Amazon’s $1.8m Claude Blunder Shows AI’s Runaway Costs
Amazon spent $1.8 million on an Anthropic Claude job that ultimately failed, highlighting the issue of AI’s escalating costs.
Internal Metrics Reveal Overbudget AI Projects
Amazon’s internal data reveals AI projects consistently exceeding budgets, including a notable example with Claude for an author-matching task. This job cost $1.8 million, a staggering 860% over budget, and went unnoticed for five months before its failure. The irony? AWS, Amazon’s own cloud platform, offers tools to prevent such overspending.
Key Findings:
-
Cost Overruns: A $1.8m Claude job, an author matching tool, ran 860% over budget, took 5 months to detect, and ultimately failed. Two other smaller projects also overshot their budgets.
-
The Hidden Bill: Unlike human errors that halt operations, AI models keep running and silently generate invoices, making it harder to track costs.
-
Per-Token Pricing Impact: The shift to per-token pricing, coupled with high token usage in AI models, exacerbates these bills.
-
AWS Solutions: AWS’s Bedrock provides batch inference, cheaper tiers, prompt caching, and routing to optimize costs. Haiku, a cheaper alternative model, was available from Anthropic.
Amazon’s Response: Amazon downplayed the incidents, attributing them to "experimentation" within a small number of teams. However, these overspends represent significant fractions of their vast quarterly revenue.