According to media reports, OpenAI is accelerating its layout of high-performance AI computing infrastructure. Recently, the head of OpenAI's infrastructure revealed that the company has been deploying AMD Helios GPU racks for three months and plans to collaborate with AMD to develop the next generation MI500 series AI chips and subsequent products. The MI500 series is expected to be officially launched in 2027, using TSMC's 2nm process technology, featuring a new CDNA6 architecture and HBM4E memory.

Memory bandwidth breaks records, performance increases a thousandfold within four years

The memory bandwidth of the MI500 series will exceed the 19.6 TB/s achieved by the MI400 series based on HBM4, further improving the efficiency of large-scale AI training and inference. At this year's CES2026, AMD CEO Lisa Su stated that from the MI300X in 2023 to the MI500 in 2027, AMD aims to achieve a thousandfold increase in AI performance within four years.

AMD also provided an optimistic market outlook: by 2030, the AI accelerator market is expected to reach $1.4 trillion, the data center CPU market will reach $220 billion, and the overall computing market is expected to exceed $2 trillion.

OpenAI's "multi-legged approach" reduces reliance on NVIDIA

This move marks that OpenAI is accelerating its diversification strategy in the AI computing supply chain. From deploying AMD Helios GPU racks three months ago to now deeply collaborating on the development of the next-generation MI500 chip, OpenAI is transforming from a simple NVIDIA GPU buyer into a computing builder with a multi-supplier strategy. This not only helps reduce dependence on a single supplier and cost pressure but also provides strong endorsement for AMD to catch up with NVIDIA in the AI accelerator market.