Agents multiply demand: AMD's Venice, Helion racks and ROCm Agentic Toolkit put the CPU back at the centre of AI

Detailed summary
AMD used its agentic AI breakout at In Depth 2026 to walk through every stage of an agent's work, show measured results against Intel, Arm and Nvidia, and reveal the racks and...
Alex Zaharov-Reutt, Global AI and Technology Editor | Published 19 November 2026, 6:50 pm AEDT
AMD used its agentic AI breakout at In Depth 2026 to walk through every stage of an agent's work, show measured results against Intel, Arm and Nvidia, and reveal the racks and open source software it has built to run hundreds of thousands of agents side by side.
For about two years, the server CPU was the least exciting chip in the AI data centre. GPUs got the headlines, the budgets and the power feeds.
That era is over, and AMD spent a full breakout session on day two of AMD In Depth 2026 in Las Vegas explaining why. The presenters were AMD Corporate Vice President, Compute and Enterprise AI Product Group, Madhu Rangarajan, and AMD Corporate Vice President, Software Development, Anush Elangovan, the Nod.ai founder who now runs AMD's ROCm software effort.
Rangarajan summed up the chatbot years of 2024 to 2026 in five words: "During that time, CPUs were boring."
The opening argument was about arithmetic. Every customer AMD talks to has to keep its existing compute running, absorb the growth in agents, and absorb the growth in tokens, all within fixed limits on power, floor space and dollars.
In the chatbot era, a person typed a prompt, a GPU produced an answer, and the CPU's job was mostly to feed the GPU. Rangarajan put the ratio for those years at roughly one CPU for every four GPUs.
Agents change the maths because they do things. They plan, look things up, run tools, check their own work and try again. AMD's view is that the ratio is already 1:1 at the data centre level and still growing, with 2:1 a real possibility.
The slide behind him carried a Goldman Sachs Research forecast that agents will multiply token consumption 24 times between 2026 and 2030, to about 120 quadrillion tokens a month. Rangarajan tied that curve to a separate number, the forecast shown in the keynote from AMD Senior Vice President and General Manager, Compute and Enterprise AI, Dan McNamara, of a server…
Opening of the article, excerpted from iTWire. To read the rest, open the article on the publisher's site.
MarketNews shows the headline and summary published by the outlet. The full article is on the publisher's website. (itwire.com)

