Picture a data-centre floor late at night. A rack of machines the size of a wardrobe hums under cold fluorescent light, and an engineer watches a single number climb on a monitor. Tokens per second.
That number has become the scoreboard of the AI age, and on November 12, AMD decided it was time to post a bigger one.
The chipmaker used its newsroom to fire two shots at once. The first was a fresh benchmark result for its flagship AI accelerator. The second was a brand-new chip, the AMD MI450X, built to give the customers buying its Helios rack systems a cheaper way in.
Together, they amount to the clearest statement yet that AMD believes it can take Nvidia’s crown in AI inference.
What AMD announced on November 12
AMD’s announcement, published on November 12, 2026, covered its Instinct MI400 Series — the family of data-centre GPUs it builds for AI and scientific computing. It had three parts, and each one matters for a different reason.
First came the headline number. AMD says its MI455X GPU, running inside its Helios rack-scale system, processed 21,333 tokens per second per GPU on DeepSeek-R1, a popular open AI model. The test used an MLPerf Inference Offline configuration, the standard setup the industry uses to compare inference speed.
Second came the comparison AMD wanted everyone to repeat. That 21,333 figure is, according to AMD, up to 30 percent higher throughput than the latest published MLPerf result for Nvidia’s Vera Rubin NVL72 system.
Vera Rubin is Nvidia’s newest platform, so beating it — even on one benchmark — is the kind of claim that makes the whole industry stop and look.
Third came the new hardware itself. The AMD MI450X joins the MI400 family as a second GPU option for Helios customers, alongside the MI455X that has powered the system since July. AMD also shared updates on the MI430X, a chip aimed at scientific high-performance computing that is still expected in 2027.
AMD MI450X: the same muscle, a different memory
The MI450X is not a faster chip than the MI455X. It is a more flexible one. AMD says it offers 288 GB of HBM4 memory per GPU while delivering the same core compute, the same memory bandwidth, and the same scale-up and scale-out connectivity as its bigger sibling.
That sameness is the point. Until now, a customer building a Helios system had one GPU to choose from. Now they have two, and the choice comes down to memory.
The MI455X, with its larger 432 GB of HBM4, keeps the extra headroom for bigger models, longer contexts, or heavier concurrency. The MI450X trades some of that memory away for a system that fits a different budget and a different workload.
Both chips run on the same open rack-scale architecture and the same ROCm software foundation, so a customer can mix and match without rebuilding its software stack. For a data-centre buyer, that kind of choice is worth almost as much as raw speed. It means you buy the memory you actually need, and nothing more.
Memory has quietly become the deciding factor in AI hardware. A model’s weights have to live somewhere, and the longer the conversation or the document an AI is working through, the more memory it needs.
A chip with too little memory forces the model to shuffle data in and out, and every shuffle costs time.
The 30 percent claim against Nvidia Vera Rubin
The benchmark number deserves a closer look, because it is the part of the announcement that will travel furthest. AMD’s testing was carried out by AMD Performance Labs in October 2026, and the comparison is against Nvidia’s published MLPerf Inference v6.1 result for the Vera Rubin NVL72, retrieved from MLCommons on November 11.
That makes it an AMD-tested result, not an independent MLPerf-certified submission. It does not make the claim false, but it does mean readers should understand exactly what was measured before deciding what it proves.
The result covers one model, DeepSeek-R1, in one configuration. It is not a finding that AMD’s chips are faster on every workload.
AMD also claims that Helios extends its infrastructure lead with up to 20 percent higher HBM bandwidth and higher scale-up bandwidth than Nvidia Vera Rubin. Again, these are AMD’s own calculations, based on published memory capacity and bandwidth specifications rather than a head-to-head test in an independent lab.
Even with those caveats, the number is a serious one. DeepSeek-R1 is a demanding, widely used open model, and 21,333 tokens per second per GPU is the kind of throughput that lets a cloud provider serve more users from the same hardware.
If the figure holds up in independent testing, it would be a genuine challenge to Nvidia’s newest platform.
The contest is not only AMD versus Nvidia, either. Google’s own TPU 8tx and 8i Max chips, announced days earlier, are chasing the same inference customers — and every vendor is now racing to prove its hardware runs today’s open models fastest.
Why the timing could not be better for AMD
AMD could hardly have picked a better week to make its case. Nvidia, the company it is chasing, is in the middle of the most turbulent period in its history.
Jensen Huang, the co-founder who led Nvidia for 33 years, resigned as chief executive on November 11, and Colette Kress has stepped in as interim chief executive while the board searches for a permanent successor.
The resignation followed weeks of turmoil around Huang, including a board investigation, a shareholder lawsuit, and a sharp fall in the company’s shares. None of that changes what Nvidia’s chips can do. But it does change the conversation in the boardrooms where billion-dollar hardware orders are decided, and AMD knows it.
The chipmaker has been building toward this moment all year. Just two days earlier, it unveiled its sixth-generation EPYC Venice server CPUs, with a flagship chip carrying 256 cores for the AI agent era.
The MI450X announcement keeps the pressure on, turning a product launch into a sustained campaign against a rival in disarray.
What AMD did not say
For all its confidence, AMD left two important questions unanswered. The announcement gives no price for the MI450X, and no shipping date.
It introduces the chip as an option for Helios customers, but it does not say when systems will be available or what they will cost.
That matters because price is where AMD has historically won its biggest victories. Its EPYC server chips took market share from Intel partly by offering more cores for less money, and its Instinct GPUs will need a similar story to win over buyers who already run Nvidia hardware and Nvidia software.
The other gap is independent verification. AMD’s benchmark claim will carry far more weight once an independent lab, or an MLPerf submission, confirms it. Until then, it sits in the same category as every vendor benchmark: impressive, plausible, and worth checking.
What happens next for the AMD MI450X
Helios itself is already in production, according to AMD, with 72 MI455X GPUs in each rack-scale system. The MI450X now gives customers a second configuration to choose from, and the first real test will be whether big cloud providers and AI labs actually buy it in volume.
Watch for three things in the weeks ahead. First, whether AMD files an independent MLPerf submission that backs up the 21,333 tokens-per-second figure.
Second, whether any major customer publicly names the MI450X in a deployment. And third, how Nvidia responds — whether with its own numbers, a price move, or a new product of its own.
The AI chip race has never been a one-company story, even when it looked like one. AMD’s message on November 12 was simple: the scoreboard is open again, and the next number posted may not belong to Nvidia.
