Dell Technologies unveiled the most complete expansion of its Dell AI Factory with NVIDIA to date at the GTC Berlin conference on Wednesday, establishing three distinct compute tiers for enterprise agentic AI — a standard-rack Vera CPU server shipping this month, a 256-CPU disaggregation rack and a rack-scale Groq 3 Language Processing Unit inference system due in Q1 2027, and a containerized AI Factory deployment concept targeting early 2027. The announcements, made during Jensen Huang's keynote at Berlin's Tempodrom, position Dell as the primary hardware integrator for the full NVIDIA Vera Rubin platform as enterprises shift from generative AI pilots to always-on agentic deployments.

The Dell AI Factory with NVIDIA has reached 6,500 customers as of August 2026 order data, up from 5,000 the prior quarter — an acceleration Dell attributes to enterprise AI infrastructure reaching an inflection point where owned compute begins to outperform hyperscaler spend.

Agentic AI Changes What Enterprise Infrastructure Must Do

Dell's four-product portfolio at GTC Berlin is built around a specific projection: agentic AI will drive 24 times today's token consumption, even as per-token inference costs fall more than 90 percent by 2030. Omdia research cited by Dell places agentic AI on track to become the single largest enterprise workload by 2028, growing nearly three times faster than generative AI by 2030.

That trajectory produces a different infrastructure problem than general-purpose AI inference. Agentic systems run continuously rather than in discrete query batches — they maintain state, invoke tools, call APIs, execute code, and orchestrate multiple specialized models in parallel pipelines that may remain active for minutes rather than milliseconds. The compute architecture that emerged for transformer-based generative AI — GPU clusters with large high-bandwidth memory capacity, optimized for loading enormous parameter weights — does not map cleanly to these workload profiles.

"AI agents reason, use tools and learn — and they need a new kind of computer," said Jensen Huang, NVIDIA's founder and CEO, at the GTC Berlin keynote. "Together, Dell and NVIDIA are building the computing foundation for the age of AI — from the deskside to the data center — so companies can build and run AI agents securely at industrial scale."

Michael Dell framed the strategic shift at the enterprise level: "AI infrastructure is no longer a cost center. It's a value creation center. Companies are expanding and reallocating budgets because AI infrastructure drives growth, productivity and competitive advantage."

What Ships This Month: PowerEdge R8822

The Dell PowerEdge R8822 ships October 2026, bringing Vera CPU performance to standard enterprise data centers in an air-cooled, two-socket 2U form factor designed for 19-inch rack environments. Dell targets the system at agent orchestration, sandboxing, reinforcement learning, data processing, and analytics, with support for up to four double-wide GPUs for acceleration-heavy tasks.

The NVIDIA Vera CPU at the R8822's core is an Arm-based processor built on 88 Olympus cores per socket, delivering 176 threads and up to 1.2 TB/s of memory bandwidth per socket. NVIDIA positions Vera as delivering up to twice the power efficiency and 50 percent higher performance than traditional rack CPUs — claims directly relevant to the orchestration and retrieval workloads that accompany GPU-based inference in agentic pipelines.

The R8822's air-cooled form factor has practical strategic value beyond its immediate specifications. The higher-density Vera systems arriving in 2027 require liquid cooling infrastructure that many enterprise data centers do not yet have; the R8822 gives organizations an immediate on-ramp to Vera CPU capacity in facilities where they can deploy today, without infrastructure upgrades.

How CPU Disaggregation Works: Why Agentic AI Demands It

Current enterprise server practice bundles CPU and GPU resources in fixed ratios: a dual-socket server with two to eight GPUs per node handles both orchestration (CPU) and inference (GPU). In generative AI workloads, this coupling worked acceptably. In agentic deployments, it becomes an efficiency problem — CPU orchestration workloads are bursty and highly concurrent, with thousands of simultaneous agent sessions each managing state, invoking tools, and calling APIs, while GPU inference is a distinct bottleneck that occurs at model execution time. The two workload types are rarely peak-utilized simultaneously, making fixed CPU-GPU ratios wasteful for either pool.

CPU disaggregation resolves this by separating the orchestration and inference pools into independently sized racks. GPU clusters handle model inference; CPU racks handle the orchestration, tool-call execution, state management, and code-execution layers that agentic pipelines require. Each pool can be scaled to its actual workload independently.

