Uber holds AI spend flat while usage rises 9.4 times

The piece frames Uber's result against a broader problem: AI bills scale with usage in a way traditional software budgets never did, so growth itself becomes a cost line. It notes that cheaper inference has so far mostly led to more AI usage rather than lower bills, which is why discipline in routing matters more than price alone.

It also warns that teams often reach for a bigger model to fix what is actually a data problem. Fixing weak retrieval, a stale knowledge base and bad context is the cheaper route, while throwing compute at those issues is described as the most expensive possible fix that only half works.

On the technology side, the piece argues small models have improved enough to make tiering practical: an 8B model today can handle work a 70B model struggled with a few years ago. It also advises setting an escalation path, since a router occasionally sending a request to the wrong tier is acceptable only if the error gets caught.