The Compute 100 Podcast with Timothy Arcuri of UBS
Introduction
Brian Shakar: Today's guest is Timothy Arcuri. He is a Managing Director at UBS. Tim had a big summer. He made a really big call on Micron, seeing their memory explosion and the path ahead. We got to hear his takes on that company and so much more in an incredible conversation.
Gabby Lorenzi: Yeah, Tim has a true bird's-eye view of the semiconductor industry and the AI capex buildout. He is really thinking at gigawatt scale and in terms of tens of billions of dollars over 2027, 2028, and what is to come. Hearing his perspective was fascinating. He was kind enough to come across 27th Street to visit us at the office. So excited for everyone to give it a listen.
Brian Shakar: The Compute 100 is a podcast about the most consequential market in AI. These are the conversations with builders and investors about the technology and the capital powering the greatest infrastructure buildout ever. I am Brian Shakar.
Gabby Lorenzi: I am Gabby Lorenzi. We are investors at Primary, supported by Gunderson Dettmer, founding legal partner of The Compute 100.
Brian Shakar: This podcast is for informational purposes only and does not constitute investment advice or an offer to buy or sell any securities. Let's dive into The Compute 100. Timothy Arcuri, Managing Director, Senior Equity Research Analyst covering semis at UBS, welcome to the office. Thank you for crossing 27th Street to come on over.
Can Power Keep Up with Semiconductor Expectations?
Timothy Arcuri: It was easy.
Brian Shakar: We appreciate you making the trek. We were just getting started in talking about pushing the frontier and slowing the frontier. Take us up to speed on how that is impacting your world. Let's dive in.
Timothy Arcuri: Yeah. Obviously, everyone looks at the expectations in semis. If you take what is baked into the numbers and you back into how many gigawatts that implies that are going to get added, and you look at the gap between what the grid can accommodate and what some of the off-grid stuff could be, you can conclude that you have to take a pretty robust view on the upside for the infrastructure to get to what expectations in semis already bake in. So the question is if you have a slowdown.
Brian Shakar: By that, you mean the amount of power that needs to come online on top of what people already think is possible is huge in order to be able to meet the projections that the labs are saying they want to grow into? Is that what you mean by that?
Timothy Arcuri: Yeah. Everyone thinks that NVIDIA is going to grow to be X size and Micron will go to be Y size. If you back into how many gigawatts that implies, the grid cannot accommodate all of that, or what has been permitted currently is nowhere close to that. So you have to believe that there is this conversion of what is not permitted currently, and you have to believe that there is a significant amount of off-grid capacity that also has to come online to get you to where semiconductor expectations already are. That is not to say that it cannot happen. It is just to say that anything that would slow that down, such as issues around safety or whatever things that have actually come out recently, would make you think maybe you do not get as much conversion because geopolitical reasons start to slow that down, along with permitting of new power and things like that.
Gabby Lorenzi: Where are we right now? I am always trying to get a kind of intuition for these things, but I do not have a good sense of how many GPUs that is, and then where we are heading over the next say three years based on NVIDIA's projections.
Timothy Arcuri: This year, we think they are going to ship somewhere in the range of 6 million to 7 million units, GPU units. Now, it is different because die now and units are not the same thing, but we think it is 6 million to 7 million units this year.
Gabby Lorenzi: Where units are really two dies?
Timothy Arcuri: Yes, yeah. If you look at ASICs this year, if you look at Broadcom, our estimate for Broadcom units looking into next year is somewhere in the range of 6 million. They are going to ship probably 3.5 million or 4 million this year. From a unit perspective, ASICs are getting close to the merchant GPUs. As I have shown in the things I published, ASICs are growing faster than merchant GPUs. Now, from a dollar perspective, NVIDIA's dollar share will be higher than its unit share obviously because the ASPs are higher.
Brian Shakar: And so that is like about 12 million units something that we are heading towards?
Timothy Arcuri: Yeah.
Brian Shakar: How much power does that need?
Timothy Arcuri: I would have to go back and look at our model as to how much power.
Brian Shakar: We are going to do this in the show notes, or show you introduce it. We are going to figure that out.
