AMD – Small Beginnings.

AMD offers a growing but still small challenge.

• AMD has followed Nvidia’s lead in teasing its roadmap early, but also launched new chips, which it claims are just as good as Nvidia’s and in some cases better, but without CUDA, it is still going to be an uphill struggle.
• However, the AI datacentre chip market is coming AMD’s way as it shifts to inference, Nvidia has already moved to lock in its position in inference as well as training.
• VivaTech 2025 was all about Nvidia, with AMD nowhere to be seen, but AMD held an event where it put its latest wares on display and stated its intent to take the fight to Nvidia.
• Top billing went to the new MI350 / MI355 chip, which is made on TSMC’s 3nm process with 185bn transistors spread across 10 chiplets.
• AMD is claiming a 4x lift in performance and that a 520bn parameter model can now be inferenced on a single chip, which is a clear sign of where AMD sees its real market opportunity.
• The MI350 will be followed up with the MI400 in 2026, which it claims will offer another large leap in performance as well as the best tokens per watt per dollar available.
• AMD claims that it gets 30% more tokens per second compared to the GB200 when running an open-source inferencing platform, but also claims that it does better (didn’t say how much) when the GB200 runs its own CUDA platform.
• This is a pretty vague claim, and I am sure that Nvidia will put up its own benchmarks, but that is not really the point.
• The main point here is that AMD’s comparisons are one generation behind, as the benchmark to beat is now the GB300, once again displaying how Nvidia’s cadence is a key weapon in its competitive arsenal.
• Despite the strengths of Nvidia’s position, it is clear that the industry would prefer a more level playing field in terms of market share which was why AMD was able to roll out a whole string of AI celebrities to tout its wares and their relationship with AMD.
• Meta, Humain, Oracle, Cohere, Red Hat, Astera Labs, Marvell, and Sam Altman from OpenAI all made personal appearances, and it is clear that for inference, the industry is looking very closely at AMD.
• It is much easier to switch away from Nvidia for inference because CUDA’s draw with developers is mainly for training, and so running an Nvidia-trained model on AMD is not very difficult, and many AI providers already do this to some degree.
• This is crucial because the rise of “reasoning” models that improve performance by consuming much more compute, mean that the AI chip market is very rapidly being driven by inference.
• RFM Research has long argued that inference would end up being more important than training, and this is already becoming a reality where AMD thinks that the AI data centre chip market could be worth as much as $500bn in a few years.  
• However, Nvidia already has a fix for this, which is one of the most recent libraries that it launched at GTC 2025 called Dynamo.
• This library can be used to optimise the whole data centre to maximise token production, which is the equivalent of making a factory run as efficiently as possible.
• As one would expect, Dynamo really only works with Nvidia chips and if it becomes a regular feature in the data centre, it will make Nvidia training and inference even more sticky than they already are.
• The net result is that AMD’s products are both high-quality and credible options for running inference in the data centre and offer Nvidia a greater challenge than it has had for a few years.
• However, with Nvidia already on the GB300 and not stopping, it will be able to make comparisons back to AMD that demonstrate that even at 75% gross margins, Nvidia products offer better value for money, especially as much of the ecosystem is already using it.
• That being said, there is a strong desire in the industry for a more balanced playing field, and I think that AMD is in a good position to take share in inference.
• When it comes to valuation, AMD is trading on 32x 2025 PER, which is pretty close to Nvidia, and there is no doubt in my mind which one of these is the lowest risk option.
• Consequently, I would still choose Nvidia over AMD any day of the week, although I still prefer my adjacencies of inference at the edge and nuclear power.