The enterprise AI spending honeymoon may be ending. Uber's COO has gone on record saying it's becoming increasingly difficult to justify the cost of what's being called "tokenmaxxing" — and if that term is new to you, it's worth understanding, because it's costing big companies serious money.
Tokens are the basic units AI language models process — roughly three-quarters of a word each. A 128,000-token context window can hold about 300 pages of text. Tokenmaxxing is the practice of deliberately filling that window to the brim: feeding a model your entire customer database, full email threads, complete codebases, whatever you have, on the theory that more context produces better outputs. It does, sometimes. But AI providers charge per token processed, so running 100,000-token requests at scale means costs compound fast.
For a company like Uber — which operates across ride-sharing, delivery, and freight in dozens of markets — even a modest token-heavy deployment touching millions of interactions daily adds up to a significant line item. The COO's comments suggest internal finance teams are starting to ask hard questions that AI vendors would rather not answer: what exactly did we get for that spend, and could we have gotten 80% of the result at 20% of the cost?
The Real Problem Is Measurement
The ROI question is legitimately hard. When AI quietly reduces customer service ticket resolution time, or helps a developer ship a feature faster, the dollar value doesn't land on a single spreadsheet line. That ambiguity has let AI spending grow unchecked at many organizations — but it also means when a CFO or COO decides to look harder, there's no clear defense ready.
Tokenmaxxing in particular is a ripe target for scrutiny because it's easy to identify and quantify. You can pull the API invoices, count the tokens, and ask whether the incremental quality gain from a 100k-token prompt over a 20k-token prompt justified the five-times cost difference. For many use cases, the honest answer is no.
This isn't an argument against AI. It's an argument for precision over brute force. The companies that will get the best long-term value from AI aren't the ones throwing the most tokens at every problem — they're the ones figuring out which tasks genuinely need a large context window and which ones don't. Uber's COO saying this publicly is a signal that the "just use more AI" phase of enterprise adoption is giving way to something more disciplined.