A New Challenger in AI Compute
AMD's Instinct MI350 GPU accelerators have begun shipping to hyperscale cloud providers including Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure. The MI350 is built on TSMC's 3nm node and features 288GB of HBM3e memory per card — matching NVIDIA's H200 on memory capacity while offering meaningful cost-per-FLOP advantages for specific workloads.
Performance Reality Check
Early benchmark results from cloud provider engineering teams show the MI350 delivering approximately 85% of NVIDIA's H200 performance on transformer training workloads at approximately 60% of H200 cost per FLOP. For inference on quantized models, the gap narrows further. ROCm 7.0, AMD's software stack for AI workloads, has matured considerably in 2025-2026, addressing the primary ecosystem complaint that drove developers exclusively to NVIDIA's CUDA.
Cloud Availability
Microsoft Azure announced MI350-powered instances will enter general availability in September 2026, priced approximately 25-30% below equivalent H200 instances. This pricing signals Microsoft's intent to compete on AI compute economics, not just performance. AWS has not yet announced MI350 availability despite active deployment work observed in job postings and data center construction patterns.
Enterprise Adoption Patterns
Early enterprise adopters are clustered in inference workloads rather than training — cost-sensitive inference at scale is where AMD's price advantage creates the clearest ROI case. Training workloads, where NVIDIA's ecosystem tooling advantage is most pronounced, remain primarily on NVIDIA hardware for now.
The Bigger Picture
The AI chip market, projected at $180 billion by 2027, is entering its first genuine two-vendor competitive dynamic since the deep learning era began. While NVIDIA remains the clear leader, a credible AMD alternative creates pricing discipline that benefits all AI compute buyers.