So, memory, memory has gone crazy. That's all. That's all I can say. Next topic.
Next. Wait, it feels like you've been saying this for the last several weeks.
No, it's gone crazier. I'll tell you why.
Hello listeners and welcome to another Semi Doped podcast. I'm Austin Lyons of Chip Strat and with me is Vik Sekar from Vik's Newsletter. All right, Vik. So, I thought let's start today with the Nvidia Marvell NVLink Fusion $2 billion investment topic. You know, I felt like this one didn't get much coverage in the press. Like, not many people talked about it. I felt like it's just like, oh, another $2 billion Nvidia investment, next.
Yeah, it's the Nvidia $2 billion cookie. Everybody gets a cookie.
Marvell gets a cookie. Go ahead. Get a cookie.
Everybody, you get, you get two billion and you get two billion and you get two billion.
Yeah, totally. So, let me read a little bit from the press release and then we'll try to unpack what does this mean and what are some implications. So, the press release stated from Nvidia, this partnership builds on NVIDIA NVLink Fusion, a rack-scale platform that enables customers to develop semi-custom AI infrastructure using the NVIDIA NVLink ecosystem. And it went on to say Marvell will provide custom XPUs and NVLink Fusion compatible scale-up networking, while Nvidia will provide the supporting technologies including Vera CPU, ConnectX NICs, BlueField DPUs, NVLink interconnect and Spectrum-X switches and the rack-scale AI compute. So kind of a lot there. Let me hand it over to you. What is your reaction in reading this and maybe what are some insights that you have?
So the first thing that came to mind when I heard this was why is Nvidia investing in its competitor? Basically I thought okay, custom ASICs are here to eat Nvidia's lunch and they are bad for Nvidia. So why is Nvidia giving a $2 billion cookie to Marvell so that they can do better in making custom ASICs which would hypothetically displace GPUs, right? I mean, it sounds like a weird deal to me. That was my first reaction. But later I think I thought about it and I was like okay, it's not quite as simple as that. So we'll talk about it.
Yes. Yes. Which I will say the $2 billion piece is also interesting too, which is like why obviously it's like let's partner together and I'm incentivizing you, but it's kind of interesting because it's like lots of companies can partner jointly on things and have skin in the game without a $2 billion investment. But um, tell me, okay, so tell me how did you get past the uh, wait a minute, Nvidia, you're investing in a competitor's XPUs?
Yeah. So I read through the whole announcement too and this was the XPU was the one aspect that stood out immediately. But then if you look a little bit closer, the press release also says the companies will also collaborate on silicon photonics technology, kind of broad. And if you like go down, you'll see that this is something to do with uh, advanced optical interconnect solutions and silicon photonics technology. So that is very interesting to me because you know, apart from this whole custom ASIC discussion which is like obviously the top of mind thing, you have to remember that recently Marvell bought Celestial AI which uh, is great for their like photonic interconnect uh, between dies. So they called it their photonic fabric technology, which is think about it like you know how like Intel has EMIB which is like a little substrate you just put inside that connects two chiplets together with like metal interconnects and it could connect like any part of a chiplet one to any part of chiplet two. And so this is like the same thing but in optics, like these are light links between chips. So this is very cool technology and Celestial got bought by Marvell for this for some time back. And my thought was okay, if you're going to work together on silicon photonics, are you also going to work on you know, integrating your HBM and GPU with like optical links at some point?
Like is this the kind of scale-up? Then the other question I had was like is NVLink in the future going to be optical?
That's really interesting. Yeah, it could be, right. It's just a protocol.
Yeah, it is. Right now, the physical layer NVLink happens is on copper and you know, they're just like metal wires that you know, connect everything up and there are these circuits at either side that like send the bits, you know, over this physical medium. There's no reason the physical medium cannot be light. NVLink in theory can work over optics.
Sure. Fascinating. Okay, I'm gonna... Yeah, that's really interesting. So I have another angle and I think I could tie that together too with it. So when I started reading this and thinking about it, NVLink Fusion. So of course, so the first question is going back to custom XPUs like Nvidia is basically saying, you know, GTC recently, we saw them say, hey, now we're going to have like this end-to-end solution that actually has kind of like multi-vendor silicon in some respects because it's Nvidia's Vera Rubin, but it's also the Groq, now Nvidia branded Groq LPU. And um, so they're sort of saying like if we disaggregate inference depending on the outcomes that you're trying to achieve, you could have a system that can get you very low latency using some of the Groq stuff for decode, but also like higher throughput with the Nvidia stuff. And naturally my head went into like, oh well, it kind of feels like they're trying to say to uh, Marvell's partners like you should put your XPU racks in a same data center with our Nvidia uh, GPUs and maybe our LPUs and if you use NVLink to connect it all, it can all talk together. So it felt like a step in the direction of even more heterogeneous... is that how you say the word? Heterogeneous? Heterogynous? I don't even know. More multi-vendor silicon, bringing sort of saying like okay fine, we admit like XPUs are going to be part of the solution but you should talk using NVLink and like fit nicely into us. And then I was trying to think more about like well who would want to do that and what is the benefit versus just having your like oh we already have a data center full of Trainium for example and a data center full of NVL or GB300s or whatever. And I say Trainium because Amazon had made an announcement. Well first of all, Amazon Trainium, they work with Marvell as a backend partner, so there's already a relationship between AWS and Marvell. And then on top of it, I found from December 2nd of 2025, and I barely remember this because so much has happened since then, but there was an announcement around NVLink Fusion that Amazon, I think Trainium 4 was going to be open, like was... I'm not sure, I should go back to exactly how they phrased it, but basically that they're going to use NVLink but they also talked about using UALink too.
