Skip to the stories
Windows & PowerShell

Dell Exascale Adds 32-Drive Liquid-Cooled Storage for AI Racks

Dell's SC26 updates add a 32-drive liquid-cooled Exascale chassis, a 20-drive 1U option, and a 512-core M8825, with most hardware due in 2027.

Dell Exascale Adds 32-Drive Liquid-Cooled Storage for AI Racks

Dell's SC26 announcements are mostly about storage density and management at rack scale, and most of the hardware won't ship until 2027. If you're planning AI racks for the next two budget cycles, the Exascale updates are the part to read closely. The rest is worth tracking, but the timelines matter more than the headline numbers.

Dell AI Factory security context: Dell AI Factory rack with NVIDIA NVL144 at Dell Technologies World, StorageReview photo

Two Exascale Platforms, Two Cooling Models

Dell Exascale runs four storage software stacks on common PowerEdge hardware: Lightning File System for parallel file, PowerScale for file, ObjectScale for object, and PowerFlex for block. The new options extend that model into two very different physical designs.

The liquid-cooled version is a 2-OU chassis for Open Rack designs, holding up to 32 NVMe drives. Cold plates cover every heat-producing component, and the unit has no fans. That lets storage sit in the same rack as liquid-cooled compute and networking and draw on the same coolant supply. Dell calls it the world's first direct-liquid-cooled storage for rack-scale AI. The company bases that claim on a November 2026 comparison of enterprise storage systems that ship with storage software and have been announced for commercial availability. Block support on this hardware is planned for the first half of 2027.

The air-cooled alternative is a 1U platform with up to 20 NVMe drives, aimed at sites where rack space is tight. Dell's 67% higher drive density per rack unit compares against the previous 2U Exascale hardware, which held 24 drives per node. Two of the new 1U systems put 40 drives in the same 2U footprint, and that's where the density figure comes from. Both platforms can be deployed as a PowerRack for Storage or as part of an integrated rack. Dell expects both in early 2027.

Separately, software-defined PowerScale on Dell Exascale has earned NVIDIA-Certified Storage validation at the Foundation level.

Dell AI Factory rack at Dell Technologies World, a close view of blue network cabling and server rear panels

Agentic Data Gateway and Pipeline Changes

The Dell AI Data Platform gains an Agentic Data Gateway: a customer-controlled MCP server and LLM gateway inside the Data Orchestration Engine. An agent connects once and reaches governed data, tools, and models through a single endpoint. Authorization, quotas, logging, and cost tracking are handled centrally. Dell says the gateway and everything behind it run inside the customer's environment, which matters for teams whose data can't leave their own facilities.

Dell's internal testing reported that reusing context through the gateway cut token consumption by up to 76% and raised answer accuracy from 60% to 94%. The test used a synthetic toll-plaza operations corpus, 20 questions across six workload types, and NVIDIA's Nemotron-3-Super-120B-A12B model served through NVIDIA NIM. These are vendor figures, and they haven't been independently verified. Treat them as a starting point for your own pilot, not a planning number.

Hybrid pipeline orchestration runs data processing where the data already sits, across on-premises private clouds, the Dell AI Factory with NVIDIA, non-Dell infrastructure, and public clouds. Large unstructured datasets don't have to move. Storage Intelligence and Data Intelligence report where data came from, its quality, and how it's used. Dell says the scanner reads only metadata, at about 11,000 files per second, capturing 61 metadata fields per file.

A unified pipeline now prepares tables, images, video, and documents together and routes vectors into the database a customer already runs. Dell cites up to 12x faster indexing. That figure comes from Elastic's published benchmark of Elasticsearch with NVIDIA cuVS, which compared GPU-accelerated and CPU-only vector indexing on a single AWS g6.4xlarge instance. Data Orchestration also adds integrations with LangChain, OpenAI Codex, and Unstructured.io. Dell lists these AI Data Platform updates for December 2026.

Dell Exascale hardware options: a 2-OU direct-liquid-cooled platform with up to 32 NVMe drives and a 1U air-cooled platform with up to 20

Management and Networking

OpenManage Rack Scale is a new management platform covering infrastructure from a single server up to complete Dell PowerRack systems. Dell says one instance manages up to 100,000 devices, compared with a ceiling of 8,000 for OpenManage Enterprise 4.8. That gap underpins the company's up-to-12x scale claim. It deploys on premises as a virtual machine, and operators can query health, alerts, firmware, inventory, power, and performance in natural language.

