The claim arrives without its workings
The sentence Dell most wants quoted is the cost one. Owning a Dell AI Factory, the release says, "can be up to 90% more cost-effective than public clouds over two years."
Everything needed to test that sentence is missing from the release's main text. No study is named there, no workload, no system size, no cloud service used as the comparison, and no statement of what the two years of ownership include. The support sits in footnote 6, which the trade-press republications of the release dropped: a "validated analysis by Signal65 and Futurum Group", The Economics of Agentic AI, September 2026. The footnote says the savings assume a two-year deployment and a range of Dell deskside and datacenter solutions running general knowledge, sales and software development workloads, and that the analysis counts estimated cloud discounts alongside infrastructure hosting, energy, infrastructure management and Dell support services costs.
The study itself is more specific, and less convenient. Signal65 benchmarked nine Dell AI Factory with NVIDIA platforms, from a single-GPU deskside system to a 128-GPU PowerEdge cluster, and modelled three agent workloads over a two-year cost horizon: a knowledge worker, a sales agent and an always-on software development agent. The comparison was against the same open-weight models running on AWS Bedrock and, separately, against selected frontier API models, which were compared by price tier and not benchmarked for equivalent capability. Against Bedrock, the modelled savings ran from 8% to 90% for the knowledge worker, 41% to 85% for the sales agent and 8% to 88% for the software development agent; the study's summary puts the overall cost reduction at 28% to 90% or more. Dell's headline takes the top of that range and restates it in different terms: a cost-effectiveness claim against "public clouds" as a whole, where the study measured modelled cost savings against one cloud's model service and a set of APIs. "Up to" does the rest.
The rest of the release's numbers are built the same way. Dell also cites projections that agentic AI will become the largest workload by 2028 and drive 24 times today's token consumption; the release's footnotes attribute those projections to Omdia, Goldman Sachs and Gartner research. They are projections either way, and none of them is evidence about the 90%.
Dell says its AI Factory customer count has risen from 5,000 last quarter to more than 6,500, an increase of over 30%. The release does not define a qualifying customer or disclose associated revenue. Dell's own footnote says the count is based on its August 2026 analysis of customer order data. A count without a definition can still mean several things at once: a customer running production agents on Dell hardware, a pilot, or a single deskside system.
Two racks are the actual news
Under the projections, two systems are interesting, and both are 2027 products.
The PowerEdge M9882 puts 256 NVIDIA Vera CPUs in one 48U rack and gives them the agent's unglamorous work: orchestration, tool use, code execution and evaluation, while separate GPU clusters handle model inference. That split is a reasonable reading of where agent workloads strain hardware. Agents spend much of their time running tools and checking results, not generating tokens, and Dell proposes not paying GPU prices for that half of the job.
The PowerEdge XE9810 goes at the other half. Dell calls it the first rack-scale inference system built on Groq 3 LPUs, with 16 liquid-cooled trays, positioned as a complement to GPU clusters for high-throughput inference. An inference rack sold under the Dell and NVIDIA names but built on LPU silicon rather than GPUs is a notable change in what a large vendor will put in a rack. The release, however, publishes no throughput, latency or cost figures for it, so the claim rests on architecture rather than on any measured result.
The calendar is the quiet part of the announcement
Read the availability list as the announcement's honest section. Of the headline systems, only the R8822 server is due this month. The 64GB GB10 desktop follows on October 23. The M9882, the XE9810 and the modular data centre land in the first quarter of 2027 or later. The GB300 desktop and the telecom edge framework are "coming soon," with no date at all.
The immediately available elements are a mixture of existing or newly available hardware, software blueprints and validated solutions. The two most distinctive new rack systems are scheduled for the first quarter of 2027.
None of this makes the direction wrong. Enterprises running agents against regulated data have reasons to own the hardware, and Dell's installed base gives it a route into those accounts. But a buying decision needs the two things this announcement still withholds: prices for the new systems, and a cost comparison for the machines actually being announced. The Signal65 study predates this launch and benchmarked GPU-based platforms; neither the Vera CPU rack nor the LPU inference rack is among them. Dell has promised the hardware for 2027. The prices should not take as long.
Sources
- [1] Dell Technologies, "Dell Technologies Advances Enterprise Agentic AI from the Deskside to Data Center with NVIDIA" (press release, October 21, 2026; read on Dell Technologies' investor newsroom, including footnote 6)Read source
- [2] Signal65 (Mitch Lewis and Ryan Shrout), "The Economics of Agentic AI: On-premises Deployments with Dell AI Factory with NVIDIA vs. Cloud" (executive summary, September 2026; hosted by Dell Technologies)Read source
- [3] DQ Channels, "Dell AI Factory expansion with NVIDIA takes agentic AI enterprise-wide" (October 26, 2026; release text cross-check)Read source