Timothy Arcuri: Because there is a big delta between the permitted power. It is not so much of a today problem. It is not even a 2028 problem, although to get to where people expect NVIDIA plus AMD plus Broadcom plus MediaTek plus Amazon, that stuff does require power. You cannot just take what has been permitted on the grid and take what you can see and say does that match what is in expectations. No, it does not. So there has to be additional things that happen between now and then on the infrastructure side to get you to where we already think semis will be in 2028, not in 2027, but more in 2028, and certainly in 2030 that actually has to happen. Public markets are weighing where the bottlenecks are. For a while, people were like, well, they are in semiconductors, and so that is where all the money flowed. Now, I think people are coming to the realization that expectations in semis are here, and expectations in some other parts of the market are here, so this becomes a more interesting bottleneck to put capital toward. You saw in this correction back in May and June, money came out of this group, my group. There has been a hesitation to regrowth back up. Most big funds I talk to have regrown back up about half of the exposure that they had prior to this correction. There is a general sense of a lot more questions being asked around this than were being asked back in May and June when it was a bunch of performance chasing and people were looking at 2028, 2029. You created this high-water mark almost, is how I look at it. Even though the stocks have come down and they are not back up to where they were back in May and June, that high-water mark has been established and it has not changed. Now I think there is this thinking that maybe we have to grow back into that high-water mark. This grid limitation is one thing that people fear could prevent you from getting to where expectations already are out in 2028 and beyond.
Brian Shakar: Is there a world in which this whole notion of pacing the frontier actually helps us on the power side, if we are either artificially or in some other way constraining the pace at which the labs are moving, so that we are able to push the power problem off a couple of years? How is the market thinking about the way that that ultimately hits?
Timothy Arcuri: I have not heard a lot of that today, but I think that there is something to it. I did a call today with NVIDIA, and they mentioned the same thing, that that could push the problem off a little bit. Then, of course, there is the open-source versus closed-source model debate. People generally think that frontier models are much more powerful and consume a lot more power, but these open-source models are pretty competitive, and they also consume a lot of power. If open source becomes the alternative to kind of a paced frontier, then that actually could mean the opposite for the power problem, where things actually move up a lot quicker.
Brian Shakar: Interesting. The lack of power could end up just being a natural pacer.
From Process Engineering to Wall Street
Brian Shakar: How did you end up doing what you are doing? You have been doing it for a long time, so you have a very prepared mind for this moment. Take us way back to how you got into covering semis, and give us the very abridged version of the journey to 2026.
Timothy Arcuri: I was a chemical engineer in college, and I worked as an engineer for a couple years. I worked for an industrial gas company, and we designed processes. I was a process engineer, and we basically developed a technology to take all these gases. When you are making chips, you use a bunch of noxious gases, and many of these gases are not consumed at all. In fact, maybe 90% gets exhausted out the stack, and some of these things have 50-year half-lives. The industry back in the mid-to-late 1990s was pushing to become cleaner. That was the initial push. They went to these industrial gas companies that were selling these gases and said, "Help us design a process to better utilize this stuff." At the company I worked at, we had a little stealth side group that basically developed a process to capture and purify this stuff. I was one of the engineers on this, and we got a patent on it. I have a patent and the whole thing. That and $5 will get you coffee. But that is sort of how I first began working with the tool companies. We were working with the chip and the tool companies, so I got to know them back in 1995, 1996, 1997. Then I realized that I did not want to be an engineer. I realized that I was good at writing and I was good at talking to people. I decided there is this industry, Wall Street, with the buy side and the sell side. I ended up on the sell side. I wrote a research report and I said, "I can do this. I don't have an MBA, but I can do this." Somebody gave me a chance back in 1998, and so I have been doing it almost 30 years.
Gabby Lorenzi: The whole time it was in semis effectively?
Timothy Arcuri: Yes, because you cannot cover the supply chain without covering what is happening in semiconductors. Now, I did not cover NVIDIA during that whole timeframe. I have covered NVIDIA for the last probably 10 years or something like that. I evolved from the supply chain into covering the actual chipmakers themselves.
Gabby Lorenzi: What was your holy moment, probably five years ago, in which I imagine your job just totally changed, or maybe it was exactly what you expected over the 15 years prior to that? I always love to hear people who were so deep in this market explain when you woke up and said the world is about to change.
Timothy Arcuri: It is funny. I do not even know if I understood at the time the gravity of this, but I remember we had Jensen in our office in San Francisco at 555 California. This was probably February 2023, and there were probably 30 investors in the room. At the time, it was me in a car with Colette, Jensen, and IR driving around San Francisco.
Brian Shakar: It was probably worth like $500 billion?
