Amazon just burned $1.8M on an AI task that never shipped. Nobody noticed for 5 months.

Amazon's own engineers called it "catastrophically expensive."

A project using Claude Sonnet to match author details with product listings ran 860% over budget.

$1.8 million gone. The deployment failed anyway. And it took five months for anyone to notice.

Two more projects blew up the same way.

A financial auditing tool: $541K over budget. A logistics system: $134K over. Total unplanned AI spend: $2.5 million.

Here's the part that should keep you up at night.

Amazon sells the infrastructure that runs these models. AWS has Bedrock, batch inference at half price, prompt caching at a tenth of the input rate. They built the tools to prevent exactly this.

And they still couldn't control their own spending.

The root cause: teams picked the most expensive model by default and left the guardrails off. Token-billed AI doesn't show up in a purchase order. It shows up five months later in a meeting where engineers are explaining why a project that never shipped cost more than most startups raise.

Amazon already scrapped an internal leaderboard that ranked staff by AI usage. People gamed it by inflating token consumption.

Now they're building automated caps on what projects can spend before the invoice lands.

Audit your AI spending controls today. If your finance team can't see token consumption in real time, you're one runaway project away from a crisis you won't detect until Q4.