Dell says its new AI workstations cost 87% less than cloud. It won't say what the workstations cost.

Dell says its new AI workstations cost 87% less than cloud APIs. It will not tell you what the workstations cost.

That is the problem with Dell's Deskside Agentic AI pitch, announced at Dell Technologies World in Las Vegas on May 18 Dell press release. The company showed up with three workstation tiers, a partnership stack built around NVIDIA OpenShell, and a cost claim that every outlet covering the event ran without modification: 87% savings over two years versus cloud API pricing, with break-even arriving in three months Dell press release. The number circulated. The math was not interrogated.

The math requires knowing three things Dell has not published: the hardware price, the power draw under load, and the utilization assumptions baked into the comparison. The company disclosed none of them.

The workstations are real. Dell Pro Max with GB10 handles models from 30 billion to 200 billion parameters and is aimed at individual prototyping Dell product landing page. Dell Pro Precision 9 supports up to five NVIDIA RTX PRO Blackwell GPUs and covers the 30 billion to 500 billion range Dell press release. Dell Pro Max with GB300, the flagship deskside machine, runs frontier-scale models from 120 billion to 1 trillion parameters Dell product landing page. That is a meaningful tiering — the GB300 is the first deskside workstation that can credibly host the full agentic AI lifecycle, from fine-tuning through deployment, on hardware that sits under a desk instead of in a colocation facility.

The specs on the GB300 are available. The machine draws 1,600 watts at full load Dell product page. It ships with 252 gigabytes of HBM3e memory, 496 gigabytes of LPDDR5X system RAM, and a 16-terabyte NVMe SSD Dell product page. The onboard GPU is an NVIDIA DGX B300 Dell product page. Computerworld found a comparable MSI GB300 workstation priced at $97,000 through CDW Computerworld. Dell has not published the Pro Max GB300 price. A machine in this class, fully configured with Dell services and support, will not come inexpensively.

This is the gap in Dell's 87% claim. The cost comparison Dell is making assumes a two-year cloud API spend that it has declined to itemize, against an on-premises capital outlay it has declined to disclose. No independent analyst has published a cost model that uses verified hardware pricing and measured power consumption. The only published analysis cited in Dell's materials is a Signal65 study commissioned by Dell itself Dell press release, which used utilization assumptions of 60 to 80 percent and cloud list pricing rather than enterprise volume rates.

The power cost alone is worth accounting for. A 1,600-watt workstation running continuously under agentic workloads — the scenario where the 87% math becomes relevant — consumes roughly 14 kilowatt-hours per day. At typical enterprise electricity rates of 10 to 15 cents per kilowatt-hour, that is $1.40 to $2.10 per day in power, or roughly $500 to $750 per year. Scale that across a small team and the annual power tab arrives fast. It does not change the two-year math dramatically, but it is part of the math, and Dell did not include it in the headline.

Jeff Clarke, Dell's vice chairman and chief operating officer, framed the logic plainly: "The most efficient token is the one produced closest to the data" Dell press release. That is architecturally sound. Enterprise data that never leaves the building, latency that does not depend on an API being available, compliance surfaces that map cleanly to data residency requirements — these are genuine advantages. The cost claim does not follow automatically from them.

Michael Dell described the product differently on LinkedIn: "The Dell Deskside Agentic AI isn't just a workstation — it's a fully governed entry point into the Dell AI Factory with NVIDIA" Michael Dell, LinkedIn. That framing is worth taking seriously. If Dell is selling an architecture, not just a box, the price of the box is not the whole story. But the company cannot have it both ways: the 87% savings pitch requires explicit hardware math, and the explicit hardware math has not been published.

NVIDIA OpenShell, the sandboxed runtime that sits at the center of the product stack, is a genuine piece of the value proposition. It lets developers build and test agents locally before pushing them to production infrastructure Dell press release. More than 50% of agentic workflows already run on open-weight models in the 30 billion to 284 billion parameter range, according to Dell's own data Dell press release, and a local sandbox makes that workflow tractable without burning cloud credits. Dell says it has more than 5,000 customers running through its AI Factory Dell press release #2 and can accelerate time-to-value by up to 84% Dell press release #2. Those are the company's numbers, but they point at a real market dynamic.

The countervailing force is cloud API pricing, which has dropped more than 90% over two years and shows no sign of plateauing. If token costs fall another 50% in the next twelve months — a conservative extrapolation given current competition among frontier model providers — the two-year payback window Dell is promising stretches. The break-even math only works if cloud inference costs stay where they are. There is no evidence they will.

A founder evaluating this pitch, an engineer pricing out an infrastructure proposal, or an IT buyer running the numbers for a finance committee will need the numbers Dell has not published. What does the GB300 actually cost configured? What does it draw under sustained agentic workloads? What utilization rate did the 87% figure assume? These are not unreasonable questions. They are the questions the 87% claim makes it necessary to answer.

The workstation hardware is credible. The architecture case has merit. The cost claim does not survive the absence of the numbers it depends on. Until Dell publishes the denominator, the 87% figure is a marketing number. The room full of reporters who ran it without asking what it was built on tells you something about how the technology press covers vendor math. This is not a compliment.

The product launches May 29. The pricing, Dell says, will be available at that time ServeTheHome.