From Inference to Execution: AI’s Next Economy
Summary
Jaya Gupta and Ashu Garg argue that AI's economic value is shifting from model inference to the execution layer, where CPU-based services handle the majority of agentic workloads.
Execution dominates agent workloads.
A single flight-watching assignment requires only six model calls but triggers over 1,000 tool calls and 1,300 database operations.
Most of the job is execution.
CPU demand scales with agents.
Intel expects the CPU-to-GPU ratio to move from one-to-eight in training toward a one-to-one mix in agentic deployments.
one-to-one in agentic deployments
Assemblers act as negotiators.
Companies managing the mix of intelligence, execution, and routing must constantly optimize across hundreds of models and sandbox partners.
the companies assembling agents become always-on negotiators.
Platform wars target the route.
Control over the path to the model—via browsers, edge, or business apps—determines which suppliers get paid and what models are reachable.
The next platform war is over who gets to turn everyone else into a supplier.
GPU ownership loses value primacy.
The link between owning the most GPUs and capturing the most value is breaking as the economy moves toward execution.
breaks the link between owning the most GPUs and capturing the most value.
Highlights
Multiply these choices together and the companies assembling agents become always-on negotiators.
My bet is that the next phase of AI breaks the link between owning the most GPUs and capturing the most value.
I expect cybersecurity to become one of the largest enterprise uses of inference (maybe even 30% next year).
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.