AMD Wins Cloud Hardware Slots
Rebellions
This shows that AMD is winning the first step of cloud adoption, hardware slots, but not yet the harder step of developer default. Oracle and Microsoft have both put AMD Instinct GPUs into real cloud products, which matters because hyperscalers want a second supplier when NVIDIA is expensive or supply constrained. But customers still build most AI workflows around CUDA libraries, profilers, and model tooling, so AMD often enters as a cheaper or higher memory option rather than the standard stack.
-
Microsoft moved early by adding Azure ND MI300X v5 instances, and described itself as the first cloud provider to bring AMD MI300X to customer AI training and inference workloads. That makes AMD part of frontline hyperscaler capacity, not just a lab experiment.
-
Oracle has gone further in product breadth. Oracle said its AMD partnership started with MI300X shapes in 2024, then expanded to general availability for MI355X, showing AMD is becoming a repeat infrastructure choice for OCI rather than a one off procurement hedge.
-
The software gap is concrete. ROCm now has official PyTorch support and published compatibility matrices, which lowers the porting burden. But CUDA still comes bundled with a much broader set of production libraries and tools through CUDA-X, which keeps model serving, optimization, and debugging easier on NVIDIA.
The next phase is a software catch up race. If ROCm keeps improving framework support while Oracle, Azure, and other clouds keep exposing AMD instances in standard workflows, AMD can become the default number two accelerator. That would turn the market from single vendor dependence into a real two supplier cloud pattern, with pricing pressure landing first on inference and memory heavy workloads.