Timothy Arcuri: The market cap was probably even less than that. It was probably $70 billion or $80 billion. ChatGPT had come out, but no one really knew how big it would be. Jensen just looked out the window and he said, "The future is so bright, I can't even tell you how bright the future is." He was so bullish in this meeting.
Brian Shakar: Which I am sure you see a lot of, though.
Timothy Arcuri: Yes, but there are certain moments where you do not really know in the moment what it means. I kind of walked out of that being like, wow, there is something that I really need to get my arms around here, because I have dealt with a lot of management teams in my career, and there is something special happening here. You go back and you look at user data on GPT and you see what is happening, and one thing led to another. Of course, at that time, I never realized how big NVIDIA would be, nor, I think, probably did even he at that time. One thing led to another, and then it evolved into getting probably 100 to 150 investor questions per day. It is really impossible to keep up. I almost have to clone myself for all these call requests, so it is a good problem to have. Then I did get some death threats, which happens when you say things that certain people do not like and you get threats. I always tell my wife I am going to write a book one day about what has happened the last five years.
Brian Shakar: We would love to see the first edition.
Measuring AI Demand and Pricing Faster Inference
Brian Shakar: I am curious about what your mental model is about how you try to understand what is going on. How do you go about thinking about it?
Timothy Arcuri: I think the first thing has been making sure that, with all this capacity being built, it is obvious to me that demand grossly exceeds supply. Why? Because I look at instances that are 7 to 8 years old, Ampere instances, A100s that are coming off lease, and all you have to know to me is that those re-lease for higher. That tells you right off the bat, you need to know nothing else other than the fact that demand so far exceeds supply if the old stuff coming off lease is re-leasing for a higher price. So I have not had to worry that much about demand because it has been all about supply. Most of my work has been around how fast TSMC is expanding wafer capacity and how many tools the equipment makers can make. It has been predominantly supply-side work. Now it is shifting a little more to stuff where I am trying to look at demand and say, here are the expectations, they have grown and grown, and now here we are. A, are the other parts that I do not look at as much able to accommodate this in the timeframe that we think they will? And B, is there demand for that? That is the next thing now where people are asking, "Tim, what about demand?" I published a number this morning that there is going to be 380 gigawatts added, but that is a power number, not an IT compute number; that includes the PUE factor. The question is how much revenue has to get generated if we are going to add 380 gigawatts over the next five years. That is where the conversation is going. You can look at some of the labs that have given ARR numbers per gigawatt, and you do all that math. Those are the questions that I am going to have to start to do more of to say how much you have to believe that this all adds to global GDP to therefore believe that you can absorb all of these gigawatts.
Gabby Lorenzi: We worked on a presentation where we were thinking about $30 billion to $50 billion per gigawatt to buy, set up, and deliver on, so that over a 4-to-5-year period you need to generate something like $80 billion in gross profit. Is that the ballpark of how you are thinking through it, in terms of what it will cost to deliver on a gigawatt of power and how much you can make off of that?
Timothy Arcuri: Yeah, yeah.
Gabby Lorenzi: One of the bigger levers there is around the price per token that people will be willing to pay. How do you think about that? How do you wrap your head around where that is going?
Timothy Arcuri: I myself have questions around this a lot because, in any technology industry, cost per token is going to come down, and the question is how fast. Now there is this market where you look at tasks like coding. There is a market now for fast inference, and it is a big market. Jensen went out and bought Groq, Cerebras went public, and there is SambaNova. There are a bunch of people servicing the fast inference market, so it is not just about cost per token. There is also a performance part of the market.
Brian Shakar: And they are always willing to pay up. There is always going to be a willingness to pay up for premium tokens and delivery of those tokens.
Timothy Arcuri: It seems like it. The question is whether it is 10% of the market or 20%. Andrew Feldman would say that is the whole market, that there is no market for slow inference and only a market for fast inference. I certainly understand what Andrew is saying about that. Jensen has said different numbers at different times, but pretty much 10% to 20% of the market is what he thinks it is. My thinking on what the market will bear in terms of price per token is evolving.
Brian Shakar: How do you build conviction in where you land on these things? Especially on the token side, is it spending time with the labs, spending time with the chip companies, or talking to developers? Where do you build your thesis around something as granular as token economics?
