Insider Perspective on Anthropic's Compute Allocation Challenges

An AI engineer provides an insider's view on the internal dynamics and challenges Anthropic faces in allocating compute resources between its research, subscription, and enterprise customers.

Details

The author, an AI engineer with personal relationships at Anthropic, explains that the company is fundamentally a research lab that hasn't learned how to be a successful product company. The original Claude AI was contracted out to external developers, and the research team's priorities often take precedence over the product team's needs. \n\nInternally, Anthropic has three groups competing for compute resources, each with very different incentive structures and value propositions. The research team requires GPU-hours to train new models, which is a pure cost with no immediate revenue. The subscription business generates flat-rate revenue regardless of usage, while the enterprise/API customers pay per token and represent the majority of Anthropic's revenue growth. \n\nWhen compute is scarce, the author argues that Anthropic is forced to prioritize enterprise customers, as they cannot afford to lose them to competitors like OpenAI. Research also cannot be slowed down, as it is the core of the company's mission. As a result, the subscription users end up bearing the brunt of the compute allocation cuts during peak hours.\n\nThe author suggests that Anthropic's failure to procure enough compute resources proactively over the past two years is the root cause of the current challenges. OpenAI's strategy of buying