NEWS TAKE: China's Optical Ban, Volta's $10B Anthropic Deal, AMD's Earnings

0:00Hello everyone, I'm Austin Lyons from ChipStrat. With me is Vik Sekar from Vik's Newsletter. And we are bringing you a news flash, news take from SemiDoped. So this is a little bit different than our regularly scheduled programming. Vik and I decided that every couple weeks there's like just some really interesting stuff that we're like really itching to talk about. And it doesn't always align with our schedule of when we are normally going to record. And so we just thought it'd be fun to have these shorter kind of news take episodes that we bring to you. So, Vic, what is it that made you want to jump on and talk this week?

0:38What are we talking about today? It's this whole Reuters news article about how the Trump administration is proposing to, you know, put a ban on Chinese optical transceivers. and it stirred up quite the panic, you know, in the markets because you really can't build out an AI data center without optical transceivers. And whether people realize it or not, like a bulk of this entire optical component assembly stuff, including the pluggable modules, are actually assembled in China. So it was like quite a shock to everybody and everybody was crambling to make out, you know, what's going on here. and at that time I was like I think we should just cover this at least because it's an interesting thing to talk about because it's just like what what's going on what if we are cut off from optical transceivers tomorrow you know what would happen or why you we think it will actually happen or not happen so that's why I was like let's do a newsflash today yes totally which is good so So let's get into it.

1:46You know, I too, I woke up and I read this and my initial reaction was, you can't just cut these off from China immediately because you're not going to backfill that supply immediately from the United States, right? So there would be, if you shut it down, there would be a period of time where it's just like, all right, we just have less optical transceivers at the point in time when we're trying to build out data centers, at scales like we never have before. But yeah, walk me through your line of thinking when you read it.

2:22Essentially, the news that was reported on August 4th says the Trump administration is drafting a ban on U.S. imports of new models of Chinese data center components as it seeks to protect the infrastructure that undergirds the AI boom. Okay, they got some part of it, right? basically it really is an important part of the AI boom, right? Because without optical interconnects, you can't do any scale out or like scale across or even like hook up to the data center. You can't even do inference or training or whatever. It's a big part of the AI infrastructure. So that's pretty good. So the specific ban actually comes down to the Chinese optical transceivers, which we've spoken on this podcast.

3:10It just converts opt. light to electricity and back again because light travels much better distances at a much faster rate. So, you know, compared to electrical transceivers, optical transceivers can connect longer distances. So we've done a whole episode on this, the basics of like transceivers and networking and all that, which you should check out if you're wanting to get more into this. So, but the whole thing was that this was the dumbest thing I saw in the article. I'm going to read this out, okay? Like after this, I'm going to ask you, what do you think about this? Because it's like really dumb.

3:49It says the move not previously reported aims to prevent Chinese firms from stealing data, installing malware or disrupting service at US data centers, which house the chips to train and run AI models. Like, why do you think that, you know, banning the shipment of optical transceivers will prevent the stealing of data and installing of malware. I mean, it literally has nothing to do with it. Right, right. I too was struggling to understand this. And I kind of went down a little bit of a research rabbit hole to figure out like, is this even possible? So the question is, if a Chinese company assembles an optical transceiver, is that actually a security risk?

4:35And so, by the way, I kind of emphasize assembles because at the end of the day, what do these optical transceivers have in them? A lot of times it's a Broadcom or Marvell DSP. It's like read timers, drivers, TIAs, all these various components, a lot of which actually will come from American companies, not necessarily Chinese companies. so you know these particular transceiver companies in a light for example or eoptolink or these other ones a lot of it is is the assembly and the final kind of cabling and putting everything together and so the question is like well where are they going to put malware like what are we talking about here if if they're sourcing a lot of the components and the one thing i could find is that There's often like a little MCU, like microcontroller, that has some firmware on it.

5:29And that firmware can be written to by the vendor. But as far as I could tell, and I would, of course, I would love some transceiver person in the industry to tell us more about that. I did find some links. It was like an OIP, Optical Interconnect Group or something, that had these slides. It was in our Semi-Dope Daily. I should have pulled it up this morning. But it was like, okay, it's a normal thing for optical transceiver companies to have an MCU that has firmware on it and it can be programmed. And sometimes they make it so you couldn't just go in and read it back out yourself to be like, what did they put on here?

