Semi Doped
The business and technology of semiconductors. Alpha for engineers and investors alike.
Semi Doped
News Take: Hyperscaler CDS, SK Hynix Earnings, China's DUV
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Austin Lyons and Vik Sekar break down the market's recent volatility. They discuss the rising fear around hyperscaler debt, visible in credit default swap (CDS) premiums, and the counter-argument that GPU ROI is actually understated. They then analyze the paradox of SK Hynix's massive stock drop despite record growth, and explain why China's new DUV lithography tool makes 3D chip stacking essential to their strategy.
Key Takeaways:
* Hyperscaler debt fears are quantified by rising credit default swap (CDS) premiums, signaling investor nervousness that the AI CapEx firehose is shifting from cash flow to riskier debt.
* The counter-take on debt is that hyperscalers are under-earning on GPU assets — as long-term contracts re-price to spot rates 2x higher, the hardware's ROI will justify the financing.
* SK Hynix's 20% stock drop despite 257% YoY revenue growth shows a market priced for perfection, where a small miss against consensus triggers a panic disconnected from fundamentals.
* The memory market paradox: commodity DRAM is currently more profitable than HBM because HBM's production is 3x less bit-efficient per wafer, consuming scarce supply while DRAM spot prices soar.
* China's reported immersion DUV breakthrough is equivalent to ASML's late 2000s-era technology, a step toward self-sufficiency but not a leap to the leading edge.
Chapters:
0:00 The Market Is Freaking Out
0:21 Hyperscaler Debt and Credit Default Swaps
3:56 The Counter-Take: Under-Earning on GPUs
6:18 SK Hynix: Record Growth, Market Carnage
8:10 The HBM vs. DRAM Profitability Paradox
10:35 Leverage, Retail, and Panic Selling
12:05 China's Immersion DUV Breakthrough
13:28 The Limits of DUV: 7nm and Multi-Patterning
14:14 Logic Folding and 3D Stacking
16:12 A Long Journey to Self-Sufficiency
17:05 Surviving Semiconductor Cyclicality
Follow Semi Doped:
Get more of Austin and Vik daily, free: https://daily.semidoped.com/
Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/chipstrat
Follow Vik:
Newsletter: https://www.viksnewsletter.com/
X: https://x.com/vikramskr
So some major things have happened in the market uh this week. I think we should talk about it. Um have you kept a tab on what's going on?
SPEAKER_01Uh everything's going on. The sky is falling. And overnight, and it doesn't stop either because overnight in Korea things like freak out. And so that's just like constant news.
SPEAKER_00Yeah, it's insane. Okay, what's what's the first thing you've heard?
SPEAKER_01Okay, let's talk uh credit default swaps. This is this is interesting, and this is related. I think Gavin Baker had a good take as well, where he was kind of downplaying a little bit. But but um the topic here is the idea that um right now there are a lot of all the hyperscalers, you know, traditionally, yes, OpenAI and Anthropic are the ones using all the compute, but it's the hyperscalers providing the compute, and the hyperscalers are the ones that have to go buy billions and billions and billions of dollars of racks from NVIDIA, basically. Um and historically, they've been funding this out of cash flow, out of their free cash flow. But as the um quarterly CapEx investments get bigger and bigger and bigger, it's getting to the point where these hyperscalers are starting to exhaust all of their free cash flow and they're starting to do things like take on debt or do these off-balance sheet special vehicle things. Like uh Meta had something with some company, I think it was called like Blue Owl or something. Um, anyway, so now all of a sudden the debt market, at first, um, debt investors were like, cool, we like hyperscalers, like we feel good about taking on their debt, but there's been such an influx and projected influx of um these hyperscalers coming to the debt markets to to take on debt to fund this, that um debt investors are starting to get a little cautious. And you see this showing up in this thing called the credit default swap, which is essentially like you can think of it kind of like as insurance. If you think that the debt has a chance of um ever not being repaid, you've got this credit default swap thing. And if uh as there's like a higher premium on that, it sort of suggests that debt investors are getting more and more nervous. And so it's kind of like more expensive to uh basically finance this. And and so that's kind of the the first thing that uh people are starting to freak out a little bit about of saying, like, whoa, there's a lot of debt. Maybe everyone's a little unsure that this is actually gonna get paid.
