Gary Cohn went on CNBC's Squawk on the Street and said six words that should give every AI IPO investor pause: "The LLM world is going to be a commoditised world."
Cohn is IBM's vice chairman. He was president of Goldman Sachs for a decade. He ran the National Economic Council under Trump. He is lead independent director on the board of Apollo Global Management, the firm that arranged Anthropic's $35bn chip financing deal.
This is not a commentator watching from the outside. This is someone embedded in the institutions financing, building and deploying the AI infrastructure that the IPO wave depends on, and he is saying the core product will become a commodity.
What commoditisation means for the IPOs
SpaceX is listing this week. Anthropic and OpenAI have both filed with the SEC. Combined, the three companies are targeting public valuations that could exceed $3 trillion. A significant portion of those valuations rests on the assumption that their AI models represent durable competitive advantages.
If Cohn is right and the models themselves commoditise, the value shifts from the model to the distribution, the data, the infrastructure and the customer relationships built on top of it. The model becomes the electricity. The value accrues to the appliances.
That is IBM's position. The company has spent decades arguing that enterprise value comes from integration, consulting and workflow, not from the underlying technology platform. Cohn's view is consistent with IBM's strategic interests.
But it is also consistent with observable market dynamics. OpenAI, Anthropic, Google, Meta and a growing list of open-source providers are producing models that benchmark within single-digit percentage points of each other across most tasks. The differentiation that existed 18 months ago is narrowing with every release cycle.
Who wins in a commoditised market
If the models commoditise, the companies that win are those that control distribution, customer access and the application layer. Apple, which has chosen Gemini for Siri and will let any provider integrate through iOS 27, is building the distribution layer. Palantir, which deploys AI through human consultants embedded in customer operations, is building the application layer. Salesforce, despite its battered share price, controls the enterprise data that models need to be useful.
The companies most exposed to commoditisation are the ones whose primary product is the model itself. Anthropic and OpenAI are both diversifying into agents, enterprise tools and infrastructure, but their IPO valuations are anchored to the premise that their models are differentiated enough to command premium pricing.
Apollo connection
Cohn sits on the board of the firm that arranged Anthropic's largest financing. Apollo structured a $35bn deal to buy chips for a company whose vice chairman's board member is publicly saying the product those chips support will commoditise.
There is no contradiction. Apollo earns fees on the financing regardless of whether the models commoditise. The infrastructure retains value even if the software running on it does not. That is the picks-and-shovels thesis applied to credit markets.
But it underscores the gap between the infrastructure bet, which is durable, and the model bet, which may not be. Investors buying AI IPOs at trillion-dollar valuations need to decide which bet they are making, because the returns on each will be very different if Cohn is right.