So tying this all together, when I thought about this investment, the more I thought about it, I was like oh this kind of feels like the customer AWS Trainium already said they're going to use NVLink and Marvell needs to kind of come along and make sure that their XPUs can work with NVLink and this is maybe Nvidia bringing it all together and saying yes you are going to make this happen for your big customer Trainium who we want to be, that's Project Rainier or however you pronounce it, it's like a million um, XPUs already. So clearly like AWS is deploying things at scale. So I think if you're Nvidia, you're thinking like we can continue to sell AWS GPUs and can we play in that expanding XPU TAM as the interconnect and maybe as the supporting CPU racks, uh, supporting KV cache storage clusters, right? So I maybe it's Nvidia seeing an expanding pie and saying we want to play in AWS's pie and this is how we're helping do that, NVLink Fusion investment into Marvell, like tighter friendship there.
So what you're saying is like this is more on a need basis because uh, AWS is like look I want NVLink and I want you to do it Marvell and Marvell's like but I can't do NVLink and so Marvell goes to Nvidia and like hey I'm going to make a lot of chips for this like and they want NVLink can we work together and Nvidia is like yeah sure why not I mean you're going to give the chips to AWS anyway but we can own the rest of the platform and lock them into that. So, it's like a win-win all around, right?
That is my hypothesis. Exactly. It's win-win all the way around. And then to take it maybe one step further to what you were saying about photonics and Celestial, if I recall, Celestial, and I need to go double check on all this, but if I recall, one of their big potential customers is Amazon Trainium. I think that was part of the whole Marvell acquiring Celestial because Celestial had lined up such a customer and my hunch is that the customer said, we want to bet on you, but we don't bet on startups. So, you know, you should get acquired and then you should be acquired by Marvell and they're a partner of ours here and then we will trust that everything will come together. So, I also don't wonder if you're not wrong in that like, hey, this is NVLink and it's working closer with Marvell into Trainium and by the way, Trainium's excited about photonic fabrics in the future. And so, yeah, maybe this is a way for Nvidia to even play there in the future with NVLink over optical die-to-die connections.
Yeah, if you look at the online reaction to this piece of news, it's always like, oh my god, Jensen is a galaxy brain. Look at him. Like, he's not really giving up anything. In fact, he is secretly locking in all the customers into the interconnect fabric and drawing them in because, you know, there's only this many GPUs or ASICs you can put into a rack. But imagine when we go to like over a thousand coherent GPUs in a rack or you know, we need basically how much networking. So you know, he's betting on the substrate instead of the compute. No point in competing with ASICs instead. Like yeah, I kind of get that but I like your story better because it's more of a need-based thing. It's more like logical or realistic to me. I don't know if it's actually real but sounds nice.
Right. Well thank you. It's my speculation and I do feel like the customer pull of Amazon saying let's make this happen is an interesting angle. We'll see if it's true.
Can I add in one more benefit I think there is to this whole situation?
So Marvell now has NVLink capability which is great for them but they also have UALink capability which means like anybody who wants to make these ASICs with Marvell has both choices now. They're not locked in and Marvell wins either way because you know what ecosystem do you want, AMD, do you want to work with UALink, you want a custom ASIC sure we'll make it for you or Google, do you want something or like you say AWS wants NVLink no problem we got you know so they can do everything now which is amazing.
Sorry. Go ahead. No, I was going to say as opposed to Broadcom who also makes ASICs but like they're tied into UALink now they don't have NVLink.
Yes. Yes. And Broadcom's definitely pushing Ethernet. So I do think this is strategically sound from Marvell to say we want to be an open partner so that you don't have to get locked into Broadcom switches. So you don't have to do Ethernet if you don't want. You can do UALink but of course UALink like AMD is pushing it. I think Astera Labs is out there. Oh, Marvell, by the way, they had an announcement a while back that they are going to support UALink and they announced a custom UALink scale-up switch back in June of 2025. By the way, they bought that XCON company kind of quietly recently. Um, and $540 million and um, they have a uh, Structer S260 lane PCIe 6.0 no switch, but I think maybe it's they're going to be able to support NVLink or UALink with this silicon. So, Marvell will say like you said, you can do NVLink, you can do UALink and we've got the silicon to do the switching no matter what protocol you want.
