AMD
Why it matters
AI compute has been dominated by NVIDIA, which means high prices and tight supply. AMD is the most credible second source. Large buyers, including Microsoft, Meta, Oracle and OpenAI, have deployed AMD chips, and in October 2025 OpenAI signed a deal to deploy AMD GPUs at multi-gigawatt scale starting in 2026. Real competition here affects the price and availability of AI for everyone.
AMD is also relevant for local AI. Its Radeon GPUs run local models through ROCm and Vulkan, and its Ryzen AI Max laptop and mini-PC chips share up to 128 GB of memory between CPU and GPU, similar to Apple's unified memory, which makes them a popular way to run large models on a PC.
Deep Dive
AMD is a US chip company best known for its Ryzen and EPYC processors and Radeon graphics cards. Under CEO Lisa Su, it has become the main challenger to NVIDIA in data-center AI with its Instinct line of accelerators. Its flagship AI chips are known especially for carrying a lot of memory: the MI300X (2023) had 192 GB of HBM, and later models raised that to 256 GB and then 288 GB, matching NVIDIA's Blackwell Ultra.
AMD's software stack for AI is ROCm, its open-source answer to NVIDIA's CUDA.
The Memory Advantage
LLM inference is often limited by memory capacity and bandwidth, not raw compute. More HBM per chip means a large model fits on fewer GPUs, and a single node can hold bigger KV caches for long contexts. That's why AMD has pitched its chips especially for inference, where its large memory can translate directly into lower cost per token.
The Software Gap
AMD's biggest challenge has been software, not hardware. CUDA has a nearly 20-year head start, and many tools, custom kernels and tutorials assume NVIDIA. ROCm has improved substantially: PyTorch, vLLM, SGLang and major libraries support it, and AMD offers HIP, a programming interface designed to make porting CUDA code mostly mechanical. But new research techniques often appear on NVIDIA first, and setting up ROCm can still be less smooth. Closing this gap has been AMD's top priority.
Building Full Systems
Training and serving frontier models needs whole racks of GPUs connected by fast networking, not just individual chips. AMD bought Xilinx (programmable chips and networking) in 2022 and ZT Systems (data-center rack design) in 2025 to compete at that level, with rack-scale systems built around its MI400-series chips shipping from late 2026.
Product Lines at a Glance
- Instinct (MI300X, MI325X, MI350 series and beyond): data-center AI accelerators.
- EPYC: server CPUs, common in AI clusters alongside GPUs.
- Radeon: consumer and workstation GPUs, usable for local AI.
- Ryzen AI: laptop and desktop chips with built-in NPUs for on-device AI.