An Amazon project that used Claude Sonnet to fill in detailed author information for the company’s website reportedly cost $1.8 million, exceeded its budget by 860%, went unnoticed for five months, and never made it into production.
According to Amazon employees cited in the source article, the assignment looked simple on paper. The company wanted Claude Sonnet to populate richer author profiles on its site. Instead, the effort turned into an expensive internal failure.
At Claude Sonnet’s public pricing, listed in the source as about $3 per million input tokens and $15 per million output tokens, a $1.8 million bill could imply as many as 600 billion tokens consumed. The article says that, by data volume alone, that would be about twice the size of GPT-3’s full training corpus.
The report says similar incidents have appeared inside Amazon more than once, and that bugs tied to AI systems are often discovered only after long delays. Employees quoted in the piece said costs around AI work are difficult to track clearly, and that small issues that would be negligible in conventional software can become unexpectedly expensive when AI is involved.
Amazon is still accelerating its AI and automation push
The overspend has not slowed the company’s broader investment plans.
Chief Executive Officer Andy Jassy said Amazon expects roughly $220 billion in capital expenditures in 2026, with most of that spending directed toward Amazon Web Services, internally developed AI chips, and power infrastructure. The article describes that figure as nearly 60% above 2025 levels and the largest single-year capital spending plan among the world’s mega-cap companies.
Amazon’s results for the second quarter ended June 30, 2026, gave that spending push a strong financial backdrop. AWS posted net sales of $42.2 billion for the quarter, up 37% from a year earlier. Out of Amazon’s total operating profit of $27.5 billion, AWS contributed about 60%, even though the cloud unit accounted for only 21% of company-wide revenue.
At the same time, Amazon has carried out two rounds of large-scale layoffs since October 2025, cutting about 30,000 jobs in total.
Jassy had already laid out the company’s ambitions in a memo to employees last year. “We now have more than 1,000 Generative AI services in progress or built, but, at our scale, that’s a small fraction of what we will ultimately build,” he wrote. “There will be billions of AI agents, across every company and in every imaginable field. There will also be agents that routinely do things for you outside of work, from shopping to travel to daily chores and tasks. Many of these agents have yet to be built, but make no mistake, they’re coming, and coming fast.”
Automation goals extend beyond office software
Amazon’s automation plans are not limited to internal digital workflows.
According to the report, the company’s robotics division is aiming to automate about 75% of warehouse operations around 2033. The article says that would mean Amazon hiring about 160,000 fewer workers in the United States by 2027 and more than 600,000 fewer by 2033.
Daron Acemoglu, the 2024 Nobel economics laureate and a professor at the Massachusetts Institute of Technology, gave a harsh assessment of that vision. The source quotes him as warning: “If Amazon’s automation dream materializes, the nation’s largest employer will shift from being a net job creator to a net job destroyer.”
Token spending blowouts are showing up across big tech
The Amazon case is presented as part of a larger pattern that began showing up across major technology companies earlier this year.
In the source article’s telling, many of these episodes start the same way: management urges teams to use as much AI as possible, build with the smallest teams possible, and extract the deepest value possible. Employees then build internal leaderboards around usage, turning token consumption into a contest. Once the bills arrive, executives shut the contests down.
The article ties that pattern to Goodhart’s law, proposed in 1975 by economist Charles Goodhart: when a measure becomes a target, it stops being a good measure.
Amazon itself once had an informal internal ranking called KiroRank, which encouraged some staff to inflate their own numbers intentionally. That ranking was later scrapped. In its place, the company introduced a metric called “normalized deployments” to measure AI-related output from employees.
In April, a Meta employee created a leaderboard called Claudeonomics that aggregated AI usage data from more than 85,000 employees and listed the top 250 token consumers. The source says Meta employees burned through 73.7 trillion tokens over a 30-day period.
Using what the article calls the public pricing of “Company A,” that amount would translate into a monthly bill of about $221 million.
In June, Meta sent a formal memo to about 6,000 employees saying token usage would be capped. The company also built a central platform called AI Gateway to monitor usage and spending across teams in real time, while setting budgets and upper limits.
Uber ran into similar issues. The report says the company exhausted its full-year AI coding budget in the first four months of 2026, then imposed a monthly spending cap of $1,500 per employee and per tool. Its COO said a measurable connection between token spending and output had not yet taken shape.
Model providers are dealing with the same pressure
Cost pressure is no longer only a customer problem.
On June 3, Sam Altman said AI cost had barely been on anyone’s radar at the start of the year, but had now become a major issue. He also disclosed that the heaviest individual user inside OpenAI consumes about 100 billion tokens a month, and that one employee at one point burned around 210 billion tokens in a single week.
The article cites a survey showing that only 26% of enterprises have full visibility into their AI costs. In practical terms, many companies still do not know how much they spent before trying to cut back.
Automation can scale mistakes as fast as it scales efficiency
The report closes with a reminder from an earlier automation failure outside the AI industry.
On Aug. 1, 2012, Knight Capital Group, one of the largest U.S. equity market makers, updated its automated trading system to participate in the New York Stock Exchange’s Retail Liquidity Program. During deployment, engineers accidentally activated an old piece of code. The system then began firing off high-frequency trades with no economic logic, buying high and selling low without any preset stop-loss mechanism or spending ceiling.
The malfunction lasted about 45 minutes from the market open until the system was manually shut down. During that stretch, it executed more than 4 million trades across 154 stocks and accumulated roughly $7 billion in stock positions the company had never intended to hold.
Once engineers found the fault and shut the system off, Knight Capital had to dump those positions at lower prices. The company’s loss came to about $440 million, roughly three times its annual profit. Over the following two trading days, Knight Capital shares lost 75% of their value. A few months later, rival GETCO acquired the firm.
Automation is usually sold on speed, lower costs, and fewer human errors. The same systems can magnify failure at the same speed and scale when something breaks. Amazon’s Claude incident, as described in the source article, is one more example of that problem.