullish" on every single variable, the compounded odds of getting the whole picture right remain very low — uncertainty multiplies, it does not add.
Once models converge, who captures the profit?
Marks invokes Buffett's classic judgment on the internet: transformative technology does not necessarily reward investors.
His logic chain: if multiple models converge in capability → AI services become undifferentiated commodities → competition shifts to price → value flows to users, not suppliers.
The price war is already underway: pricing for a given performance tier is falling roughly 13× per year — far faster than Moore's Law.
This reflects a deeper signal — Marks notes that OpenAI and Anthropic were unknown to most investors just a few years ago, casting implicit doubt on the durability of current leaders.
Is the "sell picks and shovels" trade really safe?
Marks is equally cautious about the strategy of investing in data-center infrastructure to sidestep AI stock-picking risk.
The primary cost of a data center lies in chips, not buildings. Add site selection, fiber, power, cooling, and financing, and the investment is "far from easy."
He also flags that several data-center construction bans passed last week — this reflects a split: existing capacity that is built, leased, and powered may grow more valuable, while investors holding planning approvals and land options face a very different situation.
What risk is almost no one modeling?
Marks raises a variable that sell-side analysts have almost entirely ignored: taxation.
The logic is direct: if AI's primary effect is displacing labor → governments need to replace the payroll taxes those workers used to pay → AI itself may become a new tax base.
In plain terms = "What tax bracket will AI be in?" No one is answering that question today, but it directly affects long-term profit margins for AI companies.
Why does the timing of this memo matter?
Anthropic is in the middle of its roadshow, with pricing set for November 18 and trading starting November 19. The IPO is considered one of the largest capital raises in history.
The core proposition of Marks's memo: even AI cannot give AI's own predictability a high score — so what exactly is investor pricing confidence built on?
This means → this is not a "bearish on AI" report. It is a rare public text, released at the market's most euphoric moment, asking investors to honestly confront the boundaries of what they actually know.
市场有风险，内容仅供研究参考，不构成投资建议。