PowerEdge M9882: Dell's Vera CPU Disaggregation Rack

The Dell PowerEdge M9882 is Dell's implementation of NVIDIA's Vera CPU Rack reference architecture: a single 48U rack holding 256 NVIDIA Vera CPUs. Each Vera CPU carries 88 Olympus cores, bringing the total rack to 22,528 cores. NVIDIA's reference specification for the platform calls for 400 TB of LPDDR5X memory with 300 TB/s aggregate bandwidth, 64 BlueField-4 DPUs, and support for 22,500 concurrent CPU environments. Dell has not yet published the M9882's own memory or networking configuration separately from the NVIDIA reference spec.

The M9882 is due in Q1 2027. It is the second CPU-only agentic rack announced this month: AMD unveiled its Helion CPU disaggregation blueprint at the OCP Global Summit on October 12, packing up to 32,768 EPYC Zen 6 cores in an open OCP design with availability targeted for November 2026. The back-to-back announcements from Dell and AMD confirm that CPU disaggregation is becoming a recognized infrastructure pattern for agentic AI — not a niche architectural experiment.

What Makes the Groq LPU Different: SRAM Bandwidth and Deterministic Latency

The distinction between GPU inference and LPU inference is a memory architecture question, not a raw compute question. GPUs — including NVIDIA's H200 and Blackwell series — are designed around HBM (High Bandwidth Memory), a 3D-stacked DRAM technology that provides large memory capacity (up to 192 GB per chip in some configurations) at high bandwidth by DRAM standards (up to approximately 3.35 TB/s for HBM3e). This capacity is essential for loading full parameter weights of large language models: a 70-billion-parameter model in FP8 precision occupies roughly 70 GB of memory.

Groq 3 LPUs invert this tradeoff. Each LPU chip carries 500 MB of SRAM delivering 150 TB/s per-chip bandwidth. SRAM eliminates the off-chip memory access latency that defines HBM-based systems. A full Groq 3 LPX reference rack houses 256 LPUs with 40 PB/s SRAM bandwidth and 128 GB of total SRAM — a collective bandwidth figure no HBM architecture approaches at equivalent power.

The architectural consequence is not just higher throughput — it is deterministic throughput. SRAM has no cache hierarchy, no cache miss penalty, and no speculative memory access; instructions execute in fixed, predictable cycles. GPU HBM systems have variable latency depending on memory access patterns, cache state, and batch composition. Groq LPU systems deliver consistent, predictable inference latency on every request. For enterprises running agentic AI in production — where time-to-first-token consistency determines user experience quality and service-level agreements — deterministic latency is a structural engineering property, not a benchmark figure. It is what makes LPU infrastructure specifically valuable for the high-throughput, continuously-running agentic workloads that are now the fastest-growing enterprise AI category.

Ian Buck, NVIDIA Vice President of Hyperscale and HPC, stated at GTC Berlin that the combination of Groq 3 LPX racks alongside Vera Rubin NVL72 GPU clusters could deliver 35 times higher throughput per megawatt of power, with 10 times more revenue opportunity for AI infrastructure operators — figures reflecting the complementary rather than competing architecture of the two system types.

The Groq 3 LPU was introduced at NVIDIA GTC San Jose in March 2026, following a December 2025 Groq technology license that also brought Groq founder Jonathan Ross and president Sunny Madra to NVIDIA. Groq Inc. retained its inference cloud business and is among the first deployers of Groq 3 LPX infrastructure — doing so in partnership with Dell. "Groq's inference cloud demands infrastructure built for speed and scale, and that's what Dell Technologies brings to the table," said Arunkumar Narayanan, Senior Vice President for Compute and Networking at Dell Technologies.

PowerEdge XE9810: Dell's First Rack-Scale Groq LPU System

The Dell PowerEdge XE9810 is Dell's first rack-scale inference system built on Groq 3 LPUs: 16 direct-liquid-cooled compute trays in a rack format designed for dedicated LPU inference deployments. Dell has not published the LPU count per tray or the XE9810's total aggregate throughput specification; the system is positioned as a disaggregated inference complement to existing GPU clusters, explicitly targeting high-throughput production agentic AI workloads. Availability is Q1 2027.

