saved
From Inference to Execution: AI’s Next Economy
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
In an X Article, Jaya Gupta and Ashu Garg argue that AI value is shifting from the suppliers of models and GPUs to the companies that assemble agents and decide where each step runs. Agent work is mostly CPU-based execution, bought per request across many suppliers, so GPU-heavy neoclouds like CoreWeave risk a Nortel-style mismatch between success and spending.
ideas
- Execution dwarfs inference. An illustrative two-week fare watch needs six model calls but thousands of scheduled wake-ups, tool calls and database operations, driving demand for CPUs, storage and networks.
- The assembler sets the mix. Builders pick among hundreds of models, small decision models, on-device inference, sandboxes and gateways per request, so every supplier competes for the next call.
- Every layer wants the route. Browsers, devices, clouds, labs, business apps and security vendors each try to own coordination and turn the others into suppliers.
- Capital is exposed. CoreWeave's doubling revenue sits beside heavy interest costs and near-15% bond yields; like Nortel and Lucent, suppliers can lose even as usage grows.
quotes
“Most of the job is execution.”
“Multiply these choices together and the companies assembling agents become always-on negotiators.”
“The technology succeeded while the purchases expected to pay for its buildout collapsed.”
“The AI boom can succeed while parts of the investment thesis financing it fail.”