AMD’s AI chip gains run into Nvidia’s software moat

AMD’s MI355X is presented as a genuine break from the usual Nvidia-challenger cycle, beating Nvidia’s B200 on inference in AMD benchmarks by roughly 30% on Llama 3.1 405B. SemiAnalysis testing also found the MI355X about 40% cheaper per token on one workload, largely because of lower per-GPU rental costs rather than a universal hardware advantage.

The coming MI400 line strengthens AMD’s case on paper, with 432GB of HBM4, 19.6 TB/s bandwidth and a 2nm-class process. Its real rival will be Nvidia’s Vera Rubin, due in the second half of 2026, not the B200 systems shipping today; against Rubin’s 288GB HBM4 and ~13 TB/s bandwidth, AMD’s memory lead narrows but remains meaningful.

The bigger obstacle is software. Nvidia’s CUDA platform, launched in 2006, anchors a deep ecosystem of libraries, frameworks, tutorials and developers, making migration expensive even when AMD offers faster or cheaper chips. Nvidia still controls ~80% of the AI accelerator market, while AMD is around 5-7%.

Tools such as OpenAI’s Triton and less CUDA-dependent inference workloads could help AMD gain share, and Microsoft already runs serious inference work on AMD hardware. The likely outcome is not AMD dethroning Nvidia, but becoming a stronger second supplier in a market eager for alternatives.