Groq CTO Sinclair Schuller confirmed the partnership will deliver interactive AI inference to developers building the next generation of agentic applications.

The XE9810's direct-liquid cooling requirement is a practical planning consideration: like the M9882, it requires data center liquid cooling infrastructure. Enterprises evaluating the three-tier architecture — R8822 for air-cooled orchestration now, M9882 and XE9810 for rack-scale tiers in 2027 — will need to account for the cooling infrastructure gap between current facilities and the 2027 systems' requirements.

Dell Modular Data Center with NVIDIA DSX POD

Between individual servers and the rack-scale systems sits the Dell Modular Data Center with NVIDIA DSX POD containerized design: that integrates compute, storage, networking, rack-level power distribution, and liquid cooling at the factory before the unit ships. Dell's headline deployment claim is that stand-up time at a prepared site drops from six months — the typical timeline for conventional heterogeneous rack assembly and validation — to 14 days. Capacity scales by adding containers rather than building out conventional data center space.

The "prepared site" qualification is substantive: the site must already have high-voltage power infrastructure (likely three-phase), a water or coolant loop connection, and network handoff capability. Site preparation itself may require months of facility work; the 14-day figure describes the container deployment at a site already meeting those requirements.

The DSX POD draws on Dell's modular data center heritage, a product category Dell has shipped for more than 20 years, and integrates it with NVIDIA's AI inference and agentic workload reference architecture. It is slated for early 2027.

GB10 64GB Deskside Arrives October 23

On the deskside, Dell is expanding its Pro Precision with GB10 — the compact developer workstation built on NVIDIA's GB10 Vera Nano Superchip, integrating a Vera CPU, a Blackwell GPU, and unified memory — with a 64GB unified memory configuration set to ship October 23. The system supports two- and four-unit clustering configurations, with Dell noting the ability to scale to an eight-node GB10 cluster for greater model capacity and concurrent user support. Dell has not published pricing for the 64GB configuration; it arrives on the same date as NVIDIA's DGX Spark 64GB, which carries a $4,999 list price.

Further out, the Dell Pro Precision with GB300 will bring NVIDIA's Grace Blackwell Ultra superchip to Windows desktops with up to 748 GB memory, 20 petaflops of FP4 compute — delivering a tier of deskside AI capability previously reserved for dedicated AI workstations. Dell has given no ship date beyond "coming soon."

Software, Security, and Deployment Templates

Alongside the hardware, Dell released on GitHub five validated reference deployment templates under the Dell Deskside Agentic AI Solutions for NVIDIA program, covering agentic coding, agentic VSS, NVIDIA NemoClaw, Hermes Agent on NemoClaw, and a Higher Education AI Lab platform. Each template specifies the hardware, models, and software stack for its use case and scales from a single developer to 1,000 or more enterprise systems.

The AI Factory software ecosystem is expanding in parallel: Reflection's Beam open-weight model — an Apache 2.0-licensed model designed for software engineering and long-horizon agentic tasks — and Mistral's knowledge assistant and agentic AI platform are available immediately. AI observability via Dynatrace and computer vision via Fogsphere follow in Q4 2026; SUSE Linux joins the supported OS list in early Q1 2027.

Dell's three-layer AI Factory security architecture for autonomous agents uses hardware-rooted Dell infrastructure at the base, Red Hat AI Factory software for platform orchestration in the middle, and NVIDIA's OpenShell reference architecture for runtime governance spanning runtime, network, data, and infrastructure layers.

Enterprise Traction and the Ownership Calculus

Dell's 6,500-customer count — up from 5,000 the prior quarter — reflects an accelerating adoption curve for the AI Factory platform. A cost-of-ownership analysis from Signal65 and Futurum Group, published in September 2026 and commissioned by Dell, placed the Dell AI Factory 90% cheaper than public cloud over two years over two years for knowledge, sales, and software development workloads with agentic AI support — a directional benchmark rather than a universally applicable figure, given the vendor-commissioned methodology.

The ownership thesis is Dell's central competitive argument against hyperscaler AI infrastructure: enterprises that own their compute retain control over data, operations, and cost structures in ways public cloud deployments cannot replicate. As inference costs fall and token volumes climb with agentic adoption, the economics of ownership versus cloud compute shift in ownership's favor over multi-year horizons — a trajectory Dell is betting will sustain AI Factory hardware demand through the decade.

