Operator: And your next question comes from the line of Antoine Chkaiban with New Street Research. Your line is open.
Antoine Chkaiban: Hi, thank you so much for taking my question. So as you can see, NVIDIA introduced new network computing capabilities with NVSwitch, performing some calculations inside the switch itself. Perhaps now is not the best time to announce new products, but I’m curious about whether this is something the broader merchant Silicon and internet ecosystem could introduce at some point.
Jayshree Ullal: Antoine, are you asking what is our new products for AI? Is that the question?
Antoine Chkaiban: No, I’m asking specifically about in-network computing capabilities. NVSwitch can do some matrix multiply and add inside the switch itself. And I was wondering if this is something that the broader merchant Silicon and internet ecosystem could introduce as well.
Jayshree Ullal: Yes. So just for everyone else’s benefit, a lot of the in-network compute is generally done as closest to the compute layer as possible, where you’re processing the GPU. So that’s a very natural place. I don’t see any reason why we could not do those functions in the network and offload the network for some of those compute functions. It would require a little more state and built-in processing power, etcetera, but it’s certainly very doable. I think it’s going to be six of one and half a dozen of the others. Some would prefer it closest to the compute layer and some would like it network wide for network scale at the network layer. So the feasibility is very much there in both cases. Thanks Antoine.
Antoine Chkaiban: Thank you.
Operator: And your next question comes from the line of James Fish with Piper Sandler. Your line is open.
James Fish: Hey, thanks for the question. Anshul, we’ll miss having you around. I echoed my sentiments there, but hope to see you soon. Jayshree, how are you guys thinking about timing of the 800 gig optics availability versus kind of use some systems? And you keep alluding to kind of next-gen product announcements for multiple quarters now not just this one but should we expect this to be more around adjacent use cases the core including AI or software kind of take us in the product road map direction if you can.
Jayshree Ullal: Yes James you might remember like deja vu we’ve had similar discussions on 400 gig too. And as you well know to build a good switching system you need an ecosystem around it whether it’s the NIC, the Optics, the Cables, the Accessories. So I do believe you’ll start seeing some early introduction of optical and switching products for 800 gig but to actually build the entire ecosystem and take advantage especially of the NIC I think will take more than a year. So I think probably more into 2025 or even 2026. That being said I think you’re going to see a lot of systems I had this discussion earlier. You’re going to see six of one and half a dozen of the other. You’re going to see a lot of systems where you can demonstrate high rating scale with 400 gig and go east west much wider and build large clusters that are in the tens of thousands.
And then once you need, once you have GPUs that source 800 gig which even some of the recent GPUs don’t then you’ll need not just higher ratings but higher performance. So I don’t see the ecosystem of 800 gig you know limiting the deployment of AI networks that’s an important thing to remember.
James Fish: Thank you.
Jayshree Ullal: Thank you James.
Operator: And your next question comes from the line of Simon Leopold with Raymond James. Your line is open.
Victor Chiu: Hi guys this is Victor Chui for Simon Leopold. Do you expect Arista to see a knock-on effect from AI networking in the front end or at the edge as customers eventually deploy more AI workloads based I’m sorry bias towards inferencing and then maybe help us understand how we might be able to size this if that’s the case.
Jayshree Ullal: Simon that’s a good question. We haven’t taken internet into consideration that’s basically production but you’re absolutely right to say as you have more back end then the back end has to connect to something which typically rather than reinventing IP and adaptive routing you would connect to the front end of your compute and storage and WAN networks. So while we do not take that into consideration in our 750 million projection in 2025 we naturally see the deployment of more back-end clusters resulting in a more uniform compute storage memory, overall front-end back-end holistic network for AI coming in the next phase. So I think it makes a lot of sense we just we but we first want to get the clusters deployed and then we’ll do the a lot of our customers are fully expecting that holistic connection.
And that’s one by the way one of the reasons they look so favorably at us. They don’t want to build this disparate silos and islands of AI clusters they really want to bring it in terms of a full uniform AI data center.
Victor Chiu: Thanks so much.
Operator: And your next question comes from the line of Meta Marshall with Morgan Stanley. Your line is open.
Meta Marshall: Great thanks maybe I’ll flip James’s question and just kind of ask you know what do you see as kind of some of the bottlenecks from going to from pilots to ultimate deployments. It sounds like it’s not necessarily 800 gig and so is it just a matter of time, are there other pieces of the ecosystem that are that need to fall into place before some of those deployments can take place. Thanks.
Jayshree Ullal: I wouldn’t call them Meta bottlenecks, I would definitely say it’s a time based and familiarity based situation. The cloud, everybody knows how to deploy that it’s sort of plug and play in some ways and but even in the cloud if you may recall there were many use cases that emerged. The first use case that’s emerging for AI networking is let’s just build the fastest training workloads and clusters and they’re looking at performance. Power is a huge consideration. The cooling of the GPUs is a huge part of it. You would be surprised to hear a lot of times it’s just waiting on the facilities and waiting for the infrastructure to be set up right. Then on the OS and operating side and Ken has been quiet here. I’d love for him to chime in.
But there’s a tremendous amount of foundational discovery that goes into what do they need to do in the cluster. Do they need to do some hashing? Do they need to do load balancing? Do they need to do the set layer two, layer three? Do they need visibility features? Do they need to connect it across the WAN or interconnect? So, and of course as you rightly pointed out there’s the whole 400, 800. But we’re seeing less of that because a lot of it is familiarity and understanding how to operate the cluster with the best job completion time and visibility, manageability, and availability of the GPUs. Nobody can tolerate downtime. Ken, I’d love to hear your point of view on this.
Ken Duda: Yes, thanks, Jayshree. Look, I think that what’s blocking people’s deployment is the availability of all the pieces. And so there’s a huge pent-up demand for this stuff, and we see these clusters getting built as fast as people can build the facilities, get the GPUs, and get the networking that they need. I think that we’re extraordinarily well-positioned here because we’ve got years of experience building scaled storage clusters, some of the world’s largest cloud players. And storage clusters are not identical to AI clusters, but they have some of the same issues with managing a massive scale backend network that needs to be properly load-balanced, needs a lot of buffer to manage bursts. And then some of the congestion management stuff we’ve done there is also useful in AI networks.
And in particular, this InfiniBand topic keeps coming up. And I’d just like to point out that Ethernet’s about 50 years old. And over those 50 years, Ethernet has come head-to-head with a bunch of technologies like Token Ring, SONNET, ATM, FIDI, HIPPY, Scalable Coherent Interconnect, Miranet. And all of these battles have one thing in common, Ethernet 1. And the reason why Ethernet 1 is because of Metcalfe’s law, that the value of a network is quadratic in the number of nodes that interconnect. And so anybody who tries to build something which is not Ethernet is starting off with a very large quadratic disadvantage. And any temporary advantage they have because of some detail of the tech cycle is going to be quickly overwhelmed by the connectivity advantage you have with Ethernet.
So I think exactly how many years it takes for InfiniBand to go and waive a fiber channel, I’m not sure. But that’s where it’s all headed.
