Elektrine lite

← Feed

@MichaelTBacon@social.coop

2026-04-06 12:54 UTC

To be clear, the 10x number isn't something that's based on a specific set of numbers, but a gesture at the direction things are going. Anthropic reported that its $200/month Max users were costing it $50k/month in compute. (https://www.revenuememo.com/p/how-does-anthropic-make-money) Those are the Claude Code users that are reporting amazing coding results, but it's at the expense of using the newest, biggest models burning gigantic compute. But those models *still* need lots of guard rails. How will they fix those? Very likely the best solution they'll find is a *bigger* model and even *more* agents. And we haven't seen any Moore's Law effects come in with GPUs. GPU time still costs basically what it did 3 years ago. (https://semianalysis.com/gpu-pricing-index/)

Replies (1)

  • @MichaelTBacon@social.coop 2026-04-06 13:13

    Of course Anthropic can't eat those kinds of compute losses forever, or even for very long. So they're trying to bring user pricing at least closer to operating costs and it's royally pissing off the vibe coding stans (https://www.nbcnews.com/tech/tech-news/claude-code-ai-mythos-leak-rcna266083). To get revenues to keep up with its usage growth, the big AI guys need something like 10-40x the data center capacity they have now. But note that current data center capacity is *already* generating massive blowback for driving up energy prices, creating noisy heat islands, creating water shortages, and of course heating up the climate, and that was before an Iran-induced oil crisis. There will be something left after all the dust settles, but currently people are advocating adopting tech based on gigantic subsidies with no indication that it will become affordable without subsidies. That's really worrisome.

    Open ##1732738