Root Cause

Agentic Sufficiency Curve

Nov 22nd, 2025
Chris Rohlf

Since 2022, the USG has restricted access to advanced AI GPUs through successive rounds of export controls, all aimed at limiting adversary access to large scale compute. While the specific parameters of those export controls and associated policies continue to be modified there is still general consensus around the goal of preventing access to significant quantities of high end AI compute by adversaries.

This strategy assumes the decisive variable in AI competition relies on access to the most capable GPUs in mass quantities. There is good reason to support this logic given the value of AI in high stakes domains like intelligence and warfare. But who wins the AI race won’t be decided by who can train the best models alone. There are many domains where a small number of GPUs used for inference of the latest open weight models are capable of producing incredibly valuable outputs. This is obvious to anyone who has run an open weight model, or written an agent to autonomously perform a task. While the cost of inference is coming down on a per-chip basis, the number of AI workloads is growing and along with it the volume of tokens and so costs overall are rising. The realities and constraints of those spiraling costs means there is incentive to minimize GPU compute and to exploit the capability overhang in existing models to drive agentic workloads that primarily run on CPUs at significantly lower costs.

In this piece I introduce the ‘Agentic Sufficiency Curve’, a threshold where GPU capacity becomes good enough for a specific problem domain through accurate initial inference and agentic throughput begins scaling on CPUs, memory, and I/O.



States don't necessarily need infinite frontier silicon to win the AI race, they need sufficient inference capacity and vast amounts of cheap CPU that agents can leverage. In other words, once the blue line intersects with the red line, agentic completions take off and the bottleneck shifts from GPU TFLOPs to commodity CPU cycles, memory bandwidth, and I/O. The ‘Agentic Sufficiency Curve’ is just an attempt to formalize something that is likely very obvious to anyone who has written an agent. The agent itself is a probabilistic program that is partially produced by inference on GPU but that executes on low cost CPU.

There is also good reason to believe the cost of frontier model inference will continue to fall. Model architectures improvements such as batching, quantization, speculative decoding, MoE architectures, and others have all pushed the sufficiency point left. This means older GPUs can still clear the bar for inference that meets the threshold and deliver the tokens per-second needed to drive and scale agent deployments. Semiconductor export controls aim to protect access to the latest silicon but not the area under the curve. Still, GPU compute will remain more expensive than CPU for the foreseeable future and this means it is in the best interest of companies and governments to shift as much of these workloads to agents (which execute on CPUs) after the sufficiency threshold is met. Accurate inference is still critical because the model’s first few trajectories determine the entire direction of the agentic workflow. A strong initial inference minimizes wasted CPU cycles, and prevents the agent from wandering into dead ends that require expensive replanning steps. Every low accuracy agent trajectory forces a return to GPU bound inference, so precision up front compounds efficiency across the whole agentic loop.

As an example, agent workloads in software vulnerability discovery and mitigation largely center around using LLMs to drive traditional tooling such as fuzzers and static analysis tools. These combine short durations of token generation with long tails of planning, retrieval, tool use, and verification. In these workloads, GPUs idle while agents wait on network and system calls, overall throughput and agent output improves from smarter process scheduling, concurrency, and fewer cache misses than from total GPU TFLOPs alone. This won’t be true of all problem domains, and as models get better and more accurate we may reach a point where a small number of tensor operations performed in milliseconds results in the same outcome and cost as hundreds of millions of CPU instructions over a longer period of time.

It’s all but guaranteed that frontier AI silicon will remain strategically important, but once an organization amasses enough inference capacity to solve for a specific problem domain, the winning condition shifts to agentic saturation driven by cheap CPU. If policy makers fixate only on who amasses the newest chips, they will miss the agentic crossover and the point where a domain specific advantage comes from agent deployment, not amassed TFLOPs intended for training runs.