Amazon employees spent $1.8 million on AI. The company wants to stop them
AI is “catastrophically expensive” for Amazon.
- Amazon’s AI projects reportedly exceeded budgets, including a failed Claude project that cost $1.8 million.
- The Claude project ran 860% over budget, and Amazon did not detect the overspending for five months.
- Amazon said it is learning to use AI more efficiently and developing guardrails to prevent future cost overruns.
- The incidents show how token-based AI pricing can make software costs harder for companies to predict and control.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
It is difficult to “figure out how much” anything related to AI really costs, but Amazon employees are spending millions of dollars on tokens on failed projects.
Amazon is spending a fortune on AI. According to the Financial Times, several AI projects on Amazon have dramatically exceeded their budgets.
Amazon staff described several incidents in which coding mistakes and insufficient spending controls led to what employees call "catastrophically expensive" situations.
Reportedly, senior engineers discussed during an internal staff meeting this week, warning that projects migrated from conventional software development to AI workflows had generated unplanned expenses.
The company is now developing automated guardrails intended to prevent similar overspending in future AI deployments.
One senior Amazon employee told the Financial Times that it is difficult to “figure out how much” anything related to AI really costs.
A failed AI project cost Amazon $1.8 million
The most expensive AI project cost Amazon $1.8 million. Employees used Anthropic's Claude Sonnet model to match author information to product listings on Amazon's e-commerce platform.
The deployment ultimately failed, but before it did, it generated approximately $1.8 million in costs. This was 860% over the budget. Most surprisingly, the spending went unnoticed for 5 months.
During an internal meeting, engineers reportedly warned that mistakes that would have been "trivially cheap" in traditional software systems can become "catastrophically expensive" when AI models are involved. The main reason is that every request to a large language model consumes billable tokens.
Other AI projects also exceeded budgets
The described incident was not an isolated case. One project building financial auditing tools reportedly accumulated roughly $541,000 in additional costs.
Another initiative to improve delivery speeds across Amazon's logistics network generated approximately $134,000 in unintended spending before the problem was detected. It took Amazon 2 weeks to notice overspending.
Amazon: “We're learning how to use AI efficiently”
Amazon acknowledged the incidents but emphasized that adopting new technology inevitably comes with lessons learned.
“As with any new technology, we’re experimenting, learning, and improving how we use it, including how we drive cost efficiencies,” it told the Financial Times.
The company also pushed back against the idea that the examples represent widespread problems.
"Cherry-picking small, isolated examples where teams are learning from one another and portraying them as business as usual doesn't reflect how teams across Amazon are using AI."
Amazon has already encountered problems related to AI consumption incentives. Earlier this year, the company reportedly shut down an internal leaderboard that ranked employees based on usage of its Kiro developer platform after it encouraged so-called "tokenmaxxing," with employees inflating AI usage to climb the rankings.
The overruns are relatively minor compared with Amazon's overall business. The company generates roughly $180 billion in quarterly revenue and is expected to spend around $200 billion in capital expenditures this year, much of it on AI infrastructure and data centers.
Token-based pricing is a lottery
Major AI providers, including Anthropic and OpenAI, have increasingly shifted enterprise customers away from fixed subscription models toward token-based pricing, where organizations pay based on how much data AI models process.
While the approach allows flexible scaling, it also makes costs far less predictable, especially when autonomous AI agents repeatedly call large language models.
To reduce expenses, many companies have begun replacing premium AI models with cheaper alternatives or open-weight models that can be run on their own infrastructure.
AI has been very expensive for the tech companies. Many have fired workers to cut costs, but now are paying 5x more for AI tokens than for salaries. Uber maxed out its full-year AI budget in 4 months.
Nvidia’s VP also admits that compute costs more than workers.
“For my team, the cost of compute is far beyond the costs of the employees,” Bryan Catanzaro, vice president of applied deep learning at Nvidia, told Axios.