Yeah. So, that opens up the door to like a lot of people uh, to use Marvell for whatever they want. And I like the other thing that you said which is you can deploy these ASICs as a separate rack right next to GPUs. So one of the fears that I read online was it's not like it's going to eat up what Nvidia GPUs like they're going to sell lesser, right? Why would you enable your competitor to put in ASICs now you can't sell GPUs? I don't think that's ever going to be the case. You always need GPUs. And if you see the I think the Blackwell Ultra MLPerf 6 results that recently came out, it is fantastic. Like the GPUs are fantastic. Even now they're performance-wise, they're amazing. So it's not like they're ever going to go away because you still need to train models, you still need to do stuff. And the inference is opening up all these different workloads that we didn't imagine like in the early days of AI when we were more focused on training models rather than just doing inference now that at scale like we are starting to do and you know with this agentic AI it's even worse like token explosion is like already here or is going to get even more right so in all of these situations I think Nvidia wins because GPUs are still going to be used and now if you hook it up all with NVLink, you can do this extreme co-design approach that you know always keeps talking about with even ASICs because why not?
Totally right. Why not have those workloads running really close to the orchestration CPUs and the GPUs like put it all right there? Yeah, co-design it. Totally totally agree. Very relatedly to this topic of running NVLink and UALink and Ethernet there's a startup out there Upscale AI and they've got they're building the switch Skyhammer and I've talked to them at some of the shows I think OCP and recently and anyway their pitch is like hey we're building silicon that can run either protocol on it you don't have to as a company you don't have to make an NVLink switch and a UALink switch and have team like put all the costs and design the masks and everything and you could just have our switch and there's a little bit of a trade-off which is like there's a little bit of overhead if we are in the software defining like what protocol you're running but it's and it maybe take a little bit bigger chip size because we have a little bit of overhead but it's totally worth it and therefore if you buy our switch it can run UALink it can run presumably NVLink and then it can run Ethernet and so I thought on the one hand I thought like oh this is exciting for Upscale AI because it's sort of validating like hey Marvell's saying we want to serve customers who want any of these different protocols. Of course there's questions around route to market for Upscale AI is like is Marvell going to buy you? Is Broadcom going to buy you? Are you going to be able to convince customers to buy your switch instead of just partnering with these guys? And then on top of it, bringing it all the way back to NVLink and everything, I was googling um, Upscale a little bit more to refresh on the details and I saw that Nvidia had poached their... there was an article saying Nvidia poached Upscale AI's one of their founding members and who was like a chief architect to join Nvidia's NVLink Fusion team. It's just like two months ago.
The conspiracy grows, right?
Exactly. So make of that what you will, but I do think there's a lot of interesting things happening in this space and it'll be fascinating to see in a year or two what scale-up protocols people are running and what is the switch silicon supporting it.
So tell me this, what happens if like UALink and Ethernet doesn't become the standard and NVLink because it like hooks up to everything anyway and I remember like sometime back Qualcomm also like partnered over NVLink so they can like provide their CPUs to hook up with Nvidia GPUs via NVLink. So it's like a CPU supply thing that they signed up some time back. So if it becomes so that NVLink becomes the dominant factor because they've opened it up via the fusion platform then what happens to UALink, Ethernet and what happens to Upscale AI because now we need only one platform one protocol right?
So here's my question to the proposition of could it only ever be NVLink and I think immediately I think well AMD has no incentive to support NVLink because they don't want to, Nvidia is their biggest customer. So they would hate to be selling their own racks and then giving a slice to Nvidia. Now you might argue yes but and so therefore that's why they're supporting UALink so they can have an open alternative. But what if customer pull is totally all about NVLink because customers are saying yes we're going with multi-vendor silicon but we want to reduce complexity as much as possible. So we want the scale-up to be just one standard so we don't have to mess around with it or worry about it. But then you always get into a place of whenever there's one winner then they can extract a premium and everyone gets tired of paying the Nvidia tax and then someone pops up. So I have a hard time seeing it ever settling on NVLink. But you make a fair point which is like if all these XPU people are starting to use NVLink so they can talk nicely with Nvidia. Nvidia is sort of disaggregating in their walled garden and letting other vendors come in but it's still sort of like Nvidia's garden if you will and they're kind of calling the shots and in control and that's going to make it hard for AMD to say put racks of Instinct into your big multi-vendor deployment you know.
Yeah. But you know if history has taught us anything especially when it comes to interconnect specifications there has never been like one solution like... why do we still have so many USBs right in the history of interconnect there has never been like one platform you remember XKCD comic that says like oh we have a standard specification no we have a dozen of them why do we need a 13th one and then now every now there are 13 different standard specifications something like that yeah I remember that I remember that for sure exactly.
Exactly. So it's that's we're always going to have more standards than we know what to do with.
Totally. And at the end of the day, it's never one size fits all. There's as engineers, you're always like, okay, there is something that's created that works a lot for a lot of use cases, but there's always that one use case where it's like, oh, if we strip out all this stuff, we could do this use case way faster. So it does stand to reason that there will be more than one and that there are many incentives for there to be more than one. How about that?
All right. Anything else to say on this Marvell topic?