Timothy Arcuri: As a semiconductor analyst, I am a P times Q guy, and it gets a little out of my comfort zone when I am trying to bridge the gap to what tokenomics are actually going to be. This is an area where I try to marshal resources from our software team, our internet team, and our utilities team. I cannot say that I have all the answers, but this is the question that I am trying to answer because I can tell you how much capacity is going to get added. What I have a harder time telling you is whether the demand will be there. I suspect within five years it will be there, but what will be the killer app that drives that demand? I am not really sure. If you take some of the numbers that the labs have put out in terms of revenue that they get per gigawatt, although it sounds like a big number, looking out five years, it is only a couple percentage points of incremental global GDP. If you pose it that way and said to the average investor, "Stocks are baking in that AI will add a couple percentage points within five years to global GDP," does that sound crazy? I do not think so.
How AI Is Changing Semiconductor Research
Gabby Lorenzi: I am fascinated by how analysts do their work, in the sense that you must be ingesting so much information as well as writing memos and notes. All of that is probably getting crowded in a world where you are reading a lot of AI content and trying to use AI productively to improve your role. How are you feeling your job is changing, either for the better or in ways where the industry needs to get this right around AI's impact on actually being an analyst that covers semis?
Timothy Arcuri: There is a lot of stuff we are doing that every research department is doing, very basic stuff to make our content more digestible and make it easier for us to write, and from a legal perspective, to get things out the door easier without having to read every single word. Those are all pretty basic things. But coming back to the P times Q in me, I have to start to wear hats that I have never had to wear before. I am speaking to the choir because you guys do this all day every day, wearing different hats all the time. For me, having to be an expert on things that are beyond semiconductors—I have to know about ROI and what the use case is going to be for AI—that is not a world that I have historically been an expert in. I found myself having to try to become an expert in a much wider array of things. My job used to be not an inch wide and a mile deep, but maybe a foot wide and a mile deep, or maybe 5 feet wide. Now I have to struggle with how to be 100 feet wide and a mile deep, and that is hard.
Gabby Lorenzi: When reading research notes, using AI to understand the literature has been valuable on the consumer end. But for someone writing and reading constantly, it completely shifts the workflow.
Timothy Arcuri: As Elon Musk said, there are going to be no more apps, and AI is going to do everything for you. Maybe you just say, "What should I buy?" and then there are not any more stocks; it is just one big blob. That could be one world we end up in.
NVIDIA, AMD and Intel
Brian Shakar: If we take a step back and think about the key players on the processor side—the compute landscape—NVIDIA has led the charge, AMD is competing, and Intel is having somewhat of a rebirth. What are your perspectives on those three companies, their standing, and what they are doing best? Let's get a high-level lay of the land across compute, and then jump into memory.
Timothy Arcuri: With NVIDIA, just with CUDA and the huge installed base, what is amazing to me is that every single GPU that they are shipping today—and the same thing with AMD—is still a greenfield shipment. We have not even begun to upgrade the installed base. NVIDIA in particular has this monetization of the installed base that the other guys do not have, simply because they have such a massive base. Right off the bat, CUDA and the software moats are what initially got them out in front, and that installed-base monetization opportunity is very unique to them. With AMD, initially ROCm was not there, but now most of the companies they deal with do not really care about ROCm. It is not a big impediment for the hyperscalers because they have their own software resources. ROCm is important for an enterprise customer, but not for a hyperscaler. So I do not think that is as much of an impediment anymore. AMD from a hardware perspective is very competitive for a lot of workloads. The question now is whether there is some compelling reason that makes AMD better than NVIDIA, and in most applications, I still do not think there really is. NVIDIA is also out in front with respect to rack engineering, and they have crossed a lot of the hurdles that AMD is now having to cross as you start to build racks and handle the whole ecosystem. The AI factory story is very clear on NVIDIA's side. NVIDIA is out creating quarters of capacity, so they are two steps ahead from that perspective. If they continue to execute on an annual cadence, it makes it very tough for anyone to catch up. I still have very high estimates on AMD because the market is huge, so AMD is going to be a much bigger company in the next few years. On Intel, I am much more bullish on Intel's foundry business and manufacturing than I am on their product roadmap. Lip-Bu Tan is awesome, and he is turning things around and will continue to do that. But there is a much more compelling reason for a customer to engage with Intel on the manufacturing side, whether in EMIB for packaging or in front-end foundry. That is where my excitement is on Intel.
Custom Chips, Broadcom and MediaTek
Gabby Lorenzi: What about the hyperscalers and custom silicon? That feels like an equal part of the puzzle.
