Amazon reports AI cost overruns, imposes spending caps

Amazon has incurred millions of dollars in unexpected costs from internal artificial intelligence projects, with one failed deployment alone running $1.8 million over budget, according to internal discussions. The overspending, which went undetected for months in some cases, has prompted the company to implement stricter spending controls.

The most significant overrun involved a project that used Anthropic's Claude Sonnet model to match author information with product listings on Amazon's e-commerce site. That project's spending reached $1.8 million, exceeding its original budget by 860 percent, and the system ultimately failed. The cost overage was not discovered for approximately five months.

This was not an isolated incident. Engineers reported during a recent internal staff meeting that other AI initiatives also ran far over budget. These included an additional $541,000 in unanticipated costs for developing financial auditing tools and a $134,000 overspend on a logistics project intended to speed up deliveries. Senior engineers described such errors as "catastrophically expensive," noting that similar mistakes in traditional software systems would have been "trivially cheap."