---
title: "In-depth interview with the founder of Stratechery: Warning of a \"chip shortage\" in 2029, the SaaS model will come to an end, and advertising is the ultimate business loop for AI"
description: "Ben Thompson warned that TSMC's conservative expansion could lead to a global chip shortage in 2029, and giants need to share the risks of building factories. In his comments on the giants, he believe"
type: "news"
locale: "en"
url: "https://longbridge.com/en/news/276001569.md"
published_at: "2026-02-15T10:00:36.000Z"
---

# In-depth interview with the founder of Stratechery: Warning of a "chip shortage" in 2029, the SaaS model will come to an end, and advertising is the ultimate business loop for AI

> Ben Thompson warned that TSMC's conservative expansion could lead to a global chip shortage in 2029, and giants need to share the risks of building factories. In his comments on the giants, he believes that Meta has the best execution; Google is chaotic but resilient; Amazon's self-developed chip strategy faces risks. Looking ahead, he pointed out that after the digital world becomes highly developed, "on-site" experiences may become the most scarce value

**Recently, Ben Thompson, founder of Stratechery and a well-known technology analyst, participated in an in-depth conversation.**

In this one-and-a-half-hour deep interview, Ben Thompson, drawing on his twenty years of experience in technology observation, provided sharp commentary on the computing power bottleneck in the AI era, the evolution of business models, and the moats of tech giants.

## **Core Warning: TSMC's Conservatism and the Chip Shortage of 2029**

As the proponent of aggregation theory, Thompson expressed great concern about the current pace of AI infrastructure development. He put forward a core viewpoint: **The limiting factor for global AI expansion is actually TSMC's capacity expansion speed.**

Thompson pointed out that despite the enormous market demand, TSMC, as a monopolist, has been quite conservative in expanding production. This is because the risks associated with fabs are extremely high; once there is overcapacity, the massive depreciation costs will destroy profit margins.

> “99.9% of the cost of a fab is depreciation... TSMC is actually behaving quite rationally. They would rather forgo potential long-term revenue than take on the downside risk of overcapacity.”

This conservative strategy has led to a mismatch of risks. TSMC has transferred the risk of insufficient capacity to Nvidia, Apple, and major cloud providers, while these tech giants face the risk of “losing future revenue due to insufficient computing power.”

Thompson made a startling prediction:

> **“I believe we will face a massive chip shortage around 2029.”**

He emphasized that the current growth rate of capital expenditures (for example, TSMC increasing from $40 billion to $60 billion) is still insufficient to meet the exponential demand for computing density brought about by future AI agents.

**Suggestion:** He called on tech giants (Hyperscalers) to support Intel or Samsung purely for economic reasons—rather than just geopolitical considerations—or to bear the risks of building factories through prepayments; otherwise, they will be locked into capacity bottlenecks in the future.

## **AI Business Model: Rebranding "Advertising"**

The market is generally concerned with how AI applications can monetize. Unlike Silicon Valley's widespread aversion to advertising, Thompson firmly believes that **advertising is the most effective monetization form for AI**, especially for companies like OpenAI that have huge traffic but lack a commercial closed loop.

He refuted the viewpoint that “advertising will affect the quality of AI answers” and cited a profound insight from Zuckerberg:

> **“Zuckerberg once said that the largest and most successful Agent (intelligent agent) in the world today is actually Facebook advertising.”**

Thompson explained that Facebook's advertising system is essentially an automated agent: businesses input budgets and goals, and the system automatically finds customers and delivers results, which is the ultimate form of an agent.

Regarding OpenAI, **he criticized the current attempts to place ads based on conversational context, believing it would alienate users.**

> "The best advertising model is not based on what you are talking about, but on a comprehensive understanding of you as a person (Profile). Google has a huge advantage in this regard; they don't need to stuff ads into the Gemini dialogue box, as they can leverage Gemini's data to monetize more accurately on YouTube or search."

## **Tech Giants Parade: Who is Falling Behind?**

Thompson provided a commentary on the five major tech giants in the U.S. (The Big Five), with some striking viewpoints:

-   **Meta (Strongest Execution):** Considered the company with the strongest execution. Despite market concerns about its massive capital expenditures, Thompson believes Meta's advertising model is underestimated, and it not only has cash flow but also builds infrastructure through open-source models.
    
    > "The threat of OpenAI to Facebook may be greater than to Google, but Facebook is clearly spending money to address this challenge."
    
-   **Google (Chaotic but Resilient):** Thompson likens Google to a biological structure.
    
    > "Google is like slime mold... Although it appears chaotic and suboptimal, with many redundant actions, this lack of optimization actually gives it great adaptability and resilience. When it finally comes at you, you're done."
    
