Loading summary
Transcript1 lines
- [00:02]
A
Ben Thompson of Stratecheri has a recent piece out titled TSMC Risk in which he calls out TSMC's conservatism as costing the American hyperscalers hundreds of billions in revenue. Before we continue, I want to disclose that I work with Ben. The asianometry newsletter runs on his platform passport and I am friendly with him. I'm not trying to flame him, but I'm hearing many similar views in the Silicon Valley Borg that TSMC is the brake or limiter on the AI boom. As if they're the reason why we don't have AGI yet. Because they didn't and still don't believe if we can ever say that a company that spent 41 billion on capital expenditure in 2025 with another 53 to 56 billion in 2026 planned and is sitting on its hands doing nothing. Now TSMC is a trillion dollar company. They don't need some random YouTuber defending them. Though I reckon people will still accusing me of Taiwan bias in the comments. And to be clear, I largely agree with Ben's final message. TSMC having 90% share of the AI chip market looks pretty unhealthy. That should go down, and it will. Samsung seems to be doing well so far. The point I want to make concerns the nature of hardware. What Ben and others in Silicon Valley are diagnosing as shortages signify TSMC's failure is really Semiconductors are hard and their supply chains are long having more foundry competition wouldn't have averted this compute shortage. The cold, hard reality is that shortages are a fact of life in semiconductors, as are horrific gluts. I was supposed to be working on a video about bananas, but I had to do this first. In today's video, a few scattered thoughts on TSMC taking away the AI Punchbowl I want to talk about the beer game. No, it has nothing to do with drinking beer, nor does it promote drinking because this is astronometry. I will rename it to the Boba game. It is a game developed in the 1950s by the famed MIT professor Jay Forrester, who also did pioneering work on core memories to demonstrate the concepts of system dynamics. I was introduced to it by a TSMC manager and friend. In the game, people operate in a marketplace for boba. The game's players cosplay as boba sellers in one of four retailer, wholesaler, distributor and the factory. The players must work together to minimize costs and maximize revenue. A deck of cards represents weekly customer demand for Boba. Retailers supply boba to the customers, pulling out of their inventory. Retailers want to keep inventory levels as low as possible because it costs money to hold inventory, but also want to avoid stock outs because that's lost revenue. These incur a penalty in negative dollars. The retailer refills low inventories with orders from the wholesaler. The wholesaler in turn must go to the distributor to reload, who then in turn purchases from the factory. At each step we have time delays for order processing, shipping or production. At the start of the Boba game, customer demand is steady, but as the game progresses, the cards start showing unannounced spikes in demand. This simple Boba supply chain has to scramble to adjust, creating delays and shortages or overreactions and overproduction and everyone thinking someone else other than themselves messed up. What we are flippantly labeling as TSMC we really mean is the AI supply chain. And that supply chain is as complicated as you can possibly imagine. Like an iceberg. It looks big enough on the surface of the water but goes way far deeper underneath. TSMC has thousands of suppliers in two equipment like the famed ASML lithography tools and and materials like photoresist, silicon wafers, acid etched gases and so on. These are not generalized tools and materials. They are not fungible like AWS compute units. Just within the bland term of deposition we have wild variations between tools like low pressure chemical vapor deposition, molecular beam epitaxy, atomic layer deposition and so on. These are not interchangeable and each need their own multi million dollar tool. And each tool category niche has maybe three major equipment players like maybe applied, lam, Tel or so on. I also have to mention the non semiconductor stuff too. Power, water, land and labor. Both Taiwan and the United States have issues providing all of this time is needed to build out the infrastructure to provide it. And then there are the memory guys. You cannot ship an AI system without memory. DRAM and nand. Nvidia's AI chips use a special form of DRAM called high bandwidth memory and they use quite a lot of it. The memory industry is just as consolidated as the logic industry with the major players being Samsung, sk, Hynix and Micron. There are also the Chinese memory makers, but they're not being used for AI chips for the west because they are so deep down. The chip guys are last to know when the party is getting started. But first they get baton'd in the face when the police shut things down. Baton'd. I mean bullwhipped. The bullwhip effect is an effect of the Boba game that says that a demand signal tends to amplify as it travels up the various levels of the supply chain. From 1961 to 2006, electronics consumption in the United States grew positively, but with wild volatility swings between 0 to 20%. But for the semiconductor makers, that translates to swings anywhere from negative 20% to 40%. And for the equipment makers, it is amplified even more, plus or minus 60%. The WIP hits particularly hard in the semiconductor industry because of the industry's long lead times. It takes 4.5 months to fabricate and package a chip. It takes 18 months