Anthropic is mortgaging its product quality to survive a compute shortage. To preserve scarce Nvidia H100 clusters for its own researchers, it began forcing public users onto less-optimized hardware from partners like Amazon and Google. This was a suicide run, argued Theo on Nerd Snipe.
The result was a severe degradation in Claude's performance. The engineering burden of hiding the model’s 'thinking' traces to prevent competitive training created a fragile system. When caches fail, the model loses its reasoning history, becoming less coherent. On Nerd Snipe, Theo suggested the company’s software is too fragile to handle this complexity, resulting in a dumber assistant.
"Because the model needs its own prior reasoning to maintain context, Anthropic must perfectly map thread IDs to hidden server-side data. If the ID link breaks or the cache expires, the model loses its reasoning history."
- Theo, Nerd Snipe
The perception isn't just vibes. An internal audit at AMD showed the input tokens required for similar coding tasks in Claude Code spiked 170-fold between January and March. Estimated costs for one project ballooned from $26 to over $42,000.
Desperate for capacity, Anthropic turned to a rival. It is now leasing the 300-megawatt Colossus 1 data center, a facility with 220,000 GPUs controlled by Elon Musk’s interests. On FYI, Frank Downing noted the facility is ideal for inference but marks a sharp strategic pivot for Anthropic, which once blocked Musk’s xAI from its models.
"The lease cleans up the financials for a potential SpaceX IPO. It demonstrates a path to monetization for massive capital expenditures even if his own model, Grok, doesn't achieve immediate market dominance."
- Frank Downing, FYI - For Your Innovation
The crisis exposes a core tension in frontier AI. Model builders are willing to sacrifice product reliability to secure the compute required for scaling. The deal with Musk shows that when the power runs out, competitive distance no longer matters.
