Early in June, Uber announced that its staff had burned through its entire 2026 AI budget in just four months. The phenomenon was given a name: “tokenmaxxing”, the practice of spending on frontier models with no understanding what value was realized.
Uber’s fix was similar to many other companies’: Impose a monthly token cap for each employee, with a manager’s sign-off required to get more.
Capping employee token use positions AI as a cost to be minimized rather than a tool for optimizing productivity. It punishes innovation and trains employees to avoid AI tools. It’s a bit like firms rationing printer paper or email storage. The only people who are happy are the accountants in Finance.
If cost is a legitimate concern, a more nuanced approach would be to identify runaway automated jobs, or look closely at teams that are heavy users and understand what they are doing with AI, rather than bluntly throttling individuals.
In parallel to tokenmaxxing, the frontier models started cutting prices.
The Chinese labs started first. In May DeepSeek confirmed that a 75% discount on its model would become permanent. Xiaomi cut prices by up to 99% later that month.
Cuts to the US models followed in late July. OpenAI cut prices by up to 80%. Google followed in August with cuts of up to 50%. Anthropic followed with cuts of up to 90%.
The race to the bottom on pricing is already underway.
Cheaper tokens won’t end tokenmaxxing; they’ll accelerate it. Every price cut allows companies to use more tokens and therein explore how AI can help their business grow.
Smart companies will take the time to understand token value by tying spend to specific business outcomes. Counting tokens tells accountants what AI costs, whereas valuing business outcomes tells companies if the spend was worthwhile.
Back to the frontier model companies: How will they make money if they are engaged in a price cutting war? It’s a volume game. Each token costs less, but usage continues to grow, so total spend keeps climbing. The successful frontier companies will help both their customers and themselves by making it easy for firms to tie token usage to business outcomes.
If history is any guide, the battle between frontier models will move from price to volume, then to the deepest and broadest coverage, and finally to ease of use. By then the price may be zero, but there will be a hidden cost for both corporate and personal customers.
Search engines followed the same path in the mid-1990s. AltaVista, Excite and Lycos fought a war over who could index the most of the web. Google won that war by being smarter and simpler: It served up a simple web page that loaded instantly with no screen clutter. Search became free for everyone, but the hidden cost was our personal information, used to serve up targeted ads. Advertising now brings in about 73% of Alphabet’s revenue.
When the cost of AI tokens hits zero, who is really paying, and what information are we willing to give up for free AI capabilities? Search engines learned what products we wanted in the material world. AI will learn how we are thinking inside our heads, and be there to influence us as we reason, decide and then take physical actions in the tangible world.
The social, economic and political implications of an AI model knowing how to nudge each individual’s internal thought processes, at exactly the right moment, to prompt specific action the physical world are enormous.