6:08You know, like I want to audit this. But it's not clear to me that after it gets shipped, they would have any way to get in and upgrade that firmware. So it's not like I could sell you my transceiver and then remotely connect from China and then put all this malware on it and, you know, do something. These pluggable transceivers don't have too many components. Honestly, they have an amplifier on the transmit side and they have, you know, an amplifier on the receive side. And then they have some conversion circuits. And that's about it. I mean, then you've got some optical connections like, you know, fiber attach units or whatever, if you want to connect to the optical engine.

6:52That's it. I mean, what else is here? Like in an optical engine, there's nothing. And then there's like an optic fiber that plugs into it. There's no malware or stealing data or whatever is going on here. You know what's ironic about this is that they want to make this whole ban to prevent the disruption of service at U.S. data centers. which the ban would do that exactly the ban would exactly do what you're trying to avoid yeah totally totally so you do wonder of course with things like this is it you you kind of wonder is it the government saying you know we want to continue to put pressure on china and here's yet one other angle it's not you know nvidia gpus it's somewhere else in the supply chain can we put the squeeze on, regardless of it's technically sound or not.

7:44And in fact, it's kind of a technically complicated area. So if we just tell people, hey, optical transceivers from China are bad, they could have malware. Maybe most people would just believe it at face value. So you wonder how much it actually has to even be true. Yeah, yeah. I don't know. I'm not in Washington, D.C. Oh, speaking of Washington, D.C. Yes, I'm glad you said Washington, D.C. because read this next. Transceivers definitely pose a risk, said this person who's an AI policy expert at Washington, D.C. As data center build-out scales up, you want to make sure the data center supply chain is secure from the get-go.

8:25Like, you know, again, there's nothing about security in this thing. Like this components have nothing in them that you can like hack into or like install a virus or something. You know the best analogy I have for this? It's like saying, if you hook up this USB cable to your computer, you're going to get a virus. Like it's just a USB cable. Like why would I get a virus? Exactly, you won't, you know? Right, right. Yeah, data just transmits on it. Yeah, I did find, yeah, I did find this. Okay, it's OIF, Optical Inner Networking Forum, and they had this webinar, and there are some slides around like a little MCU on these transceivers for some little bit of firmware stuff.

9:14So there is a little bit of trusting the module vendor. But again, we could put a link in the show notes. But again, it's very unclear that it is an attack vector. Yeah, actually, there are some management services that actually go on inside, you know, so that's there. Yes, there are some functions that do this kind of stuff. But one other thing is that the moment, the biggest company that was hit in China for this is Zhongji Enolite, right? They seemingly contribute to about 34 % of the global sales of transceivers, which some analysts actually revised up from what this Reuters article calls to be 27%.

10:03So down here it says InnoLight has a leading 27 % share of the global transceiver market. But some people said it's actually a little bit more than that. And there are others too in the Chinese supply chain who make these things. And if you add them all up, they constitute about 50 % of the global optical transceiver market. that is a significant amount of supply to just like ban because you can't get these things anywhere else and i have to explain what that means because the shares of like lumentum and coherent and all these things and applied optoelectronics goes up with when when this came came out but none of these companies other than maybe applied optoelectronics to some small fraction actually assembles transceivers.

10:56Yeah, I mean, Lumentum and Coherent are still going to make lasers, but they kind of make the hard to make high end content of an optical transceiver, like the 200 gig EMLs and ultra high power lasers that aren't really like accessible to too many like Chinese modules, makers or whatever. So what the Chinese module makers actually do is like they They work on this very low margin stuff, but it's a very essential portion of the market. And without them, you can't have optical interconnections in the data center, right? So I don't know how to think about this, but ultimately, if there is nobody to package these optical transceivers into a pluggable module, then I would imagine that Lumentum and Coherent would be upset, like affected by the situation because the sales would go down.

