Dell & AMD expand on-prem AI servers for enterprises

Dell and AMD have expanded their on-premises AI infrastructure offering with support for AMD Instinct MI350P PCIe GPUs in Dell PowerEdge servers. The update also introduces a new modular architecture for the Dell AI Platform with AMD.

Dell PowerEdge XE7745 and R7725 servers will support the AMD Instinct MI350P PCIe GPUs from summer 2026. The change is aimed at enterprises running generative and agentic AI workloads inside existing data centre infrastructure without redesigning facilities.

The hardware support centres on PCIe accelerators designed for standard air-cooled server deployments. Dell said the servers are intended to offer drop-in deployment, letting customers add the GPUs within current data centre setups rather than move to liquid cooling or change rack designs.

According to specifications released by the companies, the AMD Instinct MI350P PCIe GPU delivers up to 4,600 peak teraflops using MXFP4 precision and includes 144GB of HBM3e memory. Dell and AMD said that is the highest memory capacity currently available in a PCIe card accelerator.

The companies are targeting a range of AI uses, including small, medium and large model inference, retrieval-augmented generation pipelines and agentic AI. The software stack is designed to work with existing development tools including PyTorch, TensorFlow and vLLM, with limited code changes required.

Platform changes

Alongside the server update, Dell has revised the Dell AI Platform with AMD with a modular design intended to let customers expand systems over time. Organisations can start with configurations suited to current workloads and increase compute and GPU density later without rebuilding the whole environment.

The platform uses AMD Enterprise AI Suite, AMD ROCm and AMD Inference Server across training, fine-tuning, inference and agentic workflows. Dell described the environment as validated and designed for on-premises deployment where companies want tighter control over infrastructure and data location.

The announcement reflects continued demand among large organisations for AI systems that remain within their own data centres, even as cloud-based AI services continue to grow. Businesses in regulated sectors and those handling sensitive internal data have been weighing cost, governance and security concerns as they decide where to run newer AI workloads.

PCIe-based GPUs also offer an alternative for companies that want to add AI processing to existing servers rather than install entirely new specialist systems. By focusing on air-cooled hardware and standard server designs, Dell and AMD are seeking to appeal to buyers that want to limit changes to power, cooling and facility layouts.

Broader portfolio

For heavier AI workloads, Dell also supports PowerEdge XE9785 servers with AMD MI355X GPUs and EPYC CPUs. Those systems are positioned for foundation model development and large-scale inference, giving Dell a broader AMD-based range spanning retrofit-style deployments to more demanding installations.

The latest move adds to competition among server makers and chip suppliers for enterprise AI spending beyond cloud hyperscalers. Vendors have been trying to show that corporate customers can deploy modern AI models inside conventional IT estates, using combinations of accelerators, software frameworks and reference architectures that reduce the need for bespoke engineering.

Dell's emphasis on modular expansion also points to the uncertainty many buyers face when sizing AI infrastructure. Companies often begin with pilot projects or departmental deployments before deciding whether to expand to wider production use, making staged investment a key procurement consideration.

The open software stack around the platform is intended to support adoption by enterprises already building on widely used machine learning frameworks. That approach may matter for customers looking to avoid locking projects into a narrow set of tools while still keeping workloads on-premises.

Dell said organisations can "start with configurations that meet current workloads and expand compute and GPU density over time - without rearchitecting."