to 2 years to build a fab, meaning from shovels down to producing chips. And it takes 12 to 18 months to produce and install something like an EUV machine into the fab. Another six months before that machine actually starts patterning wafers. Long lead times mean having to make very long demand forecasts, which leads to extreme volatility swings during up and downturns, even if those up or downturns are relatively small. ASML just reported 2025 earnings and we see the bullwhip in full effect. TSMC raised capital expenditure 35%. The ASML announced 13.2 billion euros of net new bookings. Analysts had expected just $6.32 billion. This is because ASML collected orders not just from TSMC, but also Samsung, intel and the memory guys. When it rains, it pours right again. This is why I fear that another AI foundry would not mean our compute shortage is solved. Because ultimately, when those foundries start scaling their capacity, they all go to the same suppliers. Those suppliers then go to their suppliers and everyone gets slammed. We literally just went through all of this a few years ago during the COVID PC and remote working boom. Did we not forget? Remember when the New York Times, Wall Street Journal and the blog boys ran headlines about how the American economy was grinding to a halt because they couldn't get these little trailing edge microprocessors that the car factories are all shutting down. Asianometry was around at this time. I remember how tortured the supply chain was. The carmakers canceled orders during the first lockdowns. But then the economy came back to life over the summer and everyone needed their chips back. TSMC was trying to discern between double booked orders and real demand, which is not an uncommon experience for them. Customers lie about their own demand all the time, or at least we can say that they are eternally optimistic. TSMC tried to respond. In 2022, the Taiwanese giant poured $36 billion into capital expenditure. They went to their suppliers and pushed like no tomorrow. Mark Hayink's excellent 2024 book Focus Details an extremely tense interaction with the TSMC RD svp who is now at Intel. By the way, they even announced new trailing edge fabs. For instance, the original plans for the FAB in Kaohsiung as announced in late November 2022 would have it run a 28 nanometer process node. A trailing edge process node? How weird is that? Well, it turned out those customers really were double booking orders and artificially inflating demand. When the macro environment turned in 2022, the automotive, smartphone and PC chips that were so hot during the COVID era fell out of vogue and customers started cutting orders by the end of 2022. Silicon Valley people though, had already moved on to the next shiny thingy, chatgpt people losing their minds over bots writing poems and code, and the hyperscalers started to figure that they needed a bigger boat data center. Meanwhile, deeper down in the supply chain, TSMC and the rest of the semiconductor industry were getting bullwhipped by Covid hangover utilization at TSMC's multi billion dollar N7 fabs crashed, Semianalysis wrote in April 2023. Now semianlysis data indicates that that the 7nm utilization rates were below 70% in Q1. Furthermore, Q2 gets even worse with 7nm utilization rates falling to below 60%. This is primarily due to weakness in both smartphones and PCs, but there is a broader weakness in most segments. AFAB's break even utilization rates are about 60 to 70%. So those N7 taichung fabs were taking financial losses, potentially on the order of hundreds of millions, maybe even billions. The financial burdens of low utilization are another reason why I am skeptical. Another AI foundry could have rushed into the AI chip fray to save the day. Having slack advanced node capacity means taking massive depreciation losses. Having such pricey non performing 7 nanometer fabs could have been crippling. The TSMC stock in 2022 and 2023 looked pretty precarious, but TSMC pivoted to AI and survived. It's an indication that their product diversification strategy works. There was another semiconductor company that did not do so well during this time, intel. Between 2021 and 2023 they hired 20,000 people, announced billions of dollars of fabs and expansions around the world, and set forth an ultra aggressive process node rollout schedule. Then the COVID PC and remote working boom abruptly ended and then the Hyperscalers started buying GPUs instead of CPUs. As a result, tens of thousands of layoffs, executive turmoil with CEO Pat Gelsinger being forced out and intel took themselves competitively out of the market for what seems like years, a situation that eventually required Japan style state intervention and the mustering of market players to try and reverse the slide. We shall see if such efforts do better than Japan's efforts to save Elpida. Ben points to TSMC's stagnant capital expenditure in 2023 and 2024 and makes a gentle criticism. ChatGPT was released in November 2022 and that kicked off a massive increase in capex amongst the hyperscalers in particular. But it sure seems like TSMC didn't buy the hype. That lack of increased investment earlier this decade is why there is a shortage today and is why TSMC has been a de facto break on the AI buildout bubble. It is true that the hyperscalers started growing their capex in late 2022, but remember the BOBA game again? When does that filter down to TSMC and the rest of the industry and when could they have known? They certainly didn't know in 2023. In the April 2023 earnings call which took place some five months