A $1.8M Claude Task: Amazon Learns the Price of Runaway AI Costs

Amazon just learned an expensive lesson about runaway AI costs. According to Amazon employees, the company recently tried to use Claude Sonnet to fill in detailed author information for its own website — a seemingly simple task that ended up costing $1.8 million, overshooting the budget by 860%.

The overrun was only discovered five months later, and the deployment never succeeded. At Claude Sonnet's public pricing of roughly $3 per million input tokens and $15 per million output tokens, $1.8 million could have burned through as many as 600 billion tokens — about twice the size of the entire GPT-3 training corpus.

QbitAI reported the incident based on employee accounts, noting that similar cases have occurred repeatedly inside Amazon and that such bugs tend to surface only after long delays. The report also highlights a broader problem: in traditional systems, minor issues cost little, but with AI they can generate staggering bills — a survey found only 26% of companies have full visibility into their AI costs.

The accidents have not slowed Amazon's automation push. CEO Andy Jassy has announced roughly $220 billion in capital spending for 2026, most of it going to AWS, custom AI chips, and power infrastructure — up nearly 60% from 2025. In its second-quarter results published this morning, AWS posted net sales of $42.2 billion, up 37% year over year, and contributed about 60% of Amazon's $27.5 billion operating profit while generating only 21% of total revenue.

Jassy has written to employees that Amazon now runs more than 1,000 generative AI services and will one day run billions of agents. The automation wave has also reached its warehouses: reports say Amazon's robotics unit aims to automate about 75% of warehouse operations by around 2033, which would mean roughly 160,000 fewer U.S. jobs by 2027 and more than 600,000 fewer by 2033. Nobel laureate Daron Acemoglu has warned that if Amazon's automation dream comes true, America's largest employer would turn from a net job creator into a "net job destroyer."

Amazon is not alone. In April, a Meta employee built a leaderboard called Claudeonomics aggregating AI usage data from more than 85,000 employees; token consumption climbed to 73.7 trillion in 30 days, equivalent to a bill of roughly $221 million per month at public pricing. In June, Meta sent memos to about 6,000 employees announcing token limits and building an "AI Gateway" platform to monitor usage and spending in real time. Uber, meanwhile, burned through its entire annual AI coding budget in the first four months of 2026, then capped spending at $1,500 per employee per tool each month.

The model providers feel the loss of control too. On June 3, Sam Altman said AI costs had gone from being completely ignored early this year to becoming a big problem, revealing that OpenAI's heaviest internal user consumes about 100 billion tokens a month and that one employee once burned through roughly 210 billion tokens in a single week.

In less than a year, Silicon Valley has moved from the excitement of discovery to budgets, caps, approvals, and dashboards. The 2012 Knight Capital disaster — a 45-minute automated trading glitch that lost $440 million and ended in the firm's acquisition — is a reminder that automation amplifies losses as efficiently as it delivers gains. For Amazon, the $1.8 million tuition may not be too steep.