The deployment follows AMD's investment in Anthropic.
I've been thinking this for a while. NVIDIA's principle moat is software end ecosystem, and those are becoming less valuable by the day.
One reason I’m bullish on AMD: Advances in coding AI will naturally accelerate ROCm’s development and help close the gap with CUDA. As part of its recent investment in Anthropic, AMD has already announced that it will actively use Claude Code for chip design and software engineering.
Claude will be pointed at both ROCm code and Instinct optimization tasks, giving AMD an external AI pair programmer that already understands large-scale training runs. How much this actually narrows NVIDIA’s long-standing software lead remains to be measured once the first models ship.
The first gigawatt of MI450 racks is slated for H1 2027 in Helios systems that combine the new GPUs with Venice CPUs and Pensando networking. Any slip in that schedule would directly affect Anthropic’s expansion plans and AMD’s revenue visibility.
Positive users praise AMD leveraging Claude to speed up ROCm and challenge NVIDIA CUDA for enabling faster co-design and delivering a full stack, while negative users call ROCm unusable and fault AMD's software efforts.
No Digg Deeper questions have been answered for this story yet.
@xlr8harder So they should train great models that are co-designed with their hardware in mind Wait a sec
I've been thinking this for a while. NVIDIA's principle moat is software end ecosystem, and those are becoming less valuable by the day.
@xeophon yeah the move makes a lot of sense. more importantly, they can deliver a proved end-to-end stack to anyone who wants it.
@xlr8harder So they should train great models that are co-designed with their hardware in mind Wait a sec
@xlr8harder thats why nvidia has a vested interest to make nemotron great they wont watch their lunch getting eaten
@xeophon yeah the move makes a lot of sense. more importantly, they can deliver a proved end-to-end stack to anyone who wants it.
@xeophon competition is a wonderful thing
@xlr8harder thats why nvidia has a vested interest to make nemotron great they wont watch their lunch getting eaten
I think AMD's perceived biggest weakpoint is software ecosystem: integration complexity & expense, additional risk, efficient use of hardware, etc. Most of that is software, sustained by NVIDIA's early mover advantage and larger scale & investment. I just don't think software will remain a good differentiator going forward. NVIDIA still has other threads it can pull, though.
With these “closing the gap” trends, how can you tell when the runner-up isn’t just temporarily closing the gap that the leader won’t then extend with a new release? Like, people have expected China models to close the gap since DeepSeek, but OAI/Anthropic have kept extending that—yes even with K3, the frontier US labs are still superior. I know this is apples to oranges, so I’m not saying the LLM gap extension implies the same will happen with hardware. But I really don’t know or study hardware at all, so I’m wondering if it might follow the same trend we’ve seen with software?
@xlr8harder I meme Xi handing out Huawei Ascend cards but they can make bank selling the shovels to run optimized whale at hone.
@xlr8harder Many are saying
@xlr8harder The broader element here is that hw/sw co-design is only going to get easier and faster from here, reducing cycle times from idea to useful HW. It’s not just AMD that benefits - all the HW startups do too.
@NavinFS you can do things pretty easily today, it's mostly new model kernels and software that no one ever tested on AMD that's lacking i do agree the gap should already be tighter than it is.
@xlr8harder Won’t Nvidia move faster with AI coding as well? Or do you think there’s just a good enough point that competitors need to reach to remove most of the moat?
@xlr8harder CUDA is advancing over time so the gap never seems to close. Last I checked ROCm is still unusable and AMD still doesn't take software seriously