after ChatGPT's release, CC Wei says he noticed ChatGPT's growth, but repeats multiple times that he has no idea what AI's impact on TSMC will be. He also mentioned getting what seems to be the first orders from presumably Nvidia for more coas capacity. Just recently in these two years I received a customer's phone call requesting a big increase on the back end capacity, especially in the coas. We are still evaluating that at the next earnings call in July 2023. He says that AI accelerators are about 6% of TSMC revenue and projected to grow to low teens percent over the next few years. Wall street was looking for such numbers, so I presume they got those projections straight from customers. TSMC also projected their overall 2023 revenue to decline 10%, citing the revenue declines due to macro post Covid and China issues to be bigger than AI. Of course this changed by the end of the year. Has AI surged so much so nobody knew or thought to scale in early 2023. But what about 2024? Well, that year had all the technical issues. I recall news in mid 2024 of TSMC struggling with COAS capacity bottlenecks and yield problems, including one design issue that caused cracks in the Nvidia chip's packaging. Nvidia stock dropped when the news came out and everyone thought that we were so over A former TSMC packaging engineer told me of frantic late night experiments to figure out the right tweaks to fix the problem and Nvidia going so hard as to tell them to take every tweak option and run them on live wafers, the semiconductor version of pushing direct to produce. I also recall News in late 2024 noting how the vendors in charge of making the server racks for Nvidia's Blackwell servers struggled with overheating, liquid cooling leaks, software bugs and connectivity issues. Such technical difficulties delayed server deployment until early to mid-2025, creating a weird situation for several months where TSMC was pumping out chips that just went into storage. So that gated things because you don't scale until you first fix the technical problems. I also want to add that in 2024 TSMC and the rest of the chip industry did not know if those buying AI chips would make money on them. Recall those famous sequoia Capital articles AI's $200 billion question and then the $600 billion question? Those came out in September 2023 and June 2024 respectively. I don't think any sensible foundry would have then committed billions to new fabs. So I argue that the optimal time for TSMC and the rest of the semiconductor industry to really scale CAPEX was 2025, whereupon the BOBA game kicked into effect. Some things just take time. Ben writes that it is chips not power behind the shortage of compute capacity that that the hyperscalers are complaining about. He points the comments from CCWei as support. CC said talking about to build a lot of AI data center all over the world. I used one of my customers customer's answer. I asked the same question. So they say that they work on the power supply five, six years ago. So today their message to me is silken from TSMC is a bottleneck and asked me not to pay attention to all others because they have to solve the silicon bottleneck first. I don't interpret those comments the same way Ben does. TSMC is not a power company. I read that as basically meaning TSMC should be focusing on what they can do and they make chips not power. Also CCWE doesn't speak as carefully as Morris does, but there is no way he is going to say on an earnings call yeah dude they can't get the power connection so they don't need TSMC chips right now. And if this customer's customer is making electricity parameters based on assumptions from five to six years ago then they definitely got a power shortage because AI data centers suck way more power than a CPU centric data center specced out in 2021. And if you want to hear words from a TSMC executive, I point to you to a deleted LinkedIn post from TSMC Arizona's CFO. I don't have a screenshot because she scrubbed that fast, but The URL reads AI's real bottleneck isn't chips, it's power. In the end, I think the power shortages are real and way more serious than the silicon ones. Elon is bringing in truck mounted gas turbines to its data centers, and new gas turbines aren't available until 2029. At least the semiconductor people are trying Semianalysis said in a report that the various legacy gas turbine makers will not greatly expand their factory footprints. They seem a bit grumpy that the turbine boys aren't AGI pilled. I want to close with a thesis that's been percolating in me for a while. The gap between the hardware and software worlds are wider than ever before. I reckon that it's been a good 30 years since Silicon Valley was actually about making silicon, and there's still many silicon people living in Santa Clara, Sunnyvale, Palo Alto. But they tend to be older, retired even. I often go to the Bay Area to talk to people, software people and AI people on occasion, and I ask them how much they know about how their hardware is made. For almost all of them, even the smartest in their domain, they know virtually nothing. It is a hard silicon line. I feel like both sides know so little about the other. My message to Silicon Valley is I'm sorry that Claude code is a little slow for you right now, but the chips are coming. People are torturing themselves to make them put them into racks and start up the data centers. Let's exercise a little patience. Alright everyone, that's it for tonight. Thanks for watching. Subscribe to the channel. Sign up for the Patreon and I'll see you guys next time.