Timothy Arcuri: I view Broadcom as the enabler of the hyperscalers. The ASIC journey is one where first you start without your own infrastructure or semiconductor team, and you build it over time. You start by going to a full-service ASIC provider whom you can trust to make chips in high volume: Broadcom. In the case of Google, it took seven generations to finally figure out that they have the resources to do it themselves and procure HBM directly. Google is 7 to 10 years down this path. Amazon is probably even farther down that path because they bought Annapurna over 10 years ago. I put Amazon and Google in a different category than everybody else. The journey is the same: nobody wants to pay NVIDIA so much money, but a lot of them have no choice. They want to control their own destiny, but you are never going to bet your whole infrastructure on an ASIC. There is always going to be a blend. Within ASICs, each hyperscaler is at a different point in that evolution. If you look at Broadcom's Google business, what Hock Tan is saying implies that Google is going to be flat or grow a little bit in 2028, but the growth flattens off as incremental volume goes to MediaTek and others. Google is not going to go away because they are still going to be reliant on Broadcom for things nobody else can do, but it is an evolution.
Brian Shakar: Can you explain the move to MediaTek? What is that?
Timothy Arcuri: Customer-owned tooling, or COT. When you want to do it yourself, procure the HBM, and do not want someone to make the entire chip for you because you want to add a lot of engineering for the chip, that is what they are doing with MediaTek. MediaTek is their partner on the COT side. Broadcom comes in and says, "I'll make the entire thing for you. I'm going to get the memory, I'm going to get the wafers, and I'm going to charge you 60% margin." Google is starting to diversify away from Broadcom to do more of its ASIC development with MediaTek.
Brian Shakar: Is that because they have more buying power to strike better deals on HBM and with partners down the line, and they want to dedicate resources internally?
Timothy Arcuri: Yes, exactly.
Brian Shakar: NVIDIA also did a big deal with MediaTek recently, right? Was that around a chip for laptops?
Timothy Arcuri: At the end of the day, you look at the deals NVIDIA is striking. They are partnering with players outside the Broadcom camp. You have Broadcom on the hyperscalers, and you have NVIDIA and the Neoclouds. NVIDIA wants NVLink Fusion to be used with more ASICs. They go to ASIC players who are not Broadcom, partner with them, and say, "Sign a deal with me, and you can use NVLink Fusion and my IP." I think the MediaTek deal was around that.
Brian Shakar: It was discussed in the context of a laptop chip, but it is a broader play around fortifying capability to challenge Broadcom's rise in custom ASICs.
Timothy Arcuri: Yes. The amount of coopetition that goes on in semiconductors is wild. You think two players are at odds, and then you realize they are completely reliant on each other.
Inference Startups and Cerebras’s Business Model
Brian Shakar: What do you think about all the startups attacking this market right now?
Timothy Arcuri: It is really interesting. Fast inference and disaggregated inference are creating massive opportunities. Prefill is compute-bound and decode is memory-bound, so you are seeing all this innovation on the ASIC side. We talked about Cerebras, SambaNova, and others who have tiered memory architectures. You are seeing tons of innovation around disaggregated inference, with Etched being one company you know well. The opportunity to improve the utilization of the existing GPU fleet is huge. You talk to these clouds and they will tell you that, in practical terms, GPU utilization is 30%, 40%, or 50%. There are huge opportunities to accelerate the fleet.
Gabby Lorenzi: Fresh off the Cerebras IPO, how has that transition from the private market into the public stratosphere impacted your coverage and the market's perception?
Timothy Arcuri: Cerebras has been very successful. Their fortunes changed significantly when they signed the OpenAI deal, which is documented in my research. Now they have a framework deal with Amazon as well. Most companies want to do disaggregated inference to accelerate their existing installed base. Cerebras helped trigger the light bulb for the fast inference market. Right after they signed that deal, Jensen bought Groq. A lot has changed in the past six months, and there is a ton of opportunity for Cerebras. The market is so big that they do not have to capture much share to be a much bigger company.
Brian Shakar: The market will be watching the OpenAI deployments. What are the key details there?
