Amazon's AI spending has spiraled out of control, and the company is now trying to rein it in — FT

Amazon employees have identified several instances of “catastrophic” overspending on artificial intelligence due to errors in technology implementation and a lack of cost controls, the Financial Times reports, citing sources. The company found that under the new pricing model adopted by OpenAI and Anthropic, AI can “burn through” thousands or even millions of dollars without delivering tangible results.

At one of the meetings, senior Amazon engineers told their colleagues that they had tried to delegate tasks previously performed by programmers to artificial intelligence. This led to “unplanned” expenses. Now, engineers are developing automated control mechanisms designed to limit costs in future projects, according to sources within the company who spoke to the Financial Times. These cases illustrate the challenges even the largest tech companies face when trying to integrate artificial intelligence into their operations while avoiding budget overruns, the publication notes.

"It's hard to figure out how much anything related to AI actually costs," a high-ranking Amazon executive told the newspaper.

At the meeting, engineers described an incident in which Anthropic’s Claude Sonnet neural network spent $1.8 million cross-referencing information about authors with product listings on the Amazon e-commerce platform. The company deemed this AI implementation a failure, an FT source said: costs exceeded the project budget by 860%, and the overspending was not detected until five months later.

Amazon engineers noted that bugs in the software code—which programmers could fix “practically for free”—became “catastrophically expensive” when AI was used. The company’s lead engineers warned employees that this instance of cost overruns was not an isolated incident. In another project, the company incurred approximately $541,000 in unforeseen expenses while developing financial audit tools. Another incident occurred when AI was used to speed up deliveries in Amazon’s logistics network: unanticipated costs reached $134,000, and it took more than two weeks to identify them.

The engineers' presentation noted that the cases cited affected only a few Amazon teams. In total, the company's corporate divisions employ about 300,000 people.

“As with any new technology, we’re experimenting, learning, and refining how we use it, including ways to improve cost efficiency,” an Amazon spokesperson told the FT. “Selecting a few isolated examples where teams are learning from one another and presenting them as standard practice does not reflect how Amazon’s divisions use AI.”

The FT article on Amazon’s unchecked spending on AI was published shortly before the company released its earnings report. After the market closes in New York, the tech giant will report its second-quarter results. Amazon shares rose 5.2% to $238.5 during Thursday’s trading session. They have gained 3.3% year-to-date.

Amazon is carrying out large-scale layoffs in an effort to cut costs and finance massive investments in AI. The company’s capital expenditures are expected to reach $200 billion this year, with the vast majority of that amount going toward artificial intelligence and data center infrastructure, according to the FT.

Companies are facing rising costs for implementing AI tools as their vendors change their pricing models. In particular, the AI labs Anthropic and OpenAI have shifted some services from a fixed subscription fee to a token-based pricing model—where tokens represent units of data processed by the models.

Changes in pricing are forcing companies to abandon the most advanced models in favor of cheaper mid-range solutions and alternatives—including open-source models—where they only have to pay for the computing power they actually need.

In May, Amazon shut down an internal ranking system that ranked employees based on their activity on the Kiro developer platform, the FT reported. The system led to so-called “tokenmaxxing”: employees artificially inflated their consumption of AI tokens. Amazon executives told employees that the service was created “with good intentions” but ultimately led to increased costs rather than greater efficiency.

This article was AI-translated and verified by a human editor