AMD went GA on the Instinct MI450 yesterday, the successor to the MI350 that shipped last October, and the announcement is worth reading twice because of how the pitch is arranged. The peak-throughput number is 3.1 exaflops FP8 per eight-GPU node, 1.5x MI350 and within about 15% of Nvidia’s GB300 NVL72 on a per-node basis. That is the FLOPS slide, and AMD has finally learned to put it in the middle of the deck instead of on the first page.
The first page is a workload-economics chart. AMD ran Llama-4 405B inference and Qwen-3 235B inference at production batch sizes on both MI450 and GB300 reference systems and reported delivered tokens per dollar. On both workloads, MI450 came out 32 to 41% cheaper per delivered million tokens. The company was careful to caveat that this is versus list price, that networking costs are excluded, and that ROCm 7.2 was used, which is a version most customers do not have in production yet. Every caveat is legitimate. Every caveat is also the kind of caveat you write when your slide is otherwise strong enough to lead with.
The order book is what makes the caveats survive scrutiny. Microsoft Azure took the largest committed initial allocation, per two people briefed on the arrangement, with capacity coming online at three regions by year end. Meta took its allocation into Llama post-training and Iris co-processing, which was expected. Oracle Cloud Infrastructure took a larger-than-modeled slice, which is the surprise, and is the third consecutive quarter Oracle has out-bid comparable-tier hyperscalers on AMD silicon. There is also an undisclosed Gulf sovereign-cloud customer, which almost certainly means G42 or Humain, and is the polite way of saying MI450 is going to end up somewhere the current US export-control regime does not allow GB300 to land. AMD did not name them. Nobody expected AMD to name them.
The strategic read is straightforward. Nvidia continues to own the frontier training market, which is the wrong market to fight in. The interesting AI money in 2026 is inference, which is a per-token-cost fight, which is a fight AMD can plausibly win at 30% off with an installed ROCm ecosystem that has been quietly non-terrible for the last twelve months. If MI450 lands the way the caveated slide says it does, the next twelve months of hyperscaler capex conversations start including AMD as a real second source instead of a hedging line item. Lisa Su has been waiting a long time to give this presentation.