Timothy Arcuri: With Cerebras, the framework deal with OpenAI consists of three tranches of 750 megawatts each, with some optionality. Initially, that deal is a cloud deal because OpenAI does not have their own infrastructure, so they asked Cerebras to find a colocation facility and handle it. Over time, we will see how it evolves. My assumption is that Cerebras is 50% cloud and 50% hardware over time. That is a major point of evolution in their model because selling both hardware and a cloud makes them unique, though I think that is currently done more out of necessity than long-term preference. Some in the market argue that for very large models, because you have to store everything on the wafer, the minute you take it off the wafer, some competitive advantages go away. I am sympathetic to that argument, but I do not think it is that big of a deal. Others, including Elon Musk, have tried wafer-scale and were not able to do it. Cerebras has really lit the fuse for privates to attack disaggregated inference.
Micron, HBM and the Memory Cycle
Brian Shakar: Let's talk about memory. It was a big year for you in memory and for the memory market in general. Talk about Micron, how that company has evolved, and what it has been like covering it.
Timothy Arcuri: If you look at the stock, it has gone up a ton, but earnings have also gone up a ton. The market trades the stock as if nothing has changed, assigning it a very low multiple of around 4 times earnings. The market does not know what to do with earnings that rise that fast, particularly for a company that has historically been cyclical. NVIDIA has had cycles, like when crypto came and went and its multiple traded at 40 times trough earnings, but memory has always been cyclical because it has been driven by supply rather than demand. The market always overbuilds supply; it takes two years to bring a wafer fab online, creating huge cycles. What has changed is HBM. HBM operates on an annual cadence because every new generation of ASIC or XPU requires a different version of HBM. There is constant connectivity required with the memory producer, where you have to specify requirements every year for the new flavor. When a new chip comes out, the old flavor of HBM is generally no longer usable. There is an obsolescence that did not exist before. In the DDR world, DDR is good for 7 to 8 years, so customers build inventory when the price is low and burn it down as the price goes up, creating big cycles. HBM changes that dynamic.
Brian Shakar: So HBM ties memory directly to the compute roadmap.
Timothy Arcuri: Correct. In an AI world, inference cost is significantly memory-driven. In disaggregated inference, I use the analogy: when you were a kid and touched a hot stove, you used logic cells to learn not to touch it in the future. Every time you walk into a kitchen after that, you generate an inference token of value, but it comes out of your memory bank. Inference is heavily driven by memory, making memory far more valuable in an AI world. The market looks at Micron's earnings, assumes they will crash, and puts a low multiple on them. Currently, the earnings multiple is very close to the book multiple, which you hardly ever see because earnings have grown so fast that book value has not caught up.
Gabby Lorenzi: What is book value in this context?
Timothy Arcuri: Tangible book value consists of PP&E, cash, and balance sheet assets. During an upturn, earnings drive the stock; during a downturn, it trades on book value because people worry that memory makers will burn cash, which erodes book value. Investors revert to book value as a floor because it is less volatile than earnings. Analysts who have been around a long time see Micron at 85% gross margin and assume upside momentum is exhausted. It stops being a directional trade and becomes a duration argument: how durable are the earnings? Historically, they have not been durable, but hyperscalers are now signing long-term agreements.
What Long-Term Memory Contracts Change
Timothy Arcuri: Hyperscalers are crying uncle because costs have risen so much. They need certainty on input costs, so they commit to $10 billion prepayments to buy product over the next three years.
Brian Shakar: That is the biggest structural shift: memory has never done long-term contracts before.
Timothy Arcuri: Never like this. Previously, you sat down on January 1st to discuss annual volumes with a framework agreement, but in a price-driven market, buyers would take the volume at half the price if spot dropped. Memory capacity is now sold out for a couple of years across major HBM players, which has never happened in past cycles. This is due to the annual cadence of HBM and because producing an HBM wafer consumes about four times the capacity of a non-HBM wafer. As GPUs suck demand toward HBM, it acts as a supply anchor, removing wafer capacity from the broader market because vertical stacking with 95% yield per layer eats up fab volume. Demand went up while available standard supply went down.
Brian Shakar: So the cycle becomes more durable. Do you believe this is a structural change rather than a standard cycle?
Timothy Arcuri: There will eventually be a cycle on the other side. After 30 years in the business, every time someone says "it is different this time," it rarely is entirely. There will be a collapse in memory prices eventually. But consider how these contracts work: if you signed a deal 3 or 4 months ago, by year-end the contract price is roughly half of the open market spot price because spot kept climbing. The contracts include a floor price far below current levels. Even if open market prices collapse 70% to 80%, customers will still take product under those contracts. Micron continues earning significant money even in a major downturn, generating cash and compounding book value, which brute-forces the stock higher.