-   **Amazon (Strategic Risk):** He expressed concerns about Amazon's chip strategy in the AI era. Amazon is used to winning in commoditized markets through low costs (such as launching its self-developed chip Trainium), but in the field of AI, where each generation sees significant performance leaps, not using the strongest NVIDIA chips could lead to a decline in competitiveness.
    
    > "This low-cost alternative strategy works in markets where performance levels off (2005-2025), but will it still work in a new market with significant generational performance leaps? Probably not."
    
-   **Apple (Platform Dilemma):** Hardware remains unbeatable, but software and service platforms perform poorly.
    
    > "Apple makes great products, but is a terrible steward of platforms."
    

## **Future Outlook: The End of SaaS and the Value of 'Live'**

When discussing the software industry (SaaS), Thompson pointed out that if AI leads to a reduction in the number of employees in companies, the SaaS business model based on "per seat" charges will face a growth ceiling.

Finally, regarding the future of AI-generated content proliferation, he believes scarcity will redefine value:

> "In a world of infinite content, **everything 'live' will become more valuable.** Shared experiences, face-to-face classrooms, sports events—these shared experiences that cannot be personalized by AI will be where the premium lies in the future." The following is the interview transcript:

> Participants:
> 
> Ben Thompson: Founder of Stratechery, renowned technology analyst
> 
> John: Interviewer.
> 
> \---
> 
> John:
> 
> Ben Thompson is the founder and author of Stratechery, a communication tool that everyone in the tech industry uses to understand the current situation. He is also a pioneer of the popular paid newsletter model in today's media landscape. For many years, he has been running Stratechery independently in Taiwan.
> 
> Ben:
> 
> Cheers. Great to see you.
> 
> John:
> 
> Over the years, you have frequently discussed aggregation theory and really popularized the idea that in the pre-internet era, power was typically held by the supply side, whereas on the internet, due to factors like marginal cost dynamics, power shifts to demand aggregators. So Booking.com is much larger than any chain hotel, something like that. Booking.com is particularly interesting because they aggregate all hotels, but they themselves are also aggregated by Google. So they are one of Google's largest customers, even though they are on the other end. I think Booking.com is a severely underrated success story in the tech world. They are a European company and relatively low-key in many ways, but if you had invested $1 in both Booking.com and Google 20 years ago, you would have made much more money as a shareholder in Booking.com. I think people don't realize this. It’s a very well-run business. What I want to ask is, how does aggregation theory apply to AI? How does this framework need to be updated?
> 
> Ben:
> 
> To some extent, it remains to be seen. I mean, this is one of the debates that I personally find most anxiety-inducing, which is about the future welfare of OpenAI. A few years ago, I suggested that OpenAI could actually stop making models and still become one of the most valuable companies in the world just based on ChatGPT. That is their most valuable asset. Part of the problem they face is that this is absolutely true in 2023 and 2024. But they have never really built a business model around this. And I think, quite famously, at least from all the tweets I received since they announced they were going to launch ads, I have been grappling with this fact for a long time. I find it interesting, and I actually want to know your thoughts on this. There is a mindset in Silicon Valley that is skeptical of advertising; people have internalized this view that advertising is bad, evil. Do you feel that?
> 
> John:
> 
> I agree. There is an unconscious skepticism towards advertising. You see, I am a subscriber to YouTube Premium. When I see someone watching videos without a membership, I think, what are you doing? So I understand this unconscious reaction. To some extent, Stripe is a bit like an anti-advertising company; we are a different form of monetization. But I don’t know, I personally don’t have much of an opinion on this. I think it is a very effective form of monetization and makes a lot of sense for certain products So I think it's just a matter of personal preference. I feel like you and Stripe are on the same side, and you are skeptical. I think advertising is great, and, you know, I'm kind of endorsing my own viewpoint.
> 
> Ben:
> 
> The reason "Stratechery" has gained immense traction is that I don't hate advertising, even though I'm not in the advertising model myself.
> 
> John:
> 
> Exactly. But you are on a paid model. Well, actually, that's interesting. I actually think that the attention I've gained over the years is precisely because when others weren't discussing advertising, I was, even though it's the most important business model in the tech space. Looking back, all my early articles about advertising were terrible. I had no idea what I was talking about, but just by discussing it, it was very helpful. The reality of advertising is that, first, if you're making a product, like "Stratechery," it's very fortunate; it's absolutely a new model, or let's say a new internet-native model, because my subscribers are spread across about 200 countries. My market is actually the whole world. Stripe clearly helped make that happen. There are also a few in the Vatican, and they are paying attention too. I can check, and I bet I have at least one subscriber in the Vatican. But what I benefited from is that I am a huge beneficiary of social media, especially Twitter. In the early days of Twitter, if you want to put it that way, sharing smart links was valuable. So if I was someone who could regularly provide links that made people feel smarter, they would share, discuss, and interact back and forth. That solved my user acquisition problem. Yes. But the reality is, for content, there's one more thing I'll mention later, which is that it's something to talk about. It's a common ground between us; we can read the same things, have our own opinions about them, and react to them. The products I buy on Instagram, I wouldn't post about them, but they might be very valuable. And theoretically, these small businesses or vendors also have the same opportunity to sell products to the whole world. They just need a way to tell people. As someone who has bought too many products from Instagram, by the way, one of the benefits of moving back to the U.S. is that Instagram's ads are amazing. I think they were good in Taiwan, but much better in the U.S. Oh, that's the best part of living in the richest country in the world. Advertising, it's great. I would think, what is this native content? Give me more ads. By the way, Facebook has been very willing to do this in the past six months. There are a lot of ads these days. But I can always buy things I never thought of, even things I didn't know existed. It's great, truly amazing, a real benefit for me as a consumer. Of course, I subscribed to YouTube Premium and look down on those who haven't subscribed, but I can discover things that I didn't know about that make life better. So, as a user, I benefited. As a wealthy user, I benefited. And for the 6 billion people in the world, most of whom don't have the disposable income I do, let alone you and others in San Francisco, they can't have the same experience as I do So for AI, especially when it offers such high costs and free products are far inferior to paid ones, gaining access is a victory for them. Therefore, for a mission that believes AI makes the world a better place, not accepting ads...

