AMD has started sampling its Instinct MI450 GPUs to lead customers, with production shipments still planned for the second half of 2026 through the Helios rack platform. CEO Lisa Su says demand is already stronger than AMD originally expected, with customer forecasts now above its initial 2027 plans.

The MI450 series is becoming important because AI deployments are shifting heavily toward inference. Training still matters, but large AI services need huge amounts of inference capacity to serve models at scale. AMD says the biggest planned MI450 deployments are for inference workloads.

MI450 is aimed at large scale AI racks

AMD says it is working closely with major customers through deep co engineering. OpenAI and Meta have already signed multi gigawatt agreements involving AMD AI hardware, and Anthropic has also been reported as a likely customer for the MI400 series.

The MI450 is based on AMD’s CDNA 5 architecture and will use HBM4 memory. The top MI455X is aimed at AI training and inference, while the MI430X is designed more for HPC and sovereign AI use cases.

The memory upgrade is a major part of AMD’s pitch. MI450 moves from the 288GB HBM3E capacity of MI350 to 432GB HBM4, with more than double the memory bandwidth. That matters because large models and inference systems often run into memory limits before raw compute limits.

AMD is also comparing MI450 against NVIDIA’s Vera Rubin platform. The company claims 1.5 times the memory capacity and 1.5 times the scale out bandwidth, while matching the competition in memory bandwidth, FP4 and FP8 compute, and scale up bandwidth.

The next step is MI500. Su says many customers are already engaged with AMD on the MI500 series and future AI opportunities. That is important because large AI buyers plan infrastructure years ahead, and early commitment can influence which platform they build around.

AMD will share more about Instinct GPUs, EPYC processors, the Helios rack platform, and customer engagement at its Advancing AI event in July. For now, the message is clear: MI450 is moving from roadmap to customer sampling, and AMD sees inference as the biggest growth opportunity for its next AI accelerator generation.