AMD has announced new additions to its data‑centre portfolio at Advancing AI 2026, introducing both the 6th Gen EPYC 9006 Series server CPUs and the Instinct MI400 Series GPUs, aimed at next-generation AI infrastructure.

The EPYC 9006 Series is described by AMD as a workload‑optimised CPU lineup built to support agentic AI, where CPUs take on a larger share of work by running agents, coordinating tasks and managing data flows. AMD claims the new generation is designed to handle this increased load through higher core counts, expanded memory bandwidth and support for next‑generation I/O standards such as PCIe Gen 6 and MRDIMM.

The EPYC 9006 Series consists of four CPU families. The SP7 tier is the maximum‑density option, with core counts up to 256 cores and 512 threads, along with high‑frequency variants reaching up to 5GHz. SP8 models focus on efficiency, offering up to 128 core configurations designed for enterprise, edge and power‑constrained deployments. For HPC and memory‑sensitive workloads, AMD points to the SP7X variants, which it says offer triple the L3 cache per core compared to SP7. The LP tier serves as an AI host‑node CPU, using LPDDR memory and higher frequencies to keep accelerators supplied with data in multi-GPU racks.

On the GPU side, AMD introduced the Instinct MI400 Series, split between the MI455X for frontier‑scale AI deployments and the MI430X for sovereign AI and HPC. The MI455X is AMD's highest‑end accelerator for large‑model training, fine‑tuning and high‑volume inference, while the MI430X is aimed at workloads requiring numerical accuracy. AMD says the MI430X can deliver up to 288 TFLOPS of FP64 performance, alongside HBM4 memory and high‑bandwidth interconnects to support larger models.

The MI400 Series also incorporates security and system‑level features for sensitive AI and HPC environments, including SecureBoot, encrypted GPU‑to‑GPU links and hardware‑based protections. These GPUs support AMD’s ROCm software stack for AI development, which is also receiving updates this week through ROCm.ai – a new AI‑native developer environment that adds a unified CLI, AMD‑authored guidance for coding assistants and AI‑driven workload optimisation.

Finally, AMD is also expanding into physical AI and robotics, introducing new Ryzen AI Embedded X100 processors and Kria AI system‑on‑modules for autonomous machines and industrial systems. These platforms combine Zen 5 CPU cores, integrated RDNA graphics and an NPU, with shared access to memory to reduce latency in real‑time workloads.

KitGuru Says: What do you think of AMD's push into the AI space? 

AMD and Hammer Distribution are promoting a CPU‑first approach to AI infrastructure in the UK, arguing that power constraints are now the primary barrier to deployment. The companies say grid connection delays have become the biggest blocker to new data‑centre capacity, with recent Ofgem and National Energy System Operator (NESO) reforms shifting priority toward projects that are efficient and ready to build.

Under the UK's “First Ready, First Connected” rules introduced in December 2025, more than 300GW of stalled or non‑viable projects were removed from the UK grid queue. Government papers on AI Growth Zones also identify grid capacity as the largest obstacle to expansion. Hammer and AMD argue that this environment makes system level efficiency a critical factor in determining whether AI projects can proceed.

While GPUs dominate AI training, the companies highlight that CPUs govern data ingest, orchestration and inference throughput. They say mismatched hardware and unnecessary accelerators waste power in a market where energy availability is now a hard limit. AMD’s guidance indicates that EPYC processors can support CPU‑based inference for models up to 20B parameters, making them suitable for workloads such as document processing, retrieval‑augmented search and summarisation.

The companies position CPU‑led inference as a way to reduce reliance on accelerators, lower total power consumption and deploy AI on existing infrastructure without waiting for grid upgrades. The strategy aligns with new European Energy Efficiency Directive requirements, which introduce mandatory reporting for datacentre performance and place greater emphasis on “useful work per watt” as a measurable KPI.

KitGuru Says: AMD's EPYC processors have been gaining ground in the datacentre market. With new rules coming into place to govern the amount of stress AI datacentres can put on the grid, AMD wants to be at the forefront.