SPEAKER_00Yeah. Uh credit default swaps, if uh folks aren't familiar with it. I mean, I'm no finance person, uh, so I had to kind of look it up to make sure that I understand it right. So it's basic, basically a form of insurance on a company's debt. So if you see a company is taking on way too much debt, uh what investors do is they pay a premium. It's like getting insurance for yourself, right? Like, hey, you you're gonna fall sick maybe, so you get insurance so that you can pay for it when it happens. The same thing with like investing. When investors are investing in a company and they see that the company is taking on too much risk, they feel like you know the company is eating too many burgers or unhealthy foods, and they're probably gonna get sick. And they're like, okay, I'm gonna get this insurance so that when the company defaults and doesn't pay back its debts, I get some money back. So that's those are like credit default swaps. And uh what's happening, like you were saying, is that yeah, a lot of a lot of credit default swaps are uh uh trading at like a higher premium, the spread has gotten higher. And so everybody's nervous about the circular financing that Nvidia is paying money to uh you know people like hyperscalers and near clouds or whatever to to buy their product, right? Like to buy NVIDIA chips. And so these kinds of like circular deals in AI have been around now at least for a year, and we've spoken about them. It's making a lot of investors nervous. So this is, I think, the one of the worries and uh fears of you know fueling the sell-off this week.
SPEAKER_01Yes, and um so I thought Gavin Baker did put a nice um take, kind of pushing back on this on X, where he basically said, Hey, here's some things that aren't being considered. Yes, um, the free cash flow of these hyperscalers is like headed towards zero, which suggests there'd be more debt. And then you've got the circular financing thing where everyone's like uh just trying to get concerned. He said, but here's here's two interesting data points. One, uh the spot price for GPU rentals is much higher than these contracted rates. Um for example, he said at times it's even like 2x higher. And therefore, he thinks hyperscalers are actually under-earning for the GPUs that they've rented out because they rented out these GPUs a lot of times on like three-year contracts, long-term contracts. So as those contracts begin to roll over, they will be able to charge a lot more. So, you know, just because you're only getting $3 an hour per GPU, let's say, as those start to roll over, maybe they're gonna renew them for five or six dollars an hour or whatever. Um, I'm just making up the numbers here, but you get the point. And then secondly, he said, you know, historically, the big consumers of that compute is ultimately Anthropic and OpenAI. Early on, it was like all for training. Obviously, now we're into the um era where they're using a lot of the compute for inference, and this inference has very high margins. And especially now that they both companies have figured out this is really good for open AI. Anthropic came first and figured out like, hey, there's a ton of money to be made from enterprises and from coding, you know, and fable and charging API token costs and um open AI, which had been just focused on consumers who will not pay anything, or maybe they'll pay $20 a month, with Codex, they're catching up, they're chasing after these customers too, and realizing, like, okay, everyone's like, we can make a ton of money on inference, and we have very high margins. And so Gavin was saying, like, don't forget these end customers are now incredibly profitable. Well, you okay, they are making more than the market probably appreciates, and therefore they uh will be good buyers and will have the their own free cash flow to be able to invest in paying for more of this GPU. So a little bit concerned of like that the end users used to not be able to afford this, and that's why Nvidia and whoever had to come in and backstop it, but now actually there's ROI on the investments.
SPEAKER_00Yeah, that's a good that's an interesting take. Let's see how it pans out. Totally. Next, SK Heinex. Um, this week in SK Heinrich land, uh, they made a lot of money, and I have it in my notes here. They they have uh 257% year over year revenue and uh more than a six-fold uh jump um in operating profit, too, I believe. And then there's like a 76% operating margin. All great, right? Like it's amazing. Like, why would you but it apparently came in below consensus? So everybody believed that the uh the the outputs would be like in terms of revenue would be even higher, and they came in a bit short. So everybody freaked out, and it went terribly bad for SK Heinnix investors, and the sole listed shares dropped by like 20%, which is the one of the worst single-day uh stock falls on record. It dragged down like all Korean stocks, it dropped the cost P index. Um, and so yeah, it's and they had to like limit this kind of leverage trading only to like institutional investors and not retail investors. So it's just like carnage, right? Like the the memory market, like all of a sudden, is carnage. Which is to me a little bit like I don't know why, but I feel like it's an overreaction. Fundamentally, do you do you think the memory uh requirement has changed, like from a technical point of view?