Dell's internal testing put Day-0 firmware update preparation at about one minute in Rack Scale, against roughly five hours with manual methods. The platform is available now, and Dell says OpenManage Enterprise customers have a defined migration path.

Dell Fabric Manager, also available now, provides intent-based automation, lifecycle management, and observability across network fabrics that mix network operating systems and switch silicon. It covers digital-twin design validation, zero-touch provisioning, upgrades, and drift auditing for data center, AI, HPC, and edge deployments. Dell cites up to 80% less Day-2 operational effort and 100% accurate device inventory. Those figures come from BE Networks lab testing of manual procedures against Fabric Manager automation.

Enterprise SONiC Distribution by Dell Technologies adds support for the PowerSwitch Z9964F-ON and NVIDIA's SN4700 Spectrum platform, along with L2 VXLAN multihoming and packet trimming. Dell lists that release for January 2027.

Dell PowerEdge M8825 air-cooled AMD EPYC server sled render, shown at a three-quarter angle

PowerEdge M8825 and AMD Options

The PowerEdge M8825 is a two-socket, air-cooled AMD EPYC server with up to 512 cores and 1,024 threads per node. That matches the core count of the PowerEdge R9825, which uses two 256-core EPYC Venice processors. Dell designed the M8825 for Open Rack deployments drawing 35kW to 75kW per rack. Its internal analysis puts up to 46 servers in a rack, depending on configuration, power limits, and cooling capacity. The specifications are preliminary, and Dell targets the first half of 2027.

On the GPU side, new modular configurations for the Dell AI Platform with AMD put AMD Instinct MI350P PCIe GPUs into PowerEdge XE7745 and R7725 servers. These sit alongside PowerSwitch networking, PowerScale storage, and the AI Data Platform. Validated support is also coming for AMD Enterprise AI Suite, Red Hat OpenShift Container Platform, and Red Hat AI Enterprise. Dell says a deployment can start with one compute node and grow by adding GPUs, nodes, storage, or networking.

For the largest builds, the AMD Helios Rack-scale Solution from Dell pairs 72 AMD Instinct MI455X GPUs with PowerRack networking and storage, plus Dell's deployment and lifecycle services. Dell quotes up to 2.9 exaflops for the rack, a figure from AMD's internal analysis that matches the FP4 number AMD gave for Helios at launch. Dell hasn't given availability dates for either the modular configurations or the Helios rack.

Validated Blueprints and Services

Dell says the Dell AI Factory with NVIDIA now spans more than 420 offerings and more than 6,500 customers. Arthur Lewis, president of Dell's Infrastructure Solutions Group, framed the SC26 updates around the full stack: "AI outcomes don't come from models alone."

New and updated Dell AI Ecosystem Program blueprints cover AI security with F5 and JFrog, agentic AI with DataRobot, observability with Dynatrace, and computer vision with Ipsotek. Others address data for AI with Iternal Technologies and Teradata, knowledge assistants with Aible, Atolio, and Maven AGI, a code assistant with OpenAI, and governance and orchestration with Klique by ClearML and JFrog. The DataRobot, Dynatrace, Aible, OpenAI Codex, ClearML, and JFrog items are available now. The Ipsotek blueprint is due in Q4 2026, and the F5, Teradata, Iternal, Atolio, and Maven AGI blueprints in Q1 2027.

A newly validated modular architecture pairs Dell PowerStore with PowerEdge XE9785 servers carrying NVIDIA HGX B300 GPUs and Dell-supported NVIDIA AI Enterprise. Dell positions PowerStore as block storage for stateful agent services such as databases, queues, logs, and recovery data. The servers and software are available now, and validated PowerStore follows in Q4 2026.

Dell AI Factory Engineers is a delivery model that assigns one team, an AI strategist, architect, data specialist, and engagement manager, to cover platform, workload, data, governance, and security, with knowledge transfer to the customer's staff. Dell Services for AI Data Centers, due in Q4 2026, assesses whether a facility's power, cooling, and space can support planned AI hardware, and produces facility designs to close any gap.

My read: the Exascale liquid-cooled chassis is the item to put on the 2027 roadmap now, especially if your facility is already planning for direct liquid cooling. Check the Services for AI Data Centers assessment before you commit to rack power, because the M8825's 35kW to 75kW range will push that question quickly.

Source: Harold Fritts · www.storagereview.com

Same flag