Amazon's Vogels Sees Enterprises Shift to Open-Source AI

Runaway AI bills are starting to shape architecture decisions rather than just finance memos, and Amazon's CTO is now on record about where he thinks that ends. In an interview with Fortune, Werner Vogels said "we see a shift happening between the cheaper open source models and the bigger expensive models," and framed cost as "a very important part of your architecture" rather than a downstream problem finance sorts out later.

The two anecdotes doing the rounds are the reason CFOs are paying attention. Uber reportedly burned through its entire 2026 AI budget in four months after incentivizing employees with an internal leaderboard ranking teams by total AI tool usage. A separate, unnamed company is said to have burned through $500 million in a single month before finally restricting employee AI usage. Vogels's takeaway, per Fortune, is that companies should question whether they really need the biggest, highest-end model to solve their problems. The proprietary providers he is implicitly contrasting against are OpenAI, Anthropic and Google DeepMind, which bill by the token.

The reason this matters beyond one CTO's talking points is that billing by the token changes what an architecture decision even means. In classical cloud you sized instances, over-provisioned a little, and paid a predictable monthly bill. In token-metered AI the unit cost sits inside every user interaction, so a well-meant internal push to adopt AI tooling can, as the Uber example suggests, blow through a year of budget while everyone is congratulating themselves on adoption. Open-weight models on your own inference infrastructure convert that variable cost back into something closer to the old capacity-planning problem, which is a very different conversation with finance.

The honest caveat is that Fortune's piece is thin on specifics. The $500M-a-month company is not named, we do not know which workloads or which models Uber was running through when it hit its cap, and Vogels has an obvious institutional interest in a world where enterprises host their own open-weight models on AWS. Take the direction as reported, and the exact numbers as reported rather than settled.

For enterprise buyers the useful move this quarter is less "switch everything to open source" and more instrument the token bill by team and by task, and know your open-weight fallback price for each workload before someone else asks. That is the conversation Vogels is nudging toward, and it is a healthier default than trusting that frontier pricing curves will keep falling faster than usage grows.