SPEAKER_01No, no, fundamentally, there's so much need for HBM, there's so much need for DRAM from CPUs to GPUs, right? Bigger models, more context length, like none of this is going away. It's truly useful. Agentic AI, of course, is now you've got tens of agents and they're just spinning up all the things, doing all the things. Like this is not this is not going away fundamentally, nothing has changed.
SPEAKER_00Yeah, so that's the whole point. It's interesting you mentioned HBM because like DRAM is making more money for these companies than HBM because the bit output of HBM is like three times worse. You need three wafers of DRAM to make like the equivalent number of bits uh in HBM. So it takes sucks up more wafer supply, and the pricing, the spot pricing of DRAM has gone insane. And a lot of companies who are just selling DRAM and not HBM have ended up making a lot more money. And so this is this is the whole thing, right? Like um selling DRAM uh is is profitable, but HBM isn't. But then the demand for HBM isn't going away anytime soon because literally there is no alternative to provide the kind of performance as a memory tier. Uh there are alternatives that people are looking at when we always hear about offloading to KV cache storage and doing different techniques like turboquant or Kimmies, uh Delta attention, you know, we have these algorithmic approaches. But all of them fundamentally rely on HBM to give you the tokens per second throughput, because SRAM isn't just enough capacity. So it's sitting in that sweet spot, that sweet tier, and it's very hard to dislodge HBM. Now, all of a sudden, because you fell short a few billions or whatever, uh like everything is wrecked.
SPEAKER_01Yes, yes. So to your point, there's a memory hierarchy. Of course, SRAM is the fastest, but it it's transistors, it takes up a lot of area, it's very expensive. HBM is that nice tier right beneath it, where it is DRAM, but it's 3D stack DRAM, and it's super close to the chip, and it gives you that bandwidth you need. Definitely the sweet spot. That's not going away. You can argue that that is less of a um commodity too going forward because we'll have these custom-based dies. Um DRAM underneath, lots of capacity, less bandwidth, more of a commodity. And so historically, you know, that's not where the value capture has been. But because to your point of the whole, like we can only make so many DRAM wafers and when we stack them for HBM, it's less bit efficient. Um and then at the same time, CPUs and agentic AI making uh d just D L P D DR, DDR, whatever, like even more um in demand. It's this weird imbalance where D like normal DRAMs capturing all the value. But you can argue that in the long run, HBM in that sweet spot will still capture more value. Um, but I guess last thing I'll say on this is you know, you've got levered investors, you've got Korean investors. I mean, think about it, and you know this well because you're in India. Like in the United States, I can invest in all these companies, I can invest in Micron. SK Heinex is new via the ADR, but if you're in other countries, um maybe locally, like you just invest in the local industry. If you're in Taiwan, maybe you invest in TSMC. If you're in Korea, maybe you invest in SK Heinex. So I think a lot of people, of course, and of course, because memory has gone to the moon, you've got a lot of retail traders, everyone's in. Some people are highly levered, you've got funds in. So I think that you know, you miss your mark a tiny bit, people start to freak out. Of course, there's other things in macro things going on too that are causing people to freak out. And especially if you're levered, a little miss, a little change in the stock, and suddenly, you know, you're you were up 500%, and now of a sudden you're only up 300%, and you're like, I should just sell and get out and take, you know, take what I can. So I do think there's a lot of deleveraging and a lot of selling going on, not based on the fundamentals, more based on fear and volatility.
SPEAKER_00Yes, true, true. Everybody's looking for the memory top, and the we we kind of expected this. The first earnings call where somebody says, Oh yeah, I'm shot, that's going to be like trigger, you know, for the panic, right? We kind of expected this, but now it's happening out in real time, really. Totally. Totally. Yeah.
SPEAKER_01And of course, that's that's a I can't I can't say memory without saying consumers. So there's pressures there too, which I think are freaking people out. So but we'll carry on.
SPEAKER_00Yeah, that's true. Um the next thing we should uh probably look at is Chinese announcement of immersion DUV. Uh, what do you think? You think they've uh broken through the lithography industry now and can make chips? What's your take?