Brian Shakar: How do you view Micron against SK Hynix and Samsung?
Timothy Arcuri: I do not cover the Korean companies; our Asia team covers them. But Sanjay Mehrotra has transformed Micron into a Silicon Valley-style company. It used to be a slower-moving firm out of Boise, but Sanjay brought in great talent from SanDisk. The technology gap has closed, and in many ways Micron has become a leader rather than a laggard. There is also a geopolitical angle: the U.S. administration wants to support domestic semiconductor manufacturing, making Micron strategically aligned with U.S. interests alongside Intel.
Brian Shakar: How do their fab expansions in places like upstate New York impact your view?
Timothy Arcuri: Historically, investors hated seeing massive fabs announced because it signaled oversupply. But there were never long-term agreements backing them before. Past oversupply was caused by producers chasing spot prices with speculative capacity additions. With LTAs providing revenue visibility, capital expenditure is disciplined and tied to commitments, reducing the risk of massive overbuilding. A downturn this time would have to be driven by a collapse in demand rather than an unforced surge in supply.
Equipment, Packaging and Fab Economics
Brian Shakar: What is the market missing on semiconductor equipment companies like Applied Materials and Lam Research?
Timothy Arcuri: The correlation between Applied Materials or Lam Research and Micron is around 0.9. Equipment stocks get pulled into the memory debate regarding peak margins versus duration. But wafer fab equipment is much earlier in its capacity addition cycle than compute. TSMC has had capacity for compute to grow for years, whereas equipment companies are just beginning to fill new cleanrooms. Fundamentally, equipment allows you to underwrite multiple years of growth. Based on our math, it requires roughly $4 billion of front-end equipment per gigawatt of compute capacity.
Brian Shakar: And we estimated 380 gigawatts cumulatively through 2030 or 2031?
Timothy Arcuri: Yes. If you have an out-year with 100-plus gigawatts added, 100 times $4 billion implies over $400 billion in wafer fab equipment spend. That is beyond my official forecast horizon, but it shows the potential scale compared to this year's global wafer fab equipment market of $155 billion to $158 billion.
Brian Shakar: Where is the core innovation happening in equipment? Is it in new memory cell designs, machine learning for defect classification, or elsewhere?
Timothy Arcuri: A lot of processes that were traditionally front-end are moving to the back end. Advanced packaging, vertical stacking, and new materials are driving significant innovation. Companies are working with entirely new material sets in memory that require specialized equipment. Others are rethinking fab architecture and wafer routing. Many fabs still rely on 40-year-old batching paradigms where wafer cassettes sit idle waiting for tools. There is also innovation in high-aspect-ratio etching. Lam Research recognized early that architectures were going vertical, as seen in 3D NAND and TSVs, and positioned their portfolio accordingly.
Gabby Lorenzi: Advanced packaging, materials, power delivery, and cooling are getting integrated directly into silicon, strengthening incumbents like TSMC and Intel who control the stack.
Timothy Arcuri: It used to be that the industry transitioned to a larger wafer size roughly every 10 years to reduce die cost, since costs do not scale 1:1 with wafer area. A large chipmaker or consortium would fund equipment vendors to make tools for the next diameter. 300mm was introduced around 2002. If you told me in 2005 that we would still be on 300mm over 20 years later, I would have said you were crazy. Given TSMC's scale, I am surprised someone has not funded the next transition. However, the limiting factor has shifted from wafer size to the scanner field size of ASML's lithography tools.
Brian Shakar: What is the field size?
Timothy Arcuri: The reticle limit defines the maximum printable area of the scanner. Transistors are patterned within that fixed optical window. It is like an optical negative: you can only expose a certain area before you have to stitch dies together, which is why processors like NVIDIA's latest chips are reticle-limited multi-die packages.
Brian Shakar: Do you cover ASML or TSMC directly?
Timothy Arcuri: I do not cover ASML or TSMC directly. ASML is covered by our European team and TSMC by our Asia team, but understanding their roadmaps is fundamental to covering the rest of the supply chain.
Where New Compute Companies Can Compete
Gabby Lorenzi: What questions do you have for early-stage investors seeing new startups enter the landscape?
Timothy Arcuri: I wonder what your view is on compute architectures. Have we gone so far down the track that if you have not already invested in a compute architecture, it is too late? Will early-stage innovation over the next five years be about processors, or will it shift to systems-level technologies or photonics?