11:50So you actually want people shipping pluggable modules, right? So I'm not sure why Lumentum and Coherent went up as soon as Chinese, you know, ban was supposed to show up. Yes. Okay. So I had that take as well, which is like, okay, look, if there's a fixed supply, and then we take the Chinese supply off the table, that doesn't mean that immediately Lumentum and Coherent are going to sell more. They still have a fixed supply, right? And so it's like, okay maybe like fabronet and these other contract manufacturers could try to pull some supply online and then maybe that would be allocated to the lumensums coherence and whoever but none of that happens immediately so actually just you know just immediately the only thing that happens is that like now we've traded uh you know security concerns for actually building capacity right and So this is actually very European in some sense, which is just like, oh, no, we're scared about something.

12:49Let's just put a bunch of regulations and slow down the building of it whatsoever. You know, like because there could be a tiny attack vector that can often say doesn't really seem possible. Let's just stop building data centers because we just cut off supply to all the interconnect cables. Yeah. And there's another complicated angle to this because TerraHop, which is actually the non-China arm of InnoLight, is actually, you know, is a subsidiary of the Chinese company, InnoLight. but they are entirely operated out of Singapore and they don't have anything to do with China. Like none of the parts come from China.

13:33None of the assembly happens in China. Everything happens in like Southeast Asia. But now what would you do with those? Like how would you deal with TerraHop? Would you ban them? Would you not ban them? It's very gray area. So, so many of these things. And they also said that this ban only applies to new modules. But what does new mean? Like these things have revisions all the time. Like what is new? Like a new, slightly different revision, V2 to V3 makes it a new product? Like changing a slight orientation of a part on a module makes it a new product? Like what is it? So I don't think anybody knows, but do you actually think that this is actually going to come into play and actually last?

14:16So I don't think so. I don't see. I think enough people, I mean, come on. If you're Broadcom, Marvell, Coherent, Lumentum, and everyone in the supply chain. Oh, and by the way, if you're Microsoft Azure or OCI or anyone who's building out data centers, like everyone's going to literally say like, hold the phone. This is not right. This is not it. So I don't think, yeah, I don't think this is going to happen. And I don't think anybody is going to, any U.S. maker is going to like, I don't know, like Fabrinet or Sanmina, you know, all these companies are going to go off and immediately say, you know, let's start building capacity.

14:57Yeah, this is our business model because, you know, before you know it, like in a few months, if this reverses and then all that capacity is gone to waste, nobody going to do anything. Correct. Yes. No, you make a very good point, which is like no one can proactively take action on this because it would be manufacturers who'd have to try to ramp up supply. And that is a very expensive and kind of one way door. Right. you buy you have shells you build you buy tools you stand up a supply uh like a a fab line and then all of a sudden the government decides never mind we're not going to do this and you're like oh well now we have all this excess capacity yeah totally awesome i think we've hit on that topic pretty nicely uh i don't have anything else to say about it do you let's no okay let's carry on We have to see how it rolls out.

15:52What's your take on the whole AMD earnings scenario? They seem to have posted some really good earnings, but the stock fell and people are questioning something. What's going on there? Yeah, yeah, yeah. So, okay, AMD, they had good earnings. They are record revenue of$11.5 billion, up 50 % year over year, 13 % quarter over quarter. Their data center segment,$6.7 billion, which is up 107 % year over year. And importantly, it's now 58 % of their total revenue. So they have shifted from a sort of consumer client PC, client graphics company to a data center company, which you would hope that they would do that, right?

16:39So they're actively executing it on that. Epic CPU demand is good. um helios has a good reception and is shipping end of q3 q4 q1 right so everything at a high level looks good now the stock dropped after hours um after their earnings call and uh so then there was a lot of like oh no you know people must not what what don't investors like about amd's earnings and and there was a couple things that people were poking on um one it was okay well how much growth is actually happening with gpus um obviously cpu agentic cpu demand for epic is very strong uh helios is promising but it hasn't really shipped yet and and they you know there were some comments about like uh acceleration and like stacy rascon if you listen to the call he was poking on it saying like oh well is it actually like not accelerating as much as possible or you know because there's like quarter over quarter acceleration versus just like half over half and and you know whatever there was some surprises about um spending 800 million in capex instead of the expected maybe two or 300 million and and you know there are some legitimate on the one hand um you know some of that spend might be into their own compute infrastructure but there's some concern that some of that is actually needed to for example uh get supply for substrates and um help oems and odms like get to the finish line with these rack scale helios and stuff like that and so like the the nuanced argument there is that it's not so simple to be fabless anymore.

