原文 · 未翻译
One company reportedly spent $500 million on Claude in one month after failing to cap AI usage
AI is becoming more expensive and companies are scrutinizing their bills more closely. This makes sense, but above all they should develop more AI expertise in controlling the systems.
Microsoft reportedly recently cut internal Claude Code licenses, partly for strategic reasons but also because costs were climbing. Uber's COO said AI spending is getting "harder to justify" as long as the actual return on investment is hard to measure.
Axios now reports a particularly extreme case: an unnamed company allegedly spent half a billion dollars in a single month because nobody set usage limits on Claude licenses. Enterprise AI models often lure companies in with flat-rate pricing, but those plans typically cap the number of requests per model.
Another CTO says employees use AI systems to check the weather. It works, sure, but it costs way more than a regular search. Sophia Velastegui, a former AI lead at Microsoft, told Axios that companies tend to throw AI at tasks nobody wants to do rather than at work that actually drives revenue.
AI is getting more expensive, and companies need to get smarter about using it
These examples point to the same thing. When AI becomes part of how a company makes money, you need people who actually know how to use and steer these systems. New roles like AI agent orchestrators will matter.
The biggest cost drivers are misuse and poor model selection. Misuse often looks like a lack of context engineering, which leads to endless chats with bloated context windows. Poor model selection means throwing a powerful, expensive model at tasks a cheaper one could handle just as well.
Not every task needs generative AI, either, meaning mostly large language or reasoning models. Many things still work better in traditional software. Learning to tell the difference should be a core part of building AI skills inside any company.
AI skills aren't just about costs, of course. Quality suffers too when you don't know what you're doing. A recent example shows Copilot in auto mode completely botching a data analysis task, confidently spitting out heavily biased answers. Switching to a thinking model fixed it. And this kind of slop has its own price tag.
AI News Without the Hype – Curated by Humans