AMD Acquires Silo

The Six Five team discusses AMD acquires Silo.

Watch on YouTube

If you are interested in watching the full episode you can check it out here.

Disclaimer: The Six Five Webcast is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors and we ask that you do not treat us as such.

Transcript:

Daniel Newman: Acquisitions are fun. AMD doesn’t make as many as some companies, but when they do, they tend to make decisions that are substantial. And they made the acquisition of Silo this week, Pat. What’s your take?

Patrick Moorhead: Yeah, so about a $650 million acquisition, big enough that they had to float this out there, 300 employees, 130 of them are AI PhDs and here’s the TLDR. AMD is going to use Silo for two things. The first one is to optimize LLMs for AMD hardware and even machine learning, object and voice, AMD had to do optimizations of those frameworks. And for LLMs, they need to do the optimizations for those. And the second one is to really help end customers either optimize, tune or prune models for their specific use cases. And Dan, as we’ve seen just in all our travels and all the briefings and all our discussions, you can take a model and then you can try to get the results that you want through the data that you put in front of it. You can actually go in and customize the model and fine tune that model to see what you get. And then in some cases you have to create your own model if it’s very specific or it’s not one you can pull off the shelf and license from a Hugging Face.

So this is very much what NVIDIA gets a lot of credit for. When model creators for training create a new model, they do it first typically on NVIDIA, with the exception being Google, Gamma and Gemini and AMD has to optimize these models for their hardware. And then even out in the field when there’s an end customer that needs help in again optimizing, tuning and pruning, this is where the silo folks come into play. Not bad. 300 employees, $650 million, Dan.

Daniel Newman: Yeah, I mean look, it’s a big spend, it’s a big turn, but I mean it’s also a big bet. And the risk reward profile is pretty substantial, Pat. This talent, it’s not just like you can run over to the local university and find this 125 PhDs and these people that have all this experience developing, building, tuning and running the models and with AMD experience too. So to your point, a lot of the talent that’s out there is primarily skilled for NVIDIA if it exists at all. And they’re very busy. The people that are out there that know how to do this right now. This is a case of AMD trying to hack the speed. When I say hack, I mean it in a positive way, the speed to which they can develop software to run on instinct. This is the biggest challenge the market has when it comes to AMD.

It’s never is there hardware good enough? In fact, the benchmarks stand up pretty significantly and amply at this point in time and it’s actually put some pressure on NVIDIA to launch its newest models. But having said that, the software is always the part that people come back to and say it’s just not there. It’s not as complete, it’s not as ready, it’s not as utilized. There’s not as much development talent. And so AMD in this case is acquiring a lot of talent that can help them foundationally move quicker. And I mean that as a double entendre, both foundationally in their business and foundational models. And then they can actually connect the stacks and develop applications that people will use. Because that’s the other thing, Pat, we’re going to talk more about that in a minute. But the big problem of AI right now isn’t hardware.

It’s not an infrastructure problem. As much as there is some accessibility to GPUs and some network build out, we are in a massive infrastructure build out. But now people want to know how do we use this stuff? Well, AMD needs to, if they want to win, they want to win more business, they want to take a bigger share of this market, Pat, $200 billion in CapEx has been spent so far. AMD has gotten what, four or 5 billion of it this year. So where did the rest of that all gone? They want a bigger piece of that, they got to make it more compelling. This is part of that. So $650 million to get a bigger piece of what I think Lisa called out as a $400 billion opportunity by the end of the decade seems like a good bet. And she makes mostly good bets.