Brian Shakar: We invested in Etched's seed round. Competing head-on in general silicon against existing roadmaps is difficult given capital requirements and incumbent scale. Photonic computing represents a true paradigm shift where a compelling approach could warrant early-stage backing. Later-stage venture investors are indexing into compute, funding Series B and C rounds for pre-revenue silicon companies with substantial capital, with players like BlackRock participating.
Timothy Arcuri: Memory constraints are pushing architectural shifts. NVIDIA is adjusting specifications to build larger scale-up domains, which shifts value toward networking, optics, and interconnects.
Gabby Lorenzi: At the seed stage, we focus on advanced co-packaging, electro-optical materials, novel memory architectures, and biological computing. With private and public valuations scaling—SpaceX at $2 trillion, NVIDIA reaching $5 trillion—the sheer dollar scale pulls late-stage capital into compute derivatives, while early-stage capital looks for fundamental enabling technologies.
Timothy Arcuri: Semis represent roughly 18% of the S&P 500 now, driving investors to seek derivative plays in industrials and cooling infrastructure. How does that materialize across your firm?
Brian Shakar: Every team has a compute angle. Industrials looks at power generation and grid efficiency for data centers. Fintech evaluates compute futures and pricing indices. Enterprise AI looks at ROI attribution relative to compute expenditure. Edge compute, physical AI, and hardware security are seeing similar cross-sector activity.
Cobots, Humanoids and Demand for Chips
Gabby Lorenzi: What trends in physical AI and robotics are translating to the public markets?
Timothy Arcuri: Collaborative robots, or cobots, are commercial today. Teradyne has a substantial cobot business through Universal Robots. While growth has faced macro headwinds, adoption is expanding, as seen in Amazon's deployment of cobots for warehouse pick-and-stow operations. Humanoids are generating significant commentary from executives. From a semiconductor content perspective, a humanoid robot is essentially a walking electric vehicle. If unit volumes scale, the silicon content represents a massive market, though widespread commercialization remains further out.
Political Backlash and Customer Concentration
Brian Shakar: How are you thinking about political headwinds, such as the anti-data center push and regulatory pressure around AI safety?
Timothy Arcuri: Over the past year, people asked about black swan risks. I do not worry about near-term demand or supply. My primary concern has been social and geopolitical backlash—a modern equivalent of people showing up with pitchforks over job displacement or power consumption. That could disrupt political support for AI infrastructure. The industry needs better communication and closer collaboration with government administrations, though the competitive dynamic with China limits how much development can be slowed.
Gabby Lorenzi: There is also concern regarding economic power concentration among a handful of hyperscalers.
Timothy Arcuri: Highly concentrated capex at NVIDIA creates downstream customer concentration for Applied Materials, Micron, and the rest of the value chain. Investors frequently question the long-term sustainability of that customer base.
Brian Shakar: If you were not covering semiconductors, what would you be doing?
Timothy Arcuri: I would probably be a DJ and play golf. I love 1980s music and blending it with modern techno sounds to make mixtapes. I just do not have the time right now.
Brian Shakar: Hopefully you can produce our intro track for next season. What is one non-consensus prediction you have for the next year or two?
NVIDIA’s Installed Base and What Gets Cut in a Downturn
Timothy Arcuri: People underestimate the long-term value of NVIDIA's installed base. Every dollar of revenue today represents greenfield deployments; none of it is brownfield replacement. Over time, that installed base will be monetized through upgrades and efficiency improvements. Blackwell and Rubin shipments today are purely additive to global compute capacity, but eventually older architectures like Ampere will be replaced. That creates an upgrade cycle that cushions revenue even when overall capex slows.
Timothy Arcuri: Furthermore, I had a debate with a CIO regarding what gets cut first when hyperscaler capex flattens: custom ASICs or merchant GPUs? Most investors assume merchant GPUs get cut first to protect the gross margins of internal custom silicon. My view is that custom ASICs get cut first. ASICs are built for narrow workloads, whereas merchant GPUs provide the architectural flexibility needed during a retrenchment.
Brian Shakar: Thank you so much for your time and perspective, Tim.
Timothy Arcuri: Thank you.
Brian Shakar: As a research analyst, Tim is required to provide certain disclosures relating to the nature of his relationship and that of UBS with any company on which he expresses a view. These disclosures are available at ubs.com/disclosures.