> John:
> 
> No, I agree that advertising is an effective monetization method. What do you think is the right way for consumer-grade AI applications to advertise? For example, ChatGPT just announced they are going to do ads, and they are doing it.
> 
> Ben:
> 
> No, they are making ads a completely separate experience from the answers.
> 
> John:
> 
> No, that's why it's so bad. That's why I'm so disappointed in them. What they are doing is the easiest, simplest solution, basically banner ads.
> 
> Ben:
> 
> It is banner ads, but they are based on conversational context. The problem is, they published their advertising principles, right? "Our ads do not affect your answers." If you target ads in the most likely way, that is, based on conversational context, we will show you a roughly relevant ad. First, your market is much smaller because you have to hope someone starts a conversation that matches the ad inventory you have. Second, you will get into a situation where if the ad is obviously related to the answer, it raises user skepticism about the connection between the two. So I would prefer ads to be unrelated to the answers. The way to do this is to build a meta-style user understanding and show them things that are relevant to them. Just like on Instagram, the best Instagram ads are completely unrelated to the content I'm browsing, but come from Meta's broad understanding of me.
> 
> John:
> 
> So you're saying AI ads should be more like Facebook ads rather than Google ads? Right now, the focus is on doing targeted ads related to prompts, but it should be about analyzing the user, analyzing who this person is, and what their interests are?
> 
> Ben:
> 
> Yes, I think that would be better. It would reduce conflicts of interest and reduce user uncertainty. And I think, I'm not the biggest fan of search ads, precisely because they work too well. I think many people are confused between natural results and ads; they cannibalize each other. Why do I have to buy my own name in search ads? Because others will... you will harvest an ad click that could have been a natural result, which is fine, search provides a lot of value. But the challenge is obviously that in ChatGPT, they only have one inventory space.
> 
> John:
> 
> The defense of Google ads is that everyone complains about brand searches, you're paying for cannibalization, but Google is very focused on search quality, so the sponsored listings themselves also have a ranking, a relevance ranking. So it's like the Yellow Pages; you need to pay to be listed in the Yellow Pages. That's fine. I'm not, again, I'm an ad enthusiast. I just think Meta's ads have broader value because they show me things I didn't know I wanted
> 
> Ben:
> 
> But if AI applications are going to generate your portrait, will that portrait include all the content of your conversations?
> 
> John:
> 
> That's the problem. So Deus comes out and says, "Wow, I can't believe they added ads. We won't do that." It's funny because the whole Gemini Deep Mind system, what funds it?
> 
> Ben:
> 
> Of course. It's funded by Google's advertising machine. This might actually be the ideal model. It's actually quite interesting. Last year I was in New York. I was meeting someone in a shared office, or across the hall was a hedge fund or something. They came over and said, "I'm a long-time reader. You're responsible for one of our worst decisions." I said, "What?" He said, "Investing in Twitter." I said, "I never said to invest in Twitter. It's always been a terrible company. I don't even cover them anymore because the business is so bad." Then I remembered what it was. It was when they acquired MoPub. My theory is, I \*\*have always been concerned about Twitter ads because they are particularly text-heavy. I don't think text is the best interface for advertising. Clearly, visuals are usually better. There's also a posture issue. If I'm on Twitter, I'm ready to fight, I'm engaged, I'm searching for information. If I'm on Instagram, the whole point of seeing ads is that I don't really care what I'm looking at right now, I'm just killing time. You're actually in a better posture, like watching TV, you can absorb ads. And Twitter is not good at that. But because Twitter is at least theoretically an interest-based network, it should be able to know more about you than theoretically through pixels and SDKs. So my theory about the MoPub acquisition is that I thought it was a great acquisition because they could leverage Twitter's signals to perform in other apps through the MoPub network. Now Twitter is incompetent, they did nothing with MoPub, sold it to Apple, and Apple is now writing MoPub as the world's top, but that was my argument. I think this can also apply to AI. I think the ideal outcome for Google is that they never put ads in Gemini, but learn a lot about you through what they do in Gemini, and then monetize through ads on YouTube, Google Search, and other assets. And the challenge OpenAI faces is that they only have one place to put ad inventory, which is ChatGPT.
> 
> John:
> 
> So you're saying Google can use Gemini to improve its ad targeting across all of Google's assets, and then if you want to put ads in Gemini, maybe...
> 
> Ben:
> 
> \*\*I think you never need to put ads in Gemini. But if you do, you definitely have the portrait of you that Google has collected from the web, and you can... and you don't need to show those ads that make users feel weird, because why show ads about the questions I'm asking? \*\*
> 
> John:
> 
> But according to the scenario you just described for Google, could there be the same issue? For example, when targeting is too precise, people might worry about the conspiracy theories surrounding Meta listening through microphones. Don't you think there could be similar issues? If you use Gemini's data to create better ads, the targeting might end up being too precise, and people would find it strange?
> 
> Ben:
> 
> I think that's a fictional concern. No, it really is. But people do have that concern. If data could really achieve that, it would be a "sweet trouble" that every tech company would love to face.
> 
> John:
> 
