Agentic AI could reshuffle compute
Some analysts argue the next wave of 'agentic' AI could move heavy workloads from GPUs back toward CPU and other architectures — a potential long‑term pressure on GPU incumbents. (The note says agentic AI may shift compute to CPUs, which could pressure NVIDIA, and that ARM stands to gain from increased royalties.) (x.com) If that thesis plays out, it’s a structural investing theme — not an overnight trade — but worth tracking alongside GPU demand signals.
A normal chatbot does one expensive thing: it sends your prompt through a giant model on a graphics processing unit, which is the chip built to do huge amounts of math in parallel. An agent does a different job: it plans, calls tools, reads files, clicks buttons, and loops through many smaller steps until the task is done. (openai.com) OpenAI’s developer docs now describe agents as applications that “plan, call tools, collaborate across specialists, and keep enough state to complete multi-step work,” which is much closer to a software worker than a single answer box. The same docs show built-in tools for web search, file search, code execution, and computer use. (developers.openai.com 1) (developers.openai.com 2) That changes where the computing bill lands. The giant model still runs best on a graphics processing unit, but the surrounding work — scheduling steps, storing state, checking permissions, moving data, and calling outside software — often runs on central processing units, which are the general-purpose chips that handle the rest of the server. (nvidia.com) (developers.openai.com) NVIDIA’s own product language shows the split. Its Blackwell Ultra systems are marketed for “AI reasoning” and large-scale inference, while its open-source Dynamo stack is described as an orchestration layer that handles routing, caching, and scaling above the model engine. (developer.nvidia.com) (github.com) Once software starts breaking one request into 20 or 50 smaller actions, the bottleneck can move from raw matrix math to coordination. A system that spends time waiting on databases, browsers, application programming interfaces, and approval checks can end up caring as much about memory, networking, and CPU scheduling as about the next graphics processing unit. (openai.com) (github.com) That is the core of the “compute reshuffle” thesis. It does not say graphics processing units disappear; it says the share of spending around the model may grow faster if companies deploy more agents that behave like software pipelines instead of one-shot chat prompts. (developer.nvidia.com) (openai.com) There is a reason to be careful with that claim. The newest MLPerf Inference results added a reasoning benchmark based on DeepSeek-R1, and NVIDIA’s GB300 NVL72 posted record throughput there, which is a reminder that long-chain reasoning inside the model can make graphics processing units even more valuable, not less. (mlcommons.org) (blogs.nvidia.com) NVIDIA is acting as if both sides will grow. In March 2026 it launched Vera, a central processing unit it called “purpose-built for the age of agentic AI,” while still pushing Blackwell and Blackwell Ultra as the engine for reasoning inference. (techpowerup.com) (developer.nvidia.com) Arm sits in the middle of this because it licenses central processing unit designs rather than selling only one finished chip. In its filings Arm says the CPU is vital in AI systems whether it handles the workload itself or works with a co-processor such as a graphics processing unit, and in its fiscal 2025 fourth quarter it reported more than $600 million in royalty revenue for the first time. (sec.gov) (newsroom.arm.com) So the question is not “graphics processing units or central processing units.” The real question is whether enterprise artificial intelligence shifts from one giant answer per prompt to fleets of agents doing thousands of smaller software actions, because that would spread more value across CPUs, memory, networking, and orchestration layers even if the biggest models still live on GPUs. (developers.openai.com) (nvidia.com)
