Dell Unveils 256-CPU Vera Server and Groq 3 Rack

Dell has unveiled a family of servers built on NVIDIA’s Vera CPU, including a rack that holds 256 of the chips, alongside its first rack-scale inference system based on Groq 3 processors, the company said at NVIDIA’s GTC event in Berlin.

In its press release, Dell pitched the line-up as hardware for agentic AI, the software agents that plan, call tools and act, and which depend heavily on general-purpose CPUs as well as GPUs. Most of the new hardware will not ship until the first quarter of 2027.

Three Vera-era systems, from 19-inch to full rack

The largest is the PowerEdge M9882, which Dell says puts “256 NVIDIA Vera CPUs into a single 48U rack” for running agent workloads at high density. It is due in the first quarter of calendar 2027.

The PowerEdge R8822 is a two-socket, air-cooled Vera server in a standard 19-inch chassis that supports “up to four NVIDIA double-wide GPUs”. Dell says it will be available this month, making it the first of the new machines customers can buy.

The PowerEdge XE9810 is Dell’s first product using Groq 3, the inference chip Groq sells as a language processing unit (LPU). Dell says it is the “first to deliver rack-scale inference featuring 16 direct-liquid-cooled compute trays powered by Groq 3”. DatacenterDynamics reported that Dell did not say how many LPUs sit in each tray. The XE9810 is also due in the first quarter of 2027.

Dell also announced a Modular Data Center built around NVIDIA’s DSX POD design, due in early 2027, and a 64GB version of its Pro Precision GB10 deskside machine, available from 23 October.

Dell’s numbers come from Dell

Dell says its AI Factory with NVIDIA now has more than 6,500 customers, up from 5,000, based on its own analysis from August 2026. It also claims on-premises AI with its hardware can be “up to 90% more cost-effective than public clouds over two years”, citing an analysis by Signal65 and Futurum. Both numbers are vendor claims, and the release gives no detail on how a “customer” is counted.

“AI infrastructure is no longer a cost center. It’s a value creation center,” said Michael Dell, the company’s chairman and chief executive. NVIDIA’s Jensen Huang said: “AI agents reason, use tools and learn—and they need a new kind of computer.”

Agents are pulling CPUs back into the AI story

The most telling product here is the CPU rack, not the GPU servers. For two years AI hardware has been sold by GPU count. Agentic workloads change the mix: an agent spends much of its time on orchestration, tool calls, sandboxes and data handling, which run on CPUs, while models handle the reasoning. A rack of 256 Vera chips is NVIDIA and Dell betting that enterprises will want dense CPU capacity sitting beside their accelerators. Cloud providers are moving the same way, from Modal running sandboxes on Vera to Google Cloud’s Vera Rubin A5X instances.

The Groq rack matters for a different reason. Dell selling Groq 3 hardware alongside NVIDIA’s gives enterprises a dedicated inference option inside a mainstream server catalogue, and suggests NVIDIA is comfortable with that pairing at its own event. Without per-tray chip counts, prices or performance figures, though, buyers have little to compare yet.

The modular data centre fits NVIDIA’s push for factory-built DSX POD AI factories, and the agent focus echoes NVIDIA’s OpenShell blueprint for safe AI agents. The real test of these launches will come in the first quarter, when the machines ship and buyers see prices.