> I get it. This is something people say when you have very good targeting. I think a lot, to be honest, I think there are many discrepancies between verbal preferences and actual preferences regarding these kinds of things. The reality is that people like, even Americans would think, you can pay for Facebook, or people would rather see ads. I think most people don't care. This is true in many tech fields, and it relates to the skepticism you mentioned about ads. It's an elite circle. Regulators in Washington, everyone is thinking about these very theoretical things.
> 
> Ben:
> 
> Isn't this also one of the challenges of banner blindness? Instagram ads work so well because it's a photo feed that shows you images. Some images look like ads. So it's like Facebook, for AI applications, you're looking for answers, you don't want to see banner ads.
> 
> John:
> 
> No, this is a huge concern. This is one of the great ironies of Meta/Facebook, that Mark and everyone hates Apple for many reasons, which I think are very reasonable, but Apple saved Facebook from itself. Remember back in the day, the Facebook platform? And things like Facebook Payments. Mark always wanted to build a platform. If you're just an app, you can't build your own platform. The problem is, I think the ad-based model is often incompatible with the platform model. The whole point of a platform is to let other things shine, to let other things emerge. You're just the supporting structure, letting other things take over. So the operating system, ideally, focuses on the applications running on it. When Facebook was forced to not be a platform, but just an app, suddenly they could fully focus on becoming an advertising company. Think about Facebook ads, even back then, they were either feed ads or story ads. Literally, your entire device is an ad. But somehow, it's not a banner, a little thing on the edge. They really got users' permission to take over your entire device and show you a full-screen ad every five seconds. That's amazing. Yes. And they were forced into this by Apple.
> 
> Ben:
> 
> Okay, this reminds me of a point, and I'm curious about your thoughts. That is, many times, tech companies become very large simply because the core idea is better than the founders realize. So Meta is a very large company because they have a feed, the feed became very large, and they were very smart along the way, acquiring Instagram, they achieved incredible precision targeting and good monetization, But the result is that people spend a lot of time, many people, you know, that information flow's PV×Q, they have monetized it very well, and that's why they have become so huge. The same goes for NVIDIA; as a result, the GPU market has become very large, and they sold a lot of GPUs. So, perhaps the founders, because they are usually high-energy people, want to have many new ideas, they often think about what the next thing is, what the second act or third act is, everyone wants to invent AWS. But I'm curious about your thoughts on this idea: generally speaking, it's about making the core business very large, and the order of magnitude at the top is much more than you might imagine.
> 
> John:
> 
> Yes, I think that's always the case. And I think sometimes people end up doing things they didn't originally want to do, and then they keep pushing back. I think Meta is a perfect example. My impression is that Mark is not very interested in advertising. He had very talented people along the way helping him build these advertising products. I think Meta has suffered because he hasn't been at the forefront, advocating that advertising is good, that advertising is beneficial to society, that they are the driving force behind all consumer surplus generated by technology. The president and the average person on the street use the same search engines, the same AI, and so on. That's because of advertising... maybe not, the president might be using something like Palantir's search engine. Yes, it might be worse. Fairly speaking, Google has fallen behind. There is too much junk online. But would Trump search? I don't know. That's a good question. But he hasn't made that argument. I think Meta has suffered for failing to make that argument. Then you get things like "we're going to do the metaverse," "we're going to do XYZ," always wanting to go back to "we want to be a platform, be a platform." And Meta is an entertainment company. I wrote about this a few years ago, and I think it's a judgment that is both good and bad.
> 
> Ben:
> 
> Do you remember that famous quote by Paul Krugman? He said the internet wouldn't be very big, or it would be a fax machine, because people don't have that much interesting stuff to say. I actually defend that statement because it is actually true. Most people really don't have that interesting things to say. I brought up that quote around 2015, saying it was a fundamental limiting factor for Meta's long-term potential. As long as they still think of themselves as a social media company, they will encounter a problem that over time their information flow will become less interesting. Now, from P2P content to... well, if I may boast, that's a very clever insight. A bad insight is my suggestion that they need to collaborate more with professional content creators, like funding websites like Buzzfeed, sharing revenue, etc. User-generated content. **The real answer is what TikTok does; TikTok is not a social network at all. It is an aggregator and a platform like YouTube, the same idea. Like personalized television.** Yes, what really matters is that this is a key point where people get tangled up; relative numbers are important, but absolute numbers are even more important. So, if you have billions or trillions of content, even if only 0.1% of the content is good, it's better than having 100 pieces of content but a 10% hit rate The latter is actually worse, even if your hit rate is higher. So stimulating a large amount of creation, writing algorithms to capture good content and push it up, actually solves Paul Krugman's fax machine problem. Facebook was caught off guard by this. They were so entrenched in their identity as a social network that they allowed TikTok to take such a large chunk. This was their blind spot.
> 
> John:
> 
> Interesting. I've stated my point of view. So, is Stratechery widely read in Washington, D.C.?
> 
> Ben:
> 
> Yes, yes. Sometimes it's gratifying. When you get a call asking for your opinion, or you receive certain responses, or see the influence, it feels great. When you're criticized and people are angry with you, it's not as gratifying. But fortunately, the key to success online is that I have a lot of what I call a very high level of "not being agreeable." So you can criticize me all you want, I won't change my mind.
> 
> John:
> 
> Okay. But back to how aggregation theory applies to AI. One simple viewpoint might be that AI applications are the new aggregators, and therefore huge economic value will accrue to them, and that's it. You could also argue that this is overly simplistic in many ways, because, as we said, Booking.com should return to you the hotels you should book, but your expectations of the commercial motivations of AI applications might be a bit lower; it's a more abstract technology, and it's not easy to insert all the commercial motivations in the right way, and you would raise various objections.
> 
> Ben:
> 
> Well, I think the advertising model might be a starting point, as I mentioned earlier, it's the difference between leaning forward and leaning back. Advertising is closely tied to human psychology, and relates to what you dig into, people's reactions, and how to do something creatively. In the short term, technology often makes old business models stronger before it eliminates them. So you have something like, suddenly you're a newspaper, and I was limited to my geographical area, and now I can reach the whole world. Yes. Oh, wait, a few years later, everyone can reach the whole world. Yes, I'm in pure competition, I'm doomed. This is definitely a concern of this model. If you enter a world like agency business, where agents just buy the right things, I think this is also something driving the tech world's skepticism about advertising. People in tech tend to be quite nerdy and obsessive. They do a lot of research to find the exact right thing. Yes. Why let someone else tell me what to buy when I've already researched for two hours? That's right. But what if this obsessive deep research is now easy for everyone because AI is helping them do it? Then where does the function of advertising lie? I think this is definitely a "be careful what you wish for" situation, because it certainly means more transparency, more detail, more understanding, which sounds nice. What it actually means is pure competition, which is a very brutal game that can destroy entire categories. This is basically what happened in the newspaper industry. So that's the first point. The second point is that in such a world, you inevitably anchor on any measurable specifications or parameters. You have Steve Jobs' old adage about "speed and power consumption" and what the intersection of those means with product feel It actually says that some things are immeasurable and cannot be put into an Excel spreadsheet. Everyone you talk to acknowledges this, saying yes, some things cannot be measured. But how does it manifest in practice? Only what is measurable is important. I think this has become a huge issue in sports analytics, which is a great example. Basketball is my favorite sport, and there are many factors that influence winning, but they are hard to quantify. Baseball is very measurable. I do think there are things like performance in key moments, and I don't know if they are measured correctly. But in basketball, there is no doubt that the way teams interact, how your effort or offensive involvement affects defense, etc., are very hard to quantify. You see it time and again, just take my favorite Daryl Morey, I think there are reasons his teams haven't won. They over-optimized and sacrificed other issues. If it can't be measured, it often gets undervalued. In a world where everything is mediated by AI, how many unmeasurable things will be discarded? Because what we end up with is very utilitarian, soulless products. In some ways, this is a bit of a silly concern. It sounds silly, but I'm a person, and I expect to like and prefer a variety of humanized things in the long run.
> 
> John:
> 
> But you could also say that e-commerce aggregators like Amazon and others have led to quite anonymous manufacturers producing many everyday goods, like Amazon Basics, which are much cheaper than before but still of good quality. Isn't that a good thing?
> 
> Ben:
> 
> This is where you throw my advertising argument back in my face, which actually raises everyone's baseline level. For the average consumer, the variety of goods they can access, there is no soul in Amazon Basics, and that's okay. Everyone looks back and says, oh, my washing machine from the 60s was much better. But that's true, while there are far fewer people with washing machines now. So my argument now is exactly the opposite; I can let you play the other side.
> 
> John:
> 
> Exactly. You can switch positions. You mentioned agency commerce, and we are obviously very invested in this and announced our news with OpenAI in October. How do you think agency commerce will develop?
> 
> Ben:
> 
> I mean, the comparison between your own OpenAI release and Google's release is quite interesting and illustrates what each company is pursuing. **OpenAI wants to be the place where you do everything. They want to be the aggregator. Critics might compare them to Netscape. If you are a skeptic of OpenAI, I think a better analogy is AOL, as they want to be the interface for everything you might do, with everything going through their channels.** And Google, like they are relative to AOL, actually wants to empower everyone to know that if everyone has the capability, we are the biggest beneficiaries because we still aggregate front-end traffic demand in this regard. So how does this manifest in business specifically? The interesting thing about technology is that I don't think it will manifest in airline tickets, which is everyone's example; no one can think of a better example than this But what will AI buy? What will it get? I don't know. I think people will retain control over their purchasing decisions. But then again, we have assistants, whether in work or other areas, who will make some purchasing decisions that we may not necessarily be involved in. I think this is a good predictive indicator, suggesting that in an ideal situation, I really need to know? In fact, I have strong opinions about what kind of tissue to buy.
> 
> John:
> 
> That's exactly what I wanted to say, but once I set preferences, can this be monitored and automatically executed?
> 
> Ben:
> 
> Well, I don't know. That's a very unsatisfactory answer, except to say it has a significant impact on advertising and similar things, on whether future business models are viable, what kind of profit margins there will be, whether there will be perfect competition, and so on.
> 
> John:
> 
> Okay, let me try to present my view on agency commerce, and I would love to hear your comments. I think some skepticism is triggered by a very distant end state that people depict, which is having a lot of agency autonomy. For example, "Help me book a honeymoon to Japan and all the activities," which no one would actually do. In reality, you should start from some very basic building blocks from the ground up. The first step is just to replace filling out web forms. That's a frustrating activity that no one likes. Imagine you found a winter coat you like, copied the URL to ChatGPT, and said, "Please help me buy this." That would be a much better experience than clicking around on a website you've never been to. So this is just agency using tools on your behalf, and anyone can create this. Maybe it will clarify, "There are multiple colors, which one do you like?" But it just replaces filling out form fields. By the way, this is one aspect I'm very optimistic about, which I call "instant UI."
> 
> Ben:
> 
> Exactly. It's a better UI.
> 
> John:
> 
> Right. So this is like the first level, a better UI to perform the actions you want. Okay. Then the second level is better discovery and search. We have relied on keyword-based search in e-commerce to get to where we are today, which is crazy. Keyword-based search works well when you know the title of the book you want to buy. For example, I want to buy this specific book. But for a winter coat, I don't know, I want a down jacket, what is that called? So, you want to be able to say, "I'm looking for a jacket of this type, I'm going to this place, and it will be this cold. You know, I like these styles and so on." So the second step is better search, being able to search with parameters; existing search UIs don't allow you to specify the temperature of the place you're going to in order to get a jacket that is warm enough. But this is clearly one of the core elements for jackets. So from our perspective, a better search UI is the second level, and I think this is already becoming apparent.
> 
> Ben:
> 
> Exactly. In the early use of the ChatGPT purchasing experience, I think this is one of those super cool features.
> 
> John:
> 
> Then at the third level, what we haven't really seen yet is this idea's re-manifestation, that is, a persistent profile of users that can predict their needs. That's right. Just like I would like to fix my favorite things while browsing, or maybe I could share my browsing history, or perhaps I could share a Pinterest board with some styles I like, "recommend me a good warm winter coat based on these," "here are my photos," "based on this," and so on.
> 
> Ben:
> 
> Well, I even have a better idea. Imagine if you were using ChatGPT around October 1st, and there was an ad showcasing a great winter coat that would be perfect for me. Because it has been learning about my interests, it understands my context. I'm not searching for winter coats because I'm not good at planning. The weather will get cold, and then I will search for winter coats. But what if it could predict that and show me the ad at the right time when I need it?
> 
> John:
> 
> Well, maybe the fourth level is that. That's exactly what I've always hoped they would build. That's my whole argument. That's why they are so late. If they had started advertising products two or three years ago, this should have been launched this year. This is possible today. That's what Meta advertising is about. You need to watch more Reels. This year, I bought more ski gear that I didn't need just because it kept appearing. I'm moving back to Wisconsin, so I'm buying things for the house. Then I saw those ski rack hooks, and I thought that must be great, sounds useful. They're still in the box; I really should hang them up. But yes, experiences like banner ads have limited capabilities, while search is incredibly powerful. But I'm curious about your thoughts on the first level (just the checkout action), the second level (better search), and the third level (defining your own preference embedding space).
> 
> Ben:
> 
> No, I completely agree with this approach. I just feel like you're underestimating how much the third level has already been established. In fact, Mark Zuckerberg said something I think is very profound in a quarterly earnings call a few seasons ago: we get caught up in the definition of technology, like what is an agent? He said, actually, the largest and most successful agent in the world today is Facebook advertising. That's the right perspective. Facebook advertising, people in their minds think you're going to set demographics and targeting and all that. No, it's very automated. What you're doing is saying, the cost of acquiring a customer is worth $10 to me, I'm willing to spend up to $10, and they will deliver a customer to you for $10. Yes, their profit margins will actually increase because they will ensure it's delivered at exactly $10. They can spend more and actually make more money. You get what you want. I think to what extent this has already been very powerful, we haven't realized it yet. They just get caught up in "50% of my ads are effective, but I don't know which half."

John:

> No, they all work on Facebook.
> 
> Ben:
> 
> Yes.
> 
> John:
> 
> I think there will be a batch of very successful new companies born in the AI-driven e-commerce space. It feels like a sufficiently different product space. But are you referring to retailers, merchants, or discovery and demand-side? It could also include retailers.
> 
> Ben:
> 
> Yes. I mean, I certainly think that the new part, perhaps what you just mentioned, is this real anticipation. Now, back to Meta ads, it's amazing how it helps merchants with very specialized products find customers they would never have found otherwise. But conversely, I need a very specialized product, how do I find it? I think you mentioned this before. But to what extent is this not just about immediacy? I need this specific thing. I remember I needed a server accessory, a rack accessory to install this router because I had an extra one, and I didn't want to buy a brand new one or something. I had this extra router, and of course, there was a person in Australia on Etsy who did 3D printing, perfectly matching this, which was awesome. I found this random person, and I'm sure he sold me a $40 part for $2, making a lot of money. Good for him. But what if an AI could predict this kind of demand? So it's not "Oh, I have a need, let me go find it." Yes. But rather "I know you will need this, let me get it for you." That would be very powerful.
> 
> John:
> 
> The public market shows that by January 2026, the SAAS (Software as a Service) model has ended. Are they right?
> 
> Ben:
> 
> I think it might be mixed. One of the great things about American business, and this is one of my theories about why Europeans are so keen on data privacy and regulation, is because they often deal with European companies. Yes. So, a few years ago, I was in Paris, of course, for tourism, going to the Louvre or other museums. They all have their own developed registration systems, they collect so much data, and they’re like, yes, your age? Your pet? What color is it? Why do they need to know all this information? And it's all non-standard forms. This is where you need AI to fill all this in. They have a theoretical idea in their minds that if we capture this data, it might be useful, so they built these internal systems in the 2000s, which are very insecure. When I used them, I thought, where are the regulators? Who can manage this, it's ridiculous. So I understand that mindset. American companies don't do that. American companies are very good at, and I think one of the great advantages of American business culture is understanding, and I personally think this too. This is the life advice I give. What is the biggest mistake people, especially when young, make? They focus on their weaknesses. I have to improve my weaknesses. No, what you need to do is double down on your strengths. You will reap rich rewards from that, and then you hire people to handle your weaknesses, right? I am a faithful believer in the GTD system. Great book, "Getting Things Done." Even if you don't use this system, the book is really good. There are a lot of great insights There are various things, such as reminder documents, etc. This is an amazing system. I can't manage this system by myself at all. So there's a Mac application called OmniFocus, built entirely around this system, and I don't have a license. My assistant has a license. I text him things, and his job is to maintain my GTD files because I can't do it. What do I do? Actually, my wife is very optimized. I write three articles a week, do one interview, and do three podcasts. All my focus and energy need to be on this. If I succeed, it will make a lot of money, and I can spend money to solve all my other problems. I think American businesses do this very well. They don't waste time and energy on things they are not good at. They double down on what they are good at, focusing on the upside rather than their cost centers.
> 
> John:
> 
> It might be a result of the very large market in the United States.
> 
> Ben:
> 
> I think so too. And there's competition in a very large common market. So you don't have, you go back to newspapers, they have a lot of internally developed things. If you are an online publication, like me on the internet, I make money by commenting on large tech companies. This might be the most competitive market on Earth, right? There are many opinions about large tech companies. So you have to be very focused, and considering this, it illustrates the lasting value of paying to have people manage these business functions from a software perspective. Now, there are many SaaS applications. I'm not sure if they are all strictly necessary or if they are all worth the money. I think when people talk about technology, I can say there are "six big ones." The "six big ones" are Silicon Valley companies, basically a cookie-cutter venture capital model, finding this founder to solve this specific business case with a SaaS business model. Everyone loves to talk about changing the world. This is actually the most predictable thing. That's why venture capital returns have been compressed, but at the same time, it's also because they are very predictable in terms of engines. A big problem is that they are all seat-based. Anything that is seat-based, there is a likelihood that the seats will decrease. If the alternative is more small-scale products, ideally, the internet overall, writing or content is a good example. In the past, you wanted to be in a big pond, where everyone in the big pond could share a piece of the pie. If you worked at Condé Nast and wrote for one of their magazines, you lived well even if you wrote just a little. Today, if you want to become a content producer, I've always advised people. I say, look, you don't want to be in the same pond as me, right? Bill Simmons was like the first internet sports writer. You don't want to mimic Bill Simmons on the internet because he got there first. What you want to do is create your own pond. The internet makes it possible to create millions of different ponds. So you can define your own pond and become the only big fish in that pond. That's how you succeed. If AI makes this possible, I think that's an optimistic scenario, where AI enables not just in the content space but in all types of businesses to have more small-scale individual entrepreneurs or small teams that are not well suited to a seat-based model driven by Salesforce. So there might be a trend back to self-service, or maybe they will take matters into their own hands because their needs are not that large So that is a larger structural change. But the problem is that it's easy to say "companies will be fine." If you eliminate growth, that's a big problem. I think that's the root of all the compression through headcount growth, which is growth itself. If these are just stable companies with astronomical stock-based compensation built on the premise that we will become very large...

> John:
> 
> I can think of two criticisms you have about the software space and why everything has been sold off. **One is that everyone will use Claude Code to internally rebuild versions, so the software model will diminish. The second is that many products are priced per seat, so if (client companies) headcount growth slows or even shrinks...**
> 
> Ben:
> 
> Exactly. The first point, Anthropic just adopted Workday, which is well-known. That's right. So I don't think, you know, Claude is a record system, that's that category. People don't seem to be quickly using Claude Code to rebuild one of those record systems. We see this at Tribul as well. I don't think anyone will quickly code one of those record systems with Claude. Right now we use Workday. I don't know how to evaluate the second criticism, but similarly, for a very broad and deep record system, it's hard to argue that its business has weakened compared to a year or two ago, right?
> 
> John:
> 
> But that's also my point. **Saying they will go to zero is wrong. But if the assumption is that you're doing fine, but you won't grow infinitely.** This shift from being considered a growth company to a stable company, I see that as the reason valuations are being cut, along with the entire compensation structure.
> 
> Ben:
> 
> Exactly. You're now being valued on EPS rather than revenue. So, yes.
> 
> John:
> 
> Can we talk about your business? Of course. You are a pioneer of the independent writer model. I think you were one of the earliest paid newsletters.
> 
> Ben:
> 
> I think so too. There are two pioneers worth mentioning. One is on Wall Street, with a long history of fax newsletters, like Grant's. The difference is that those were very expensive, and the market was very small. So, the difference with Stratