Could AMD Finally Challenge Nvidia With Its MI400 AI Chips?

Advanced Micro Devices (AMD -0.20%) has built a sizable AI accelerator business over the past few years, but it’s a drop in the bucket compared with market leader Nvidia (NVDA 0.87%). While Nvidia’s data center segment generated more than $39 billion in revenue during the most recent quarter alone, AMD managed just $5 billion in AI accelerator revenue in all of 2024.

AMD’s latest Instinct MI350X and MI355X graphics processing units (GPUs) are powerful, delivering four times the AI compute performance and 35 times the AI inferencing performance compared to the company’s last-generation products. However, specifications and performance are in line with Nvidia’s latest Blackwell-based GPUs. AMD’s chips feature 288 GB of HBM3E memory and 8 terabytes per second (TB/s) of memory bandwidth, the same as Nvidia’s GB300 Superchip. Raw compute performance across different types of operations is also similar to Nvidia’s latest. While AMD’s MI350X and MI355X are capable AI accelerators, they fail to leapfrog Nvidia.

Upping the ante

In 2026, AMD will release its next-generation MI400 series AI accelerators. The company claims that the flagship MI400X will be 10 times more powerful than the MI300X, an enormous leap in just a few years. The chips will feature up to 432 GB of HBM4 memory, 19.6 TB/s memory bandwidth, and significant increases in raw computing performance.

Of course, Nvidia won’t be sitting still. Nvidia’s Vera Rubin chips are expected to ship in the second half of 2026, bringing significant performance gains over Blackwell. The initial Rubin AI accelerator will keep memory per GPU at 288 GB but up the bandwidth to 13 TB/s, while Rubin Ultra in the second half of 2027 pile on additional memory and bandwidth. Rubin is expected to more than triple compute performance over its predecessor.

While companies’ claims about the performance of future products should be taken with a grain of salt, AMD’s MI400X looks like it will be a solid contender when it launches in 2026. It should beat Nvidia’s current generation of products handily in terms of performance, although Nvidia’s next-gen Rubin chips may ultimately win out. One advantage AMD will have is memory, with the MI400X set to offer more capacity and bandwidth than Rubin. For certain types of workloads, that will be appealing.

In addition to MI400 family, AMD is planning to launch a new rack-scale AI solution called Helios in 2026. Helios will feature up to 72 MI400 GPUs, Venice EPYC server CPUs with up to 256 cores, and AMD’s next-generation Vulcano AI network interface card to enable fast data transfer in high-density clusters. Rack-scale solutions are becoming more popular among AI infrastructure companies as they face challenges building out large enough clusters to handle the latest AI models, and Helios should be a compelling option when it launches.

Nvidia will still be tough to beat

While AMD will have far more powerful AI accelerators and a new rack-scale AI solution in 2026, Nvidia will probably remain the overwhelming market leader. Not only is Nvidia’s hardware powerful, but the company’s CUDA software ecosystem provides a critical competitive advantage. AMD is attempting to build out its own ecosystem with ROCm, and the latest version brings increased performance and new features to the platform. However, AMD continues to fight an uphill battle.

AMD sees the market for AI chips reaching $500 billion by 2028. With annual AI accelerator revenue at just $5 billion for 2024, the company doesn’t need to beat Nvidia to vastly grow its AI chip business over the next few years. More powerful AI chips will help, as will a more fleshed-out rack-scale solution that should appeal to the largest customers of AI accelerators.

With the MI400 series and Helios slated for 2026, AMD is positioned to grow its AI-related revenue substantially over the next few years. But just like in the gaming GPU business, the company is almost certain to remain in a distant second place behind Nvidia.