AMD launches the Instinct MI450X and claims 30% faster inference than Nvidia's Vera Rubin

Nvidia had the worst week in its modern history — its co-founder and chief executive resigned on November 12 — and AMD chose the same day to land its hardest punch yet in the AI chip race. The company introduced the Instinct MI450X, a new data-center GPU, and published benchmark numbers claiming its flagship MI455X beats Nvidia’s brand-new Vera Rubin platform by up to 30% on inference.

The announcement, published in AMD’s newsroom on November 12, covers the Instinct MI400 Series, the company’s family of data-center GPUs for AI and scientific computing. It has three parts: a new chip, new measured performance for the existing flagship, and a new entry for high-performance computing. The timing was not subtle. But the numbers deserve a closer look than the headline allows.

What AMD actually announced#

The MI450X is the most practical piece. Helios is AMD’s rack-scale system: dozens of GPUs linked together in a single rack, designed to run very large AI models. Until now, the rack’s GPU option was the MI455X, which carries 432GB of HBM4 memory per GPU. The MI450X slots into the same rack, on the same open architecture and the same ROCm software foundation, with the same core compute, memory bandwidth and scale-up and scale-out connectivity — but with 288GB of HBM4 instead.

That is a deliberately boring difference, and that is the point. Not every workload needs the biggest memory configuration, and HBM is the most expensive ingredient in a modern AI accelerator. A lower-memory option lets customers match the chip to the job — smaller models, shorter contexts, or lighter concurrency — instead of paying for headroom they will never use. AMD says the MI450X exists because customers asked for exactly that: an AI GPU with less HBM4.

The third piece of the announcement, the MI430X, targets high-performance computing and AI for science rather than commercial inference, rounding out the MI400 family across both markets.

The 30% claim, and the asterisk#

The number that will travel furthest is 21,333: tokens per second per GPU on DeepSeek-R1, the popular open model, measured on MI455X hardware inside a Helios rack using an MLPerf Inference Offline configuration. AMD says that is up to 30% higher throughput than the latest published MLPerf Inference v6.1 result for Nvidia’s Vera Rubin NVL72 in the Closed Division.

It is a real measurement, and the comparison is against a real published result — AMD’s footnotes cite the MLPerf entry and say the figures were retrieved from MLCommons on November 11, the day before the announcement. But the test was conducted by AMD Performance Labs in October 2026, and it is not an independent MLPerf-certified AMD submission. That does not make the claim false, but it means readers should understand exactly what was measured before deciding what it proves: one model, one benchmark configuration, in the vendor’s own lab, against a competitor’s published number.

AMD also claims Helios itself has up to 20% higher HBM bandwidth and higher scale-up bandwidth than Vera Rubin, based on its own calculations from published specifications. Again, the direction is plausible — the MI455X’s 432GB of HBM4 is a genuine memory advantage — but the framing is AMD’s.

None of this is unusual. Every chipmaker benchmarks its own hardware against its rival’s published numbers. What matters is what happens next: whether independent MLPerf submissions and cloud deployments reproduce the gap, and on which workloads. A 30% lead on DeepSeek-R1 inference does not automatically mean a 30% lead everywhere.

Why memory is the real story#

Strip away the benchmark duel and the MI450X tells a clearer story about where the AI hardware market is going. Inference — running models for users, rather than training them — is now the volume business, and inference is a memory game. The model has to fit, the context has to fit, and many users’ requests have to fit at once. The MI455X’s 432GB of HBM4 is roughly 1.5 times the memory of its predecessor, the MI355X, with around 2.9 times the peak memory bandwidth, according to the figures reported when the MI400 series launched earlier this month.

The MI450X exists because that much memory is not always the right purchase. It is AMD selling choice inside its own rack rather than forcing every customer onto the flagship — the same logic that has made its Epyc server CPUs a fixture in data centers: same platform, different configurations, buy what you need.

There is also an ecosystem signal. Cloud provider Vultr has already said it will offer MI455X GPUs and support Helios on its platform, with preorders open for systems starting in the fourth quarter of 2026. For AMD, the long-standing problem has never been the silicon alone; it has been whether anyone can actually rent it. A cloud launch in the same quarter as the hardware matters more than a benchmark slide.

What to watch#

First, the missing facts. AMD announced no price and no ship date for the MI450X. The chip is an option for Helios customers, not a product with a launch date, and the market’s first reaction — shares up 5.4% on Thursday — was to a benchmark claim, not to an order book.

Second, the context. Nvidia’s leadership crisis is real, but its hardware lead is not decided by a press cycle. Vera Rubin is shipping into the world’s largest AI deployments, and AMD’s claim will be judged when independent testers run the same workloads on both platforms. If the gap holds up outside AMD’s lab, it is a genuine inflection point in the accelerator market. If it narrows to a few percent on other models, the MI450X’s real contribution will be the cheaper configuration, not the headline.

Third, the timing. AMD launched the MI400 series only earlier this month, and the MI450X arrives days later — a sign the company is moving fast to press an opening while its rival is distracted. The question for the next quarter is whether customers follow: watch for Helios orders, cloud availability beyond Vultr, and independent MLPerf results. The benchmark war is entertainment; the order book is the verdict.

Sources: AMD Newsroom — AMD Advances the Instinct MI400 Series for AI (Nov. 12, 2026); Engineering.com — AMD expands Instinct MI400 Series with new GPU options (Nov. 13, 2026); DaySec — AMD Adds MI450X to Helios, Reports New MI455X Inference Results (Nov. 12, 2026). Prior coverage: Jensen Huang resigns as Nvidia CEO; CFO Colette Kress named interim chief.