AMD Expands Datacenter Portfolio
AMD is broadening its datacenter offerings with new EPYC processors, including the high-frequency 4564P and the cost-effective 4124P, to target a wider range of AI and enterprise workloads. Supporting this push, GIGABYTE launched new HPC servers designed to house up to eight AMD Instinct MI300X accelerators, purpose-built for LLM training and inference.
- The new EPYC 4004 series processors are designed to compete directly with Intel's Xeon E-2400 series for small business servers and the dedicated web-hosting market. AMD highlights that the 16-core EPYC 4004 can fully utilize the 16-core limit for a Windows 2022 Server base license, a key value proposition for this segment. Prices for this series start at $149, aiming to provide enterprise-level features at a more accessible price point. - AMD's Instinct MI300X accelerator is gaining significant traction with major cloud providers, including Microsoft Azure, Meta, and Oracle Cloud Infrastructure, which is using the chip to power new AI superclusters. The GIGABYTE G593 series servers can house eight of these MI300X GPUs, providing up to 1.5TB of total HBM3 memory. - The broader trend of hyperscalers developing their own custom silicon (ASICs) presents a long-term competitive dynamic for merchant silicon providers like AMD. Companies like Google, Amazon, Meta, and Microsoft are investing heavily in custom chips to optimize for specific AI workloads and reduce costs, with projections suggesting AI ASIC shipments could triple between 2024 and 2027. - While hyperscalers are building their own custom chips, the immense and immediate demand for AI compute has also led them to lease capacity from specialized "neo-cloud" providers, creating a complex "build vs. buy" ecosystem. This hybrid strategy allows hyperscalers to meet urgent customer needs while their own multi-year data center construction projects are underway. - AMD's roadmap includes the "Venice" generation of EPYC server processors, built on the Zen 6 architecture using a 2-nanometer process, expected in 2026. For AI accelerators, the upcoming MI400 series, planned for 2026, will feature the CDNA 5 architecture and HBM4 memory. - For the full year of 2025, AMD's Data Center segment accounted for over half of its total revenue for the first time, driven by high demand for Instinct MI300 series GPUs and the 5th Gen EPYC "Turin" processors. The company has secured significant deployments with major AI labs like OpenAI and Meta for training next-generation models. - The Radeon Open Compute platform (ROCm) is AMD's open-source software ecosystem designed to compete with Nvidia's CUDA. AMD is working to expand this ecosystem through partnerships and acquisitions, such as Nod.ai, to support popular AI frameworks like TensorFlow and PyTorch. - As AI workloads become more prevalent, enterprises are reconsidering a cloud-only strategy, with many planning to build more on-premises data centers or increase colocation deployments due to factors like cost predictability, data gravity, and security. This shift is driven by the high and often unpredictable costs of running large-scale AI models in the public cloud over extended periods.
