As token costs shoot up, Dell doubles down on desktop AI

As enterprises confront spiraling cloud inference bills, on-prem AI computing is emerging as the decisive lever for making agentic AI economically viable at scale.

Dell Technologies Inc.’s announcement of Dell Deskside Agentic AI put on-prem AI computing at the center of its enterprise strategy, framing local compute not merely as an alternative to the cloud but as the essential foundation for agentic workflows where token costs can make or break return on investment. The calculus is straightforward: Research agents that burn $600 per cloud run in a single session become a very different financial proposition when the compute is owned outright and sits at the desk, according to Marc Hammons (pictured, right), senior distinguished engineer at Dell.

“The cloud is where frontier models go first. It’s where cutting-edge AI goes first,” Hammons said. “It’s also where your costs are going to be buried if you don’t do something about that. The opportunity … is to bring some of that compute locally on the machine and start to adjust the tokenomics of the situation.”

Hammons and Charlie Walker (left), senior director and general manager for Dell Pro Max and Dell Pro Rugged, spoke with theCUBE’s John Furrier and Dave Vellante at  Dell Technologies World 2026, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how agentic AI is landing at the desktop, the tokenomics of on-prem versus cloud inference and use cases ranging from sovereign research agents to deskside supercomputers. (* Disclosure below.)

On-prem AI computing and the tokenomics of agentic workloads

The economic argument for local compute becomes clearest when the recursive nature of agentic work is examined. Unlike simple prompt-response exchanges, agents iterate continuously — reading, reasoning and acting in loops that compound token consumption rapidly. Enterprises are increasingly discovering that the hardware investment pays for itself in a matter of months once that burn rate is relocated from the cloud to a local machine, Walker noted.

“When you think about a research agent, it’s going out, finding one paper, then it’s finding something else, analyzing it, and so it’s continually learning, continually developing. That burns tokens,” Walker said. “People look at [the Dell Pro Max] and they’re like, ‘That’s a very expensive device,’ but something like a GB300, you pay for it within three to six months. It becomes very economical when you start to actually think about the tokenomics behind it versus just the hardware investment.”

Dell’s portfolio spans from the Dell Pro Max with GB10, capable of running persistent local agents in a compact form factor, all the way to the Pro Max with GB300 — which packs 20 petaFLOPS of performance and 748 gigabytes of coherent memory into a deskside tower. That upper end of the portfolio effectively brings datacenter-scale compute to the individual desk, enabling trillion-parameter-scale agents to run entirely on-prem, Hammons explained. The software stack scales seamlessly across the range, meaning a team can prototype on a GB10 and graduate to a GB300 without changing their tooling or workflow.

“You do need those frontier models in the cloud for repository-scale reasoning where it’s planning at scale,” Hammons said. “But then it can delegate that down to these machines that are deskside and really let it take over and drive the individual efforts.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of Dell Technologies World 2026:

(* Disclosure: TheCUBE is a paid media partner for Dell Technologies World. Neither Dell, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE