TSMC Risk
Recorded: Jan. 26, 2026, 3 p.m.
| Original | Summarized |
TSMC Risk – Stratechery by Ben Thompson By Ben ThompsonAboutEmail/RSS@benthompsonExploreConceptsCompaniesTopics Stratechery Plus Subscribe Log In Learn MoreMember Forum Menu Search× Stratechery Plus Subscribe Log In Learn MoreMember Forum By Ben ThompsonAboutEmail/RSS@benthompsonExploreConceptsCompaniesTopics TSMC Risk Listen to this post: Log in to listen You probably think, given this title, you know what this Article is about. The most advanced semiconductors are made by TSMC in Taiwan,1 and Taiwan is claimed by China, which has not and will not take reunification-by-force off of the table. Anthropic Chief Executive Officer Dario Amodei said selling advanced artificial intelligence chips to China is a blunder with “incredible national security implications” as the US moves to allow Nvidia Corp. to sell its H200 processors to Beijing. “It would be a big mistake to ship these chips,” Amodei said in an interview with Bloomberg Editor-in-Chief John Micklethwait at the World Economic Forum in Davos, Switzerland. “I think this is crazy. It’s a bit like selling nuclear weapons to North Korea.” The nuclear weapon analogy is an interesting one: a lot of game theory was developed to manage the risk of nuclear weapons, particularly once the U.S.S.R. gained/stole nuclear capability, ending the U.S.’s brief monopoly on the technology. Before that happened, however, the U.S. had a dominant military position, given we had nuclear weapons and no one else did. Perhaps Amodei believes the U.S. should have advanced AI and China should not, giving us a dominant military position? You’re going to see us continue to be very aggressive investing in capacity because we see the demand. As fast as we’re adding capacity right now, we’re monetizing it. Here was Microsoft CFO Amy Hood on the company’s earnings call: Azure AI services revenue was generally in line with expectations, and this quarter, demand again exceeded supply across workloads, even as we brought more capacity online. Here was Google CFO Anat Ashkenazi on the company’s earnings call: In GCP, we see strong demand for enterprise AI infrastructure, including TPUs and GPUs, enterprise AI solutions driven by demand for Gemini 2.5 and our other AI models, and core GCP infrastructure and other services such as cybersecurity and data analytics. As I’ve mentioned on previous earnings calls, while we have been working hard to increase capacity and have improved the pace of server deployments and data center construction, we still expect to remain in a tight demand-supply environment in Q4 and 2026. Here was Meta CEO Mark Zuckerberg on the company’s earnings call: To date, we keep on seeing this pattern where we build some amount of infrastructure to what we think is an aggressive assumption. And then we keep on having more demand to be able to use more compute, especially in the core business in ways that we think would be quite profitable than we end up having compute for. Earlier this month, TSMC CEO C.C. Wei admitted that the shortage was a lack of chips, not power; from the company’s earnings call: Talking about to build a lot of AI data center all over the world, I use one of my customers’ customers’ answer. I asked the same question. They told me that they planned this one, 5-6 years ago already. So, as I said, those cloud service providers are smart, very smart. So, they say that they work on the power supply 5-6 years ago. So, today, their message to me is: silicon from TSMC is a bottleneck, and asked me not to pay attention to all others, because they have to solve the silicon bottleneck first. But indeed, we do get the power supply, all over the world, especially in the US. Not only that, but we also look at, who support those kind of a power supply, like a turbine, like, what, nuclear power plant, the plan or those kinds of things. We also look at the supply of the rack. We also look at the supply of the cooling system. Everything, so far, so good. So we have to work hard to narrow the gap between the demand and supply from TSMC. The cause of that gap is obvious if you look at TSMC’s financials, specifically the company’s annual capital expenditures: After a big increase in CapEx in 2021, driven by the COVID shortages and a belief in 5G, TSMC’s annual CapEx in the following years was basically flat — it actually declined on a year-over-year basis in both 2023 and 2024. Note those dates! ChatGPT was released in November 2022; that kicked off a massive increase in CapEx amongst the hyperscalers in particular, but it sure seems like TSMC didn’t buy the hype. To put it another way, if Altman and OpenAI are the ones pushing to accelerate the AI infrastructure buildout, it’s Wei and TSMC that are the brakes. The extent to which all of Altman’s deals actually materialize is dependent on how much TSMC invests in capacity now, and while they haven’t shown their hand yet, the company is saying all of the right things about AI being a huge trend without having yet committed to a commensurate level of investment, at least relative to OpenAI’s goals. That Update was about the future, but it’s important to note that the TSMC brake has — if all of those CEO and CFO comments above are to be believed — already cost the biggest tech companies a lot of money. That’s the implication of not having enough supply to satisfy demand: there was revenue to be made that wasn’t, because TSMC didn’t buy the AI hype at the same time everyone else did. Make no mistake, $54 billion is a big number, one that Wei admitted made him nervous: You essentially try to ask whether the AI demand is real or not. I’m also very nervous about it. Yeah, you bet, because we have to invest about USD52 billion to USD56 billion for the CapEx, right? If we did not do it carefully, that will be a big disaster to TSMC for sure. So, of course, I spent a lot of time in the last three-four months talking to my customers and then customers’ customers. I want to make sure that my customers’ demands are real. Wei made clear that he was worried about the market several years down the line: If you build a new fab, it takes two and three year, two to three years to build a new fab. So even we start to spend $52 billion to $56 billion, the contribution to this year is almost none, and 2027, a little bit. So we actually, we are looking for 2028-2029 supply, and we hope it’s a time that the gap will be narrow…So 2026-2027 for the short-term, we are looking to improve our productivity. 2028 to 2029, yes, we start to increase our capacity significantly. And it will continue this way if the AI demand megatrend as we expected. First off, this delayed impact explains why TSMC’s lack of CapEx increase a few years ago is resulting in supply-demand imbalance today. Secondly, notice how this year’s planned increase — which again, won’t really have an impact until 2028 — pales in comparison to the CapEx growth of the hyperscalers (2025 numbers are estimates; note that Amazon’s CapEx includes Amazon.com): Remember, a significant portion of this CapEx growth is for chips that are supported by TSMC’s stagnant CapEx growth from a few years ago. It’s notable, then, that TSMC’s current and projected CapEx growth is still less than the hyperscalers: how much less is it going to be than the hyperscalers’ growth in 2028, when the fabs being built today start actually producing chips? Our mythical startup, however, doesn’t exist in a vacuum: it exists in the same world as TSMC, the company who has defined the modern pure play foundry. TSMC has put in the years, and they’ve put in the money; TSMC has the unparalleled customer service approach that created the entire fabless chip industry; and, critically, TSMC, just as they did in the mobile era, is aggressively investing to meet the AI moment. If you’re an Nvidia, or an Apple in smartphones, or an AMD or a Qualcomm, why would you take the chance of fabricating your chips anywhere else? Sure, TSMC is raising prices in the face of massive demand, but the overall cost of a chip in a system is still quite small; is it worth risking your entire business to save a few dollars for worse performance with a worse customer experience that costs you time to market and potentially catastrophic product failures? Becoming a meaningful customer of Samsung or Intel is very risky: it takes years to get a chip working on a new process, which hardly seems worth it if that process might not be as good, and if the company offering the process definitely isn’t as customer service-centric as TSMC. I understand why everyone sticks with TSMC. Share on X (Opens in new window) Share on LinkedIn (Opens in new window) Email a link to a friend (Opens in new window) RelatedAI Promise and Chip PrecariousnessDateTuesday, February 25, 2025Chips and GeopoliticsDateTuesday, May 19, 2020Chips and ChinaDateTuesday, October 25, 2022 Search× Stratechery Plus An Interview with Netflix co-CEO Greg Peters About Engagement and Warner Bros. TSMC Earnings, The TSMC Brake Revisited, Why AI Needs Foundry Competition Ads in ChatGPT, Why OpenAI Needs Ads, The Long Road to Instagram View All Stratechery Plus TVs and Siri ChatbotDithering | Jan 23TVs and Siri Chatbot(Preview) A Call to Action for TSMC’s AI Customers, Wall Street’s Netflix Anxiety, Q&A on Tech’s Cignetti, OpenAI, StarbucksSharp Tech | Jan 23(Preview) A Call to Action for TSMC’s AI Customers, Wall Street’s Netflix Anxiety, Q&A on Tech’s Cignetti, OpenAI, StarbucksJimmy Butler Out for the Year, All-Star Reserves in Both Conferences, Trouble in Paradise for the Buss Family?Greatest Of All Talk | Jan 22Jimmy Butler Out for the Year, All-Star Reserves in Both Conferences, Trouble in Paradise for the Buss Family? Stratechery Plus An Interview with Netflix co-CEO Greg Peters About Engagement and Warner Bros. An Interview with United CEO Scott Kirby About Tech Transformation An Interview with Jeremie Eliahou Ontiveros and Ajey Pandey About Building Power for AI View All Sharp Text Was Nico Harrison Wrong?Friday, January 23, 2026 News in the Era of Irreconcilable DifferencesFriday, January 16, 2026 Notes from Schrödinger’s Cold WarFriday, January 9, 2026 View All More by Ben Thompson View All All Articles View All All Content View All Stratechery Plus An Interview with Netflix co-CEO Greg Peters About Engagement and Warner Bros. TSMC Earnings, The TSMC Brake Revisited, Why AI Needs Foundry Competition Ads in ChatGPT, Why OpenAI Needs Ads, The Long Road to Instagram View All Stratechery Plus TVs and Siri ChatbotDithering | Jan 23TVs and Siri Chatbot(Preview) A Call to Action for TSMC’s AI Customers, Wall Street’s Netflix Anxiety, Q&A on Tech’s Cignetti, OpenAI, StarbucksSharptech | Jan 23(Preview) A Call to Action for TSMC’s AI Customers, Wall Street’s Netflix Anxiety, Q&A on Tech’s Cignetti, OpenAI, StarbucksJimmy Butler Out for the Year, All-Star Reserves in Both Conferences, Trouble in Paradise for the Buss Family?Greatest Of All Talk | Jan 22Jimmy Butler Out for the Year, All-Star Reserves in Both Conferences, Trouble in Paradise for the Buss Family? Stratechery Plus An Interview with Netflix co-CEO Greg Peters About Engagement and Warner Bros. An Interview with United CEO Scott Kirby About Tech Transformation An Interview with Jeremie Eliahou Ontiveros and Ajey Pandey About Building Power for AI View All © Stratechery LLC 2026 | Terms of Service | Privacy Policy |
The most immediate and concerning risk facing the artificial intelligence industry, particularly as reflected in recent earnings calls, isn't the geopolitical tension surrounding TSMC’s operations in Taiwan—though that remains a significant worry. The real brake on AI’s development is the company’s inability to keep pace with the surging demand for chips. Amazon’s Andy Jassy, Microsoft’s Amy Hood, Google’s Anat Ashkenazi, and Meta’s Mark Zuckerberg all echoed this sentiment: demand consistently outstripped supply. This isn't merely a short-term constraint; it’s a structural issue rooted in TSMC’s strategic conservatism and delayed investment. The trigger for this bottleneck was the rapid deployment of AI infrastructure following the release of ChatGPT in November 2022. This unleashed an unprecedented wave of capital expenditure by major tech companies – the hyperscalers – seeking to build out their AI capabilities. However, TSMC’s capital expenditures remained stubbornly flat for several years, despite this surge in demand. C.C. Wei, TSMC’s CEO, directly attributed the shortage to a customer’s answer: “Silicon from TSMC is a bottleneck.” The customer, a large cloud service provider, had planned for this surge five to six years prior. This reveals a crucial disconnect—TSMC was essentially reacting to a future that many others had already anticipated, and it was too late to catch up. The company’s reaction, or rather, lack thereof, had effectively put the brakes on the entire AI buildout. TSMC’s capital expenditures, now significantly increased to $52–$56 billion for this year, are being viewed with cautious optimism, but the timing is crucial. These investments won't begin to significantly impact supply until 2028, and even then, the gap between demand and capacity is likely to persist for several years. Wei’s nervousness regarding these investments underscores the magnitude of the challenge. Building a new fabrication facility—a ‘fab’—takes two to three years, meaning the impact of a $54 billion investment will only be realized in 2028 and beyond. The consequence of this delayed investment is already being felt in terms of revenue foregone. The hyperscalers are essentially waiting in line for chips that TSMC isn’t producing quickly enough, and this translates into a tangible loss of potential revenue. The extent of this loss is significant – projections suggest it could reach hundreds of billions of dollars if AI realizes its full potential, and it's a burden primarily borne by TSMC's customers. The numbers aren't merely theoretical; they represent a concrete, measurable cost. The nature of TSMC’s risk—inherently higher than traditional business models—further exacerbates this issue. Unlike companies with a greater percentage of marginal costs, TSMC’s costs are predominantly capital expenditures, which don’t diminish even if demand stagnates. This makes the company particularly wary of over-investing, and a failure to meet demand would represent a major financial blow. Wei's concern isn’t simply about revenue loss; he's worried about a potential “big disaster” for TSMC. A key element of the issue is the timing. The hyperscalers’ investments were initiated based on the assumptions generated by the rapid adoption of AI, while TSMC’s delayed response means that its capacity is lagging behind. This isn’t a simple supply-side problem; it’s about a fundamental divergence in strategic outlooks. To unlock the full potential of AI, the industry needs competition in the foundry space. Currently, TSMC controls the market, and its cautious approach – prioritizing risk aversion over rapid expansion – is stifling innovation and growth. Samsung and Intel are potential competitors, but they face significant hurdles achieving TSMC's performance and customer service levels. The most effective catalyst for change would be increased competition, forcing TSMC to accelerate investment and innovation. The risk TSMC is managing isn't simply geopolitical; it’s the risk of being left behind by a rapidly evolving industry. |