The AI coding boom hit a wall. Developers are flooding GitHub with complaints that Claude Code, the agent that triggered a 'SaaS apocalypse,' is becoming unusably slow and expensive. An internal audit at AMD reveals the scale: between January and March, input tokens for the same tasks increased 170-fold, and one project’s estimated cost exploded from $26 to over $42,000.
Nathaniel Whittemore of The AI Daily Brief frames this as the 'second moment' for AI - a shift from assistants to autonomous builders. Claude Code grew from $1 billion to $2.5 billion in annualized revenue in just two months, its success upending the traditional software stack. Companies like Figma, down 85% since its IPO, are feeling the pressure.
“You realize this is a suicide run... routing users away from hardware they like, to hardware they want their researchers to use.”
- Theo, Nerd Snipe
Theo and Ben on Nerd Snipe argue Anthropic’s problems stem from a hardware crisis. They claim the company is aggressively rerouting public users from optimized Nvidia H100 clusters onto Google TPUs and Amazon Trainium chips to hoard scarce GPUs for internal research. This infrastructure arbitrage explains why Anthropic eliminated the price premium for its 1-million-token context window - it forces traffic onto the cheaper, non-Nvidia hardware where performance degrades.
Anthropic’s engineering moves compound the issue. To prevent competitors from distilling its reasoning, the company began redacting 'thinking' traces from API responses. This requires perfect mapping of thread IDs to hidden server-side data - a brittle system that, when it fails, leaves the model disoriented. A recent cache timeout reduction from 60 minutes to five exacerbates the 'dumbness,' as the model forgets its own logic mid-task.
“The era of using AI to save minutes is over; the era of using it to invent new business models has begun.”
- Nathaniel Whittemore, The AI Daily Brief
The broader market is moving faster than Anthropic’s stability. Whittemore notes enterprise focus has pivoted from time-saving 'efficiency AI' to revenue-generating 'opportunity AI.' The capability overhang is widening: 40% of enterprises are predicted to have working agents by year’s end, while others flounder on data quality. Anthropic’s internal culture, described by Theo as research-first and safety-obsessed, now faces the operational reality of serving millions of paying developers who expect reliability.
The cracks in Claude Code are a stress test for the entire agentic future. As AI shifts from a feature to a foundational competitor, the platform that wins won't just be the smartest - it will be the one that doesn’t break under its own success.