"The market is entering a new phase where customers are focused less on building AI infrastructure and more on generating measurable business outcomes," said Matt Kimball, Vice President and Principal Analyst at Moor Insights & Strategy. "The continued expansion of the Dell AI Factory with NVIDIA reflects a mature approach that combines validated solutions, ecosystem integration, automation and services, helping organizations move from pilots to production with greater confidence and less risk."

GTC Berlin's hardware lineup represents Dell completing the NVIDIA Vera Rubin portfolio picture at the infrastructure layer — from the air-cooled R8822 shipping now, through the 256-CPU M9882 rack and the Groq 3 LPU-powered XE9810 arriving in Q1 2027, to the DSX POD containerized deployment model targeting early 2027.

Whether the full portfolio can differentiate Dell against hyperscalers building their own agentic stacks, and against AI-native infrastructure providers like CoreWeave targeting the same enterprise workloads, remains the defining competitive question heading into 2027. Dell's customer growth trajectory and AMD's parallel Helion announcement together signal that the agentic infrastructure market is large and diverse enough that no single architecture wins every deployment — and that the vendor with the broadest validated portfolio across CPU orchestration, GPU inference, and LPU inference tiers is positioned to capture the most of it.

Frequently Asked Questions

What is the difference between a Groq LPU and a GPU for AI inference?

GPUs are optimized for memory capacity — they carry tens to hundreds of gigabytes of HBM (high-bandwidth DRAM) to hold large model weight sets in memory during inference. Groq 3 LPUs are optimized for memory bandwidth: each chip carries 500 MB of on-chip SRAM that delivers 150 TB/s versus H200's 3.35 TB/s — roughly 44 times the per-chip bandwidth of an H200 GPU. The tradeoff is capacity: an LPX rack's 128 GB of total SRAM cannot host frontier models exceeding that threshold, but for instruction-tuned models that fit, LPUs deliver higher throughput and — critically — deterministic inference latency that GPU HBM systems structurally cannot match. That predictability matters for production agentic deployments where consistent time-to-first-token performance defines service-level agreements.

When does the Dell PowerEdge R8822 ship?

The PowerEdge R8822 ships October 2026, making it the only system from Dell's GTC Berlin announcement with immediate availability. The PowerEdge M9882 (256-CPU disaggregation rack) and PowerEdge XE9810 (Groq 3 LPU inference system) are both due in Q1 2027. The Dell Modular Data Center with NVIDIA DSX POD is slated for early 2027. The Dell Pro Precision with GB10 64GB deskside system is set to ship October 23.

How does CPU disaggregation improve agentic AI infrastructure efficiency?

Traditional enterprise servers bundle CPU and GPU resources in fixed ratios — meaning both pools are sized together even though agentic workloads make very different simultaneous demands on each. In agentic deployments, CPU orchestration workloads (tool calls, state management, agent coordination) and GPU inference workloads are rarely peak-utilized at the same moment, so fixed bundling produces chronic underutilization of one or the other. Disaggregation separates them into independently sized rack pools, allowing each to be scaled to actual demand. AMD's Helion CPU rack blueprint, announced October 12 at the OCP Global Summit, reflects the same architectural reasoning through an open OCP design approach — confirming CPU disaggregation as a broadly recognized infrastructure response to agentic AI workload profiles, not a Dell-specific claim.

Is the Dell AI Factory more cost-effective than public cloud for AI workloads?

A September 2026 analysis from Signal65 and Futurum Group — commissioned by Dell — placed the Dell AI Factory at up to 90% cheaper than public cloud over two years for knowledge, sales, and software development workloads with agentic AI support. The vendor-commissioned methodology means assumptions favor Dell's typical deployment profile; independent total-cost-of-ownership comparisons for enterprise-owned AI infrastructure against hyperscaler AI services remain sparse. The underlying economic argument — that falling inference costs combined with rising token volumes shift the ownership calculus in favor of on-premise infrastructure over multi-year windows — is broadly accepted in industry analysis. The specific 90 percent figure should be treated as a directional benchmark rather than a universally applicable estimate.