Editorial portrait illustration of Mark Zuckerberg for Jeremy Stern's profile on Meta's AI strategy

Jeremy Stern’s profile of Mark Zuckerberg in Colossus explains why Meta’s AI strategy could work even without having the best model. I agree with much of his argument. It also helps separate two questions that are often conflated: how useful AI becomes and how much pricing power model developers retain.

I’ve written about this in AI Models Are the New Rebar and AI Models as Standalone P&Ls. A model can remain useful after competitors reproduce its capabilities. The premium customers will pay for access can then approach zero. I’m increasingly convinced that this describes the economics of much of the model market.

Why Meta does not need to sell access to its AI models

Stern describes two objections to Zuckerberg’s spending. If AI does not commoditize, Meta will remain behind the leading labs. If it does, Meta will have spent a fortune developing something competitors can also supply. He summarizes the critics’ position:

Heads, his rivals win; tails, he loses.

The second objection assumes that Meta needs to earn its return by selling access to models. Its advertising business gives it another way to recover the investment. A fall in model prices can weaken a model provider’s business while improving the economics of companies using those models. We should assess Meta’s spending against the revenue and savings it produces across the company.

Commoditization does not send OpenAI and Anthropic to zero

Stern asks competing researchers and investors to consider how Zuckerberg might succeed. He writes:

If AI commoditizes, then Anthropic and OpenAI go to zero, and value accrues instead at the complements Meta already dominates, like distribution, attention, personalization, and commerce. If it doesn’t commoditize, then at least he is not his competitors’ prisoner the way he’s been with Apple, and all he has to do is remain within six months of the frontier, which he’s already close to. Heads, he wins; tails, he wins.

“Anthropic and OpenAI go to zero” goes further than the argument supports. Both can build products that customers prefer even when the underlying models become interchangeable. Reliability and integration can justify payment independently of a model’s capability lead. Competition at the model level does not settle the value of the businesses built around it.

Meta’s AI strategy earns its return through advertising and independence

Meta already has businesses through which it can earn a return. Better ad targeting can generate revenue without charging users for AI. Improvements to its apps can also justify some of the expense. Stern puts the potential user base at 3.6 billion people. For many of their tasks, a model behind the frontier may be sufficient.

Owning a competitive model also reduces Meta’s dependence on another company’s permission to build products. Stern’s comparison with Apple explains why Zuckerberg might pay for that independence even without expecting model sales to cover the cost.

Stern makes a persuasive case that Meta’s AI strategy can benefit from commoditization. Whether those benefits justify its spending remains a separate question. Having a weaker model is insufficient evidence that the strategy has failed, just as having billions of users is insufficient evidence that it will pay off.