Amazon Engineers Flag $1.8M Claude Bill, 860% Over Budget

The most quietly damning line in the Financial Times' report on Amazon's internal AI cost meeting is not the eye-catching dollar figure. It is the admission from Amazon's own engineers that "it's very difficult to understand how much anything related to artificial intelligence costs." That is coming from the company that runs the cloud everyone else buys AI on.

The specifics they walked through were sobering. One project to automate the mapping of author data to listings on an e-commerce platform, built on Anthropic's Claude Sonnet, ran to $1.8 million before anyone caught it, an 860 per cent overrun on the original budget, and it took management roughly five months to spot the hole. The project never shipped. Other examples on the same slide included about $541,000 on financial audit tool development, and $134,000 lost on a logistics AI project. Engineers reportedly described the pattern as "catastrophically expensive."

This lands on top of the earlier story the FT also broke about Amazon shutting down "KiroRank," the internal leaderboard for its Kiro agentic development platform, after staff started assigning agents to pointless tasks to climb the standings. Senior vice president Dave Treadwell told employees, "Please don't use AI just for the sake of using AI." Amazon's response to the cost-overrun story was that these cases occurred in only a few teams and did not reflect broader organisational usage.

The honest caveat is that this is one FT report drawing from what sounds like a single internal meeting, and the company disputes how representative it is. What the reporting doesn't give you is a base rate, how many projects came in on budget for every one that blew up, or what the total wasted spend looks like as a share of Amazon's roughly $200 billion 2026 AI capital expenditure commitment. Take the specifics as reported, not as a settled accounting.

The forward-looking thought is that "AI FinOps" is going to actually mean something in 2026. If Amazon can lose track of a $1.8 million bill for five months, the enterprises buying agentic tools from Amazon, Anthropic, and OpenAI should assume they can too, and the vendors who ship real per-project spend controls and hard budget caps by default are the ones who benefit.