A while back I posted a link to an interview with Dylan Patel of SemiAnalysis talking about Nvidia’s advantages over AMD in the AI space. He was pretty brutal across the board, saying it wasn’t even that AMD’s chips weren’t as fast, it was that they didn’t have the system offerings or architecture, their software was poor and their software development efforts lacking.
Well, it turns out by 7am the following morning, Lisa Su of AMD contacted Patel and setup a series of meeting to learn more about his views. And they took it to heart. There’s new article from Patel on AMD’s AI effort now out, which you can read the first big chunk of for free:
AMD is now in a wartime stance, but there are still many battles ahead of it.
In this report, we will discuss the many positive changes AMD has made. They are on the right track but need to increase the R&D budget for GPU hours and make further investments in AI talent… We will also discuss how AMD’s product launch cadence has put their current generation products against Nvidia’s next-gen products. Launching the MI325X at the same time as B200 has led to mediocre customer interest. Customers are now comparing the 8-GPU MI355X to the rack-scale 72-GPU GB200 NVL72 solution.
We now believe that, there is renewed interest in AMD GPUs from OpenAI via Oracle and a few other major customers, but still not Microsoft, on the condition that they reach a sweet-heart pricing with AMD. We will also outline how AMD’s window for fully catching up to NVIDIA could open in H2 2026, when AMD will finally bring a rack-scale solution to production.
Of course, H2 2026 is 15 months away, and Nvidia certainly isn’t standing still, as their Rubin line will be introduced by then (Blackwell is probably finished ramping by now). I think it’ll be a tall order for AMD to innovate faster than Nvidia given Nvidia’s blistering pace, but perhaps they can have good products at lower cost and carve out more than the small niche they have today/