Kicking off this week is Dell Technologies World, the prolific OEM’s annual tech conference and trade show. While Dell has major product offerings across a multitude of client and server segments, in a sign of the times and of where Dell’s chief priorities lie, the first day of the show was all about AI, specifically the ecosystem Dell is building on top of NVIDIA’s hardware and AI Factory model. With a combination of client hardware, server hardware, software, and services, Dell is going all-in on on-premises enterprise AI as they continue to expand their offerings under the Dell AI Factory with the NVIDIA umbrella.
Dell Sovereign & On-Premises AI
With a day-one keynote already focused on AI, the central theme of Dell’s biggest announcements was around sovereign and on-premises AI. Certainly one of the biggest set of buzzwords in the AI infrastructure market at this point, Dell believes that AI has moved beyond experimentation and that clients are now ready to deploy AI hardware throughout their businesses and nations. To do this, Dell argues that many of their customers are going to need to be able to do all of that AI crunching on-premises, whether to secure valuable data against potentially being stolen, secure their ability to develop and run AI models on their own, or just in order to have more consistent and predictable AI costs.
In building out their on-premises AI infrastructure, Dell’s customers will need not only hardware but also software and services on top. This is where Dell is looking to add value to the ecosystem for customers, as well as to make its case as an OEM. In Dell’s eyes, this means giving customers access to the latest frontier AI models and making them as easy as possible to run on Dell hardware.
On the frontier model front, Dell announced and/or reaffirmed partnerships with virtually all of the major frontier model developers during today’s keynote. OpenAI, Palantir, SpaceXAI, Google, Reflection, Mistral AI, and others have announced that they will make their frontier models available on-premises, giving customers access to the latest and greatest models without giving those model owners access to their data. As Dell touts it, these partnerships make their ecosystem “second-to-none” thanks to the large number of different frontier models available through Dell (and not other OEMs). Though “where is Anthropic in this?” is left as a valid question.
Alongside access to frontier models, the second half of Dell’s big on-premises AI push will come through its Dell AI Ecosystem Program, a validation and blueprinting framework for deploying AI models and software on Dell’s AI Factory hardware. As with other OEM validation programs, Dell’s formal Ecosystem Program aims to simplify the deployment of AI software by providing enterprise customers with packages and blueprints validated for compatibility with Dell’s hardware, and by employing best practices for security and reliability. In other words, making it easier for customers to get their own AI factories up and running.
Ultimately, Dell is looking to make their combined hardware and software offerings a critical differentiator for enterprises and even nations setting up their own on-premises AI infrastructure.
Server Hardware: Dell PowerRack, PowerFlex, & PowerCool
Alongside the company’s software/services offerings, Dell is also releasing several new and updated pieces of hardware for their AI Factory ecosystem.
First and foremost is a whole new family of rack-scale products from Dell, aptly named the PowerRack family. For their rack-scale products, Dell offers both 19-inch and 21-inch racks, with a choice between air and liquid cooling based on density and cooling needs. Within that, Dell is offering multiple rack configurations, with individual designs focused not only on compute but also on networking and storage, everything in a complete rack-scale system from Dell.
For customers who would normally need to purchase a large amount of hardware and then set up their own racks, Dell will be pushing those customers towards their PowerRack setups as a more simplified, turn-key option that is essentially sold as a single system. According to Dell, a PowerRack can be set up and deployed in as little as 6 hours, which is a fraction of the time it would take to build a bespoke rack. At STH, we toured how Dell does this with a look inside Dell’s AI Factory, where these integrated racks are made.
PowerRack will not replace Dell’s existing rack-scale systems, but it will supplant it in a big way. That is especially true on the networking side, where Dell has not previously offered networking gear by the rack. The PowerRack networking configuration will employ eight of Dell’s PowerSwitch SN6600-LD, a branded version of NVIDIA’s Spectrum SN6600-LD Ethernet switch. The networking PowerRack will offer over 800Tb/sec of switching capacity for east-west traffic.
Meanwhile, on the storage front, Dell is expanding their offerings in multiple ways. Besides offering a whole rack of storage with the PowerRack for Exascale Storage, a 10+ PB storage rack that can house SSDs as large as 245TB each, the company is also improving the flexibility of the underlying storage hardware by adding block storage to the array of data types that can be stored on the racks. Dell dubs this PowerFlex, and it joins the current support for file storage (PowerScale) and object storage (ObjectScale).
Finally, to help keep these dense racks cool, Dell is also introducing a new rackmount cooling distribution unit for liquid cooling, the PowerCool CDU C7000. The C7000 is designed to support racks for NVIDIA’s forthcoming Vera Rubin NVL72 platform, which is going to set new records for GPU/AI performance – and for power density. Of note, the 4U CDU is designed to cool upwards of 220kW of heat, and Dell says that the CDU can handle relatively warm water intake temperatures, supporting temperatures as high as 40C.
Client Platforms: Dell Deskside Agentic AI
Alongside servers, Dell is making an equally big play for on-premises inference at the client level with a platform they are calling the Dell Deskside Agentic AI. A combination of hardware and software, Dell’s client AI offerings are heavily leveraging NVIDIA’s range of client hardware and coupling that with the necessary (and newest) software stacks to support agentic AI.
On the hardware front, Dell is now offering a complete range of NVIDIA-powered systems, both big and small. These include Dell’s SFF GB10-based system, the Dell Pro Max with GB10, as well as various Precision tower PCs that can be loaded up with discrete NVIDIA video cards. Meanwhile, at the high-end of Dell’s hardware offerings, their first GB300-based DGX Station system, the Dell Pro Max with GB300, recently started shipping. The GB300 represents the peak of workstation-class hardware – and from a power consumption standpoint, it is as powerful a system as Dell can ship, as anything else would be too much for a standard North American power circuit.
All of this hardware, in turn, will be leveraged to run agentic AI systems locally – not just the inference powering their AI reasoning, but also the execution of the agents. The software side of this is largely Dell’s implementation of NVIDIA’s software stack, with Dell offering a stack based around NVIDIA’s NemoClaw, which is their OpenShell runtime coupled with Nemotron models. NVIDIA has been quick to embrace agentic AI following the explosion in popularity of OpenClaw, and, by extension, so has Dell.
As with Dell’s server announcements today, the Dell Deskside Agentic AI is all about giving customers the tools needed to run AI workloads on-premises, with Dell laying their security software and services on top of the NVIDIA AI stack. Ultimately, Dell sees their agentic AI stack as being used to drive a multitude of client AI roles, including coding, research, and other assistant functions, all without the security and legal risks of using private data in the cloud.
Finally, Dell will be hitting the ground running with its client efforts. With the hardware and software already shipping, Dell can offer its Dell Deskside Agentic AI stack immediately.
