OpenAI Board Chair Makes a Bold Prediction: In a Year, No One Will Worry About Token Costs

Token costs are eating into corporate AI budgets — but OpenAI’s board chair says that anxiety has a shelf life, and it expires in about 12 months.

Bret Taylor, who also co-founded AI startup Sierra, told CNBC on Sunday that the entire conversation around token pricing and management will soon feel as dated as worrying about minutes on a cell phone plan. The reason is straightforward: the market is still in its awkward early phase.

“Right now, the way we bill for AI is tied to a technical unit that nobody outside engineering should have to think about,” Taylor said in the interview. “Twelve months from now, IT departments will be very good at deploying different AI tools for different scenarios — marketing uses one product, software engineers use another. Companies simply won’t need to think about tokens anymore.”

Token costs spiked earlier this year, pushing some companies to scale back AI spending and reexamine whether the investments were delivering measurable returns. CFOs in particular have been asking hard questions about ROI, and the volatility of usage-based AI pricing has made budget planning a headache.

Taylor sees a pattern that’s played out before. Building a website in the early days of the internet cost far more than it did once the market matured. The same trajectory, he argues, applies to AI infrastructure. Entrepreneurs should build solutions that let businesses offload token management entirely, rather than forcing every company to become an expert in model pricing.

Some startups are already moving in that direction. Ramp recently launched a tool for tracking token expenditure, and legal AI company Harvey handles token management on behalf of its clients. These are early signals of a broader shift: instead of billing by the token, providers will charge for outcomes.

The math also works in the direction of cheaper tokens. As model inference efficiency improves — through better architectures, quantization, and purpose-built silicon — the cost per token is expected to keep falling. That natural decline, combined with abstraction layers that hide the complexity from end users, makes Taylor’s timeline plausible.

Taylor’s larger point is that the current state of AI is not the steady state. The token is a temporary unit of accounting that reflects an immature market — not a permanent feature of how businesses buy AI.