Amazon uncovers costly AI budget overruns in internal projects

Amazon has identified several cases of uncontrolled spending on internal artificial intelligence initiatives, including one AI budget overruns incident in which a project exceeded its planned budget by 860% and went unnoticed for five months, according to a Financial Times report.

The most expensive case involved a project built on Anthropic’s Claude Sonnet model to automate the matching of author data with product listings on Amazon’s e-commerce platform. The initiative generated $1.8 million in costs before the overrun was detected and was never deployed into production.

Engineers also reported two additional projects with significant unexpected expenses. One financial auditing tool accumulated roughly $541,000 in unforeseen costs, while an AI logistics project designed to improve delivery times resulted in losses of approximately $134,000.

Company engineers described the situation during an internal review meeting as “catastrophically expensive.” One engineer said it remained extremely difficult to determine the actual cost of many AI-related workloads. The issue highlights a broader challenge facing organizations adopting large language models. Unlike traditional software, where inefficient code often has limited financial consequences, poorly optimized generative AI applications can consume large amounts of computing resources and rapidly inflate operating costs.

Amazon said the reported incidents affected only a small number of teams and did not reflect the company’s broader use of artificial intelligence. Since then, senior engineers have introduced stronger budget controls and warned internal teams about the financial risks associated with AI development.

The disclosures come just before Amazon is scheduled to release its quarterly earnings, at a time when investors are closely monitoring the company’s spending on artificial intelligence. Amazon has committed to investing hundreds of billions of dollars in AI infrastructure, making questions about spending efficiency increasingly important.

The incidents also reflect a wider issue across the technology industry. As companies accelerate the adoption of generative AI, weak financial safeguards can allow relatively small technical mistakes, such as inefficient API calls or poorly configured prompt loops, to generate substantial cloud computing bills before conventional monitoring systems detect the problem.

The situation is particularly notable because Amazon’s cloud business offers cost management tools to enterprise customers, highlighting the complexity of controlling AI-related spending even within one of the world’s largest technology companies.