SPEAKER_01I I I thought I thought you had a good take there, which was like, hey, this is like uh ASML tool circa 2008 or 2003 or something. And I think it was a good reminder because just like with any startup, you know, if someone in a lab says, hey, I built this thing, it doesn't mean they're taking down a company instantly. It doesn't even mean that it's necessarily competitive with where the frontier is for any given company. And so, you know, yes, of course China's going to work on DUV and of course they're gonna try to work toward EUV, but just because they can ship something that works doesn't just mean that ASML is dead and the monopoly is over and so on and so forth. But yeah, take us into the technology a little bit.
SPEAKER_00Yeah, so uh we did a deep dive on this podcast about lithography. So if anybody is like wondering what's immersion versus not immersion DUV, we have like a nice full podcast with pictures about this stuff. But essentially immersion DUV was what was used like in the early 2010s to make a whole lot of chips below the 45 nanometer node. I think it could pattern 28 nanometer node like in a single shot. Uh now if you want to go to 7 nanometers, we have to do multi-patterning, which is another uh topic we've discussed on the lithography episode. So check that out. But essentially what you can do is you can only get to like 7 nanometers, okay? Uh you can't go to 2 nanometers, and that doesn't automatically mean that uh China has like advanced or whatever. So take everything with a grain of salt. It's a good milestone if it's true. I mean, if this thing actually works, I'd like to see chips coming out of it. So I'm taking everything with like a like a a lot of hesitation about what has already been done, where their capabilities are, because none of this is really public and a lot of it is just announcements without any proof. So I'm always skeptical. But when you can make a 7 nanometer node, you know, they have to do their logic folding and stacking in 3D chips to make anything useful out of this uh node uh to be competitive in today's markets and technologies. And we've spoken about logic folding as well in another episode uh where Huawei is planning to take like seven nanometer technologies and put them vertically on top of each other so that you can get more transistors per unit area, right? It's like instead of building suburban homes, you want to build basically apartment complexes uh on wafer so you can get like more people or aka more transistors in a given footprint of urban space, which has now become very, very, very expensive as advanced nodes have gone into the EUV land, right? Uh so yeah, this is if it's true, it's a good good advancement to make um in-house lithography of reasonably advanced nodes. But and some people on X had the belief that UV is just three years away, which I disagree with because UV is a substantially different problem and an extremely difficult one because you have to blast those tin droplets twice uh to generate 13.5 nanometer uh extreme ultraviolet light. That's very difficult. It took ASML 20 years or so to engineer it without all of these supply chain restrictions. Okay, so I don't think it's EUV is going to just like come out of it. There are options that like free electron lasers uh that like X Lite is doing, right? Like maybe there are other technologies that could work. I'm not really sure. Maybe they'll figure out X-ray lithography before Substrate does, which is another company in the Bay Area. Um so there are options that may come out of this, but I don't think EUV is around the corner, right? So that's my whole that's my whole take on this thing.
SPEAKER_01Yep, yep, I hear you. You know, I think um if you're rooting for China to have DUV here for our East Asian listeners, you know, I think probably the best sign here is uh the the way like what a long journey starts with one step, right? And so if you want to make progress, you make progress toward DUV and then figure out how to ramp that and produce it at scale and see that it's working and improve on it, and that will continue to help you take steps. Yes, DUV doesn't necessarily directly lead to EUV because it's a different technology, and there's ways to leapfrog with free electron lasers or X-ray lithography, like you said. But I think they can celebrate, hey, this is DUV. Um, you know, I think it said they're trying to ship five machines this year in 20 and 2027, and so it is definitely a step in the right direction towards self-sufficiency, but it is a you know a far cry from being at the leading edge and necessarily being competitive.
SPEAKER_00But yeah, overall it's been like a hard week uh for semiconductor investors. And with the recent run-up, like there are a lot of people who haven't seen the cyclicality of this industry before, and is probably their first taste of it. Memory investors, for a long time, for example, have seen these things happen over and over again. Uh, but long term, at least I believe, that there is a lot of potential in the technology side. The technology works. What we needed last week, we still need all those things. We still need indium phosphide, we need lasers, we need better computers, we need more memory. Nothing has changed. So it's not investment advice, but you know, if if you're suffering in the market, you know, things will get better. So that's that's my uh closing line, really.
SPEAKER_01There you go.
SPEAKER_00Hang in there, people.
SPEAKER_01All right, thanks.
