Agents Move From Inference to Execution, and Buyers Gain Power

Jaya Gupta and Ashu Garg, investors at Foundation Capital, argue in a long-form X article that the money in AI is moving away from whoever owns the most GPUs and toward whoever assembles agents (X article). Their core distinction is between inference and execution. Inference is the model call that interprets a request or judges an unfamiliar option, and it runs on GPUs. Execution is everything else an agent does to finish a job: scheduled checks, database writes, network calls, retries and confirming whether a transaction went through. That work runs on ordinary CPU-based services. Their example is an assistant told to watch flights for two weeks. Most of that job is a worker checking fares every hour and calling a model only when something needs judgment.

The authors' point is that the company assembling the agent decides what each step needs. It can choose a frontier model, a small cheap model, cached or reused results, or no model at all, and it can pick where the code runs and whose sandbox, gateway or browser it uses. Because agent work is increasingly bought per request, they write, the assembler can change that mix on every run. They cite signs that execution demand is rising, including agent requests on Cloudflare up more than 1,700% in a year and Intel describing the CPU-to-GPU ratio moving toward one-to-one in agent deployments. The risk they see falls on suppliers who financed capacity against expected purchases. They use CoreWeave as the example: revenue more than doubled while interest costs ate about a quarter of it and its bond yields neared 15%. Their historical parallel is Nortel and Lucent, whose equipment orders collapsed even as the internet kept growing.

This is an investment thesis, and the authors present it as a bet. The article's figures are their citations, and the stored text does not include the sources behind them. Predictions such as cybersecurity becoming one of the largest enterprise uses of inference, possibly 30% next year, are explicitly hedged as their expectation. The rest of the thread is a one-word reply and a post labelled Example that links to another account's post, with no further argument from the authors.