Claude Is Melting Down. AI's Compute Crisis Explained.

By AI For Humans

Categories: AI

Summary

Major AI labs face an immediate compute crisis: Anthropic's Claude has demonstrably degraded performance as tokens used dropped 50% from January to March, forcing companies to ration compute for existing models while preparing expensive launches of bigger models like Mythos—a bottleneck that could delay widespread adoption of next-generation AI.

Key Takeaways

  1. Claude's reasoning tokens dropped from thousands to hundreds on basic queries between January-March, a 50% reduction that directly correlates with degraded model performance as Anthropic manages compute constraints.
  2. Compute constraint mechanics mirror legacy telecom pricing: companies now incentivize off-peak usage and limit reasoning power per query to manage congestion, sacrificing user experience for infrastructure margins.
  3. The traditional AI adoption cycle is broken: benchmark performance degrades as user adoption increases, creating a trough period where released models become demonstrably worse before next-gen launches.
  4. Major AI labs face a strategic timing trap: they must allocate scarce compute between serving millions of current users and provisioning infrastructure for upcoming flagship models, forcing trade-offs that harm existing products.
  5. Users and developers can detect performance degradation by monitoring token usage patterns; AMD's own AI director confirmed Claude's token reduction, validating user complaints as measurable, provable constraints rather than perception.

Topics

Transcript Excerpt

Big AI models from Anthropic and OpenAI are coming. We know this, but will anybody be able to use them? >> OpenAI's new Spud model might come out this week, and Anthropic's Opus 4.7 is actually right around the corner, but Anthropic right now today is already struggling to serve its current models. Claude feels downright useless to some people right now. And this could all delay a wider release of the big daddy model that is >> Mythos. I like to think of Mythos as mommy Kevin. >> Okay. What is M...