18:29Like you used to be a fabless company and not have to take on a lot of that, have that skin in the game for like getting allocation to certain things. But now that it's hard to get substrates, it's hard to get memory or anything like that, like that. The point is like, Hey, wait a minute. AMD is supposed to be a fabless company, but they're having to, you know, put their own money in to, to get allocation. But again, that doesn't necessarily say that AMD is an unhealthy business. That's just like a dynamic of the time. So I personally, the stock had run up ahead of the earnings and then it fell off a little bit, but I wasn't necessarily concerned about anything.

19:08I think if I was trying to push on AMD and take like a more skeptical, bearish angle and try to really push on them, the line of thinking that I came up with is after re-listening to the earnings call, there was definitely talk about demand outpacing supply. And Lisa had said something like it was unforecastable. It was just like, wow, this demand is crazy. So then I would ask, okay, what demand are we talking about? So the answer there is CPUs, right? Like agentic AI came on stronger than we expected. And we need to sell. they said they had double digit growth on cpu unit volume so this is epic server cpus and epic asps and so that's why they they were very strong is because they sold a lot more cpus and a lot higher price than they expected but also they they're leaving money on the table because they didn't have enough supply because they didn't see it coming out so strong okay good good that's great.

20:11AMD is a very strong server CPU business. But then I would say, okay, but what about GPU demand? Because we're in the GPU era, and that's where the money is made. Does your GPU demand outstrip your supply? And so the answer they would probably give is like, yes, demand is very strong, and Helios is a great product. You say, okay, but does demand outpace supply? And the answer is like, well, right now, of course it does, because we haven't shipped any, right? So it's like, okay, we're shipping a little bit at the end of Q3. There'll be a step up in Q4 and a step up in Q1. And we've seen some nice conversation and customer partnerships around Helios where, you know, OpenAI, Meta, lots of folks have stepped up Anthropic and said they're going to deploy Helios.

21:00And so, you know, you'd say, okay, great. There's strong demand for Helios and you're ramping supply. But the question is, how many quarters ahead do you think that the demand for Helios will still outstrip supply? When Helios is at full ramp, maybe it's like Q2, will there still be more new demand for it than supply? And so, you know, I think that the answer would probably still be yes, especially because AMD took a particular, they made a particular design bet as they did with the 300 series, which is let's have more memory, HBM memory capacity than the competition. And so AMD, I think would say, well, yes, there will be certain workloads where we'll still have a ton of demand because it's, you get better TCO.

21:55If for example, you can run inference of this particular model on eight AMD GPUs versus 16 NVIDIA, right? But here's where it gets really interesting. Okay, so you're betting on HBM being the differentiator here, but HBM, the price of HBM has gone crazy, right? So it's like really expensive for HBM, but it's worth it for certain workloads. But what about workloads where it's not worth it? So for example, Prefill and Decode. pre-fill is very compute bound, compute heavy. You don't actually need as much HBM. Decode, very memory bound. So that's where the HBM shines. So then the next logical question is like, okay, but you kind of only have one SKU.

22:43So, and you decided to put a ton of memory on it. So is it going to be really expensive to use Helios MMI 455 chips for pre-fill? And so on the recall actually um there was some interesting conversation where lisa said um hey we are flexible we're a chiplet based company we have the flexibility to change the compute to memory ratios and so for certain workloads if our customers if the tco of the way we've designed the chip doesn't work we could actually reduce the hbm footprint for them um okay okay this is there's this promising. So that I think what that leads to is I wouldn't, my prediction would be, I won't be surprised if AMD comes out and says, we're going to start to have more of a portfolio of offerings in the Instinct family.

23:40Obviously they already have these other ones for like, that are more focused on enterprises and air cool and stuff like that. But even with their high end and my 450 series, I wouldn't be surprised if they start to say, like, we've got some offerings that don't have as much HBM if you want to use it for pre-fill and you can use our current ones for decode and maybe end up moving toward a portfolio of offerings, which aligns with the direction the industry is heading anyway. I think the memory costs are really just putting pressure on AMD's design decision and sort of pulling them in this direction.

24:17I mean, NVIDIA had this whole CPX thing for pre-fill, but we never heard of it again. So maybe disaggregating pre-fill and decode into separate hardware SKUs wasn't the greatest idea? Well, okay, so that is a great question, a great pushback. The solution was instead of to say, let's have GPUs, some with a lot of HPM, some without, then they actually, of course, came in and did the whole SRAM thing and said, wait a minute, what if we use SRAM for decode and then we keep our HBM for the pre-fill? And so that would also continue to point out that AMD doesn't really have an SRAM offering, although they did just launch or make that announcement at Advancing AI about partnering with Cerebris to sort of fill that out.

25:07But again, I think maybe we'll see more from AMD here on flushing out? What is their strategy? Where does SRAM come in? Where does GPUs with less HBM come in? So I think there's probably going to be more coming from AMD. That's my takeaway. Okay. Yeah. That's an interesting approach. It's a good explanation, actually. I don't have anything to add to it, but we'll see how it rolls out from here. Because I wanted to go on to the next piece of interesting news, which is this company called Volta Infrastructure. Have you heard of this company? I didn't until you texted me, to be honest. Okay. So let me run through what this thing does.

25:50This is quite amazing to me, okay? Because there is this company called Volta Infrastructure, which is a London-based company. Not to be mistaken for another company of the same name, which is Singapore-based and also does something very similar. I was totally confused when I saw the website. Okay. So this company was very interesting only because this company has been around only for about six or seven months. Okay. This company, let's say, as old as our podcast as we record this right now because we started it. Okay. And they seemingly have gotten a$10 billion deal with some unnamed AI lab or something, which later, I think Bloomberg, was it?

26:40Yeah, reported that it's actually Anthropic who has put in a$10 billion computing deal with this new cloud startup that's been around only six or seven months. So you've got this like big player coming in and just like putting in a compute deal that's worth, you know, nine figures. And you've got this company that becomes instantly a multi-billion dollar valuation company. So I was like, wait, what does this company even do that? Like, how can they get like a$10 billion deal in compute in like six months? What can I learn from this for Semi-Dope podcast? Yes, yes. My original thought was like, oh man, they're the same age as Semi-Dope.

27:17Dude, we picked the wrong industry. I know. Podcasting. Why can't we start a compute? Start a NeoCloud. Yeah. Okay. Let me explain the idea of this. The fundamental idea behind this company is that they treat it like it's infrastructure. And they treat this as basically institutional grade infrastructure build out. And I had no idea what that meant. So I was like, okay, let me dig into this a little bit more. And it's interesting because whenever you borrow money to build infrastructure, which can be like roads or, you know, something that, you know, you can just get basically money out of on a regular revenue basis.

28:04Right. Just let's use the example of roads. Right. Infrastructure debt to build out roads is very different from what you would borrow to do something like a VC or like a corporate financing deal. And so the whole idea is that you treat this asset, whether it's roads or compute, as a contracted but predictable cash generator. So it generates cash. Like a toll road collects tolls. And it's going to continue to do so for decades. Like a power plant sells electricity for a very long time, often 15, 20 years. So the bet is that AI is not like a data center. You know, AI is not, you know, we call it data centers, but it's really not that because there's a lot of sharing and there's a lot of virtualization and you get part of the compute in a data center, typically in the old days.

28:59Compute is very different. Like it is a token factory. It's an AI factory. Think of it as infrastructure and that is factory. So whenever you say that I can produce revenue, like collecting tolls for a long period of time, institutional lenders, right, they tend to give this loan to you at lower interest rates, which means that they are not betting on growth or some fixed outcome, but they know that you're going to start generating revenue by collecting tolls immediately. So the cost of the capital, right, because you are getting this at a lower interest rate means that the person building out with them gets computed at a lower price somehow.

29:38Right. So that's where this whole thing is built out. So this is their whole idea behind building data centers from capital to tokens for, you know, companies like perhaps Anthropic, who's coming through the door. Sure, totally. And I know we're short on time, so I'll keep my take short. So first of all, I'll say that Jensen would be very happy to hear you talk about token factories and financing it like a utility. And that's how he thinks about it, too. So, you know, congrats on thinking the same way as Jensen. So, okay, there's one other way that I would approach this. When I started to think about this, I thought like, okay, let's say it's Anthropic.

30:22Anthropic says, oh, wow, we need a bunch of NVIDIA GPUs. And let me remind people that Anthropic traditionally used Tranium. And then they also said, we need more compute. And so they started using CPUs. And like nine months ago, there's an announcement where Microsoft, NVIDIA, and Anthropic said like, hey, there's this three-way partnership. and Anthropics going to commit to 30 billion of Azure compute plus up to one gigawatt additional initially on Grace Blackwell and Vera Rubin. And NVIDIA was investing, Microsoft investing. Okay. And so it was like, oh, wow, this is actually upside TAM for NVIDIA, which I think people forgot was one of the two biggest model lab companies wasn't actually really using NVIDIA yet.

31:03Okay. So now let's say you're Anthropic and you're like, great, we want even more NVIDIA. How do you find it? You, you, every Neo, large Neo cloud and large CSP who has NVIDIA compute has already has customers that it's allocated to. So where are you going to find it? Okay. Well, you just, maybe Jensen says, Hey, I'll allocate you some. And then you say, okay, well, I have to find power. Okay. So where do you get the power and someone to operate it? Well, basically in this case, it, um, this Volta is the balance sheet company that just started, but the operator is actually this crypto miner, like Bitdeer or something like that.

31:45And so there's an existing crypto miner that has experience operating data centers, but they're not going to be able to take on any of the balance sheet. They're not going to get that low cost of capital financing. And so I think this is also just a clever way to say, let's start a new company, clean balance sheet, and they're going to receive the money and they're going to be sort of like the parent, but actually they will pay BitDeer for power and operations. And NVIDIA will help make sure that some new GPUs get built and allocated and set up there. And then Anthropic can, to me, it feels like the fastest way for Anthropic to spin up more NVIDIA GPUs was essentially needing a clean balance sheet to borrow against someone to come in and play that role.

32:35But it is obviously, yeah, it's very financially creative. It's amazing. Yeah, it's financially creative. And I'm like, wait, what's going on? Like people just drop like 10 billion, like just like that nowadays into a company I've never even heard of. But it's fascinating. We live in fascinating times. It's great. Like I've had to work for a long time for that much money, you know? I know. But now think about if you're Volta, you're like, yeah, we've got one customer and we've got$10 billion. But on the other hand, it's like, dudes, you guys have one customer. Like, what's the long term game plan?

33:15And I'm sure the game plan is like, just keep making that customer happy as long as possible. And then maybe eventually win some other customers. But then that's essentially every other Neo cloud in the world. It's like we've got one big customer and we're hoping to differentiate. That's true. So basically two things, right? Like most startups can't even say, yay, one customer, right? Right, right. One customer is amazing. So the second thing is that if the customer is like Anthropic, there's a big customer to nail like on like level one and you've like already nailed Anthropic, everyone is going to follow suit, right?

33:51The pack of cards and the dominoes will fall. Everybody's going to throw compute money everywhere and CapEx is going to continue going to the trillions. yeah I don't know when this this is very interesting so I thought we should cover it on the podcast anyway enough of that let's see what happens it's nice to keep a tab on these kinds of creative financing deals we are not really finance guys but I love to understand the financial creativity behind how some of these things work or maybe it's just new to me either way I have fun absolutely it is going to be It is the story that's woven through the period of history we live in right now, which is very interesting technical innovations as we move to data center scale computers and then very interesting financial innovations to finance it all.

34:46And so with that, we'll call this episode a wrap. Thanks for listening, guys.