Financial Times Reveals Amazon AI Projects Significantly Over Budget
The Financial Times cites insiders saying that multiple AI projects at Amazon have severely exceeded their budgets.
The most serious case involved matching product author information with Claude Sonnet, which failed and cost $1.8 million, exceeding the budget by 860%, and was only discovered five months later.
Additionally, a financial auditing tool overspent by approximately $541,000, and a logistics optimization project mistakenly spent $134,000. The company is developing automated measures to control costs.
Source: Public Information
ABAB AI Insight
Amazon previously shut down its internal AI usage leaderboard because employees were having the Agent perform meaningless tasks to climb the rankings, leading to excessive Token consumption, indicating a misalignment between incentives and cost control.
In terms of capital pathways, the cloud business AWS sells Tokens and cost control tools, but internally, the lack of safeguards has caused the Frontier model's default calls to spiral out of control, exposing a disconnect between deployment and billing mechanisms.
This is similar to early cases in cloud computing where companies misallocated resources, leading to explosive bills. As enterprise-level AI transitions from experimentation to production, it remains in a high-risk phase due to the absence of safeguards.
Essentially, this is a restructuring of cost structures during the technological substitution process, where traditional near-zero-cost mistakes become catastrophic expenses under Token billing, forcing automation constraints and more refined model selection.
ABAB News · Cognitive Laws
- Token bills turn small mistakes into large invoices.
- Unfettered intelligence equals uncontrollable invoices.
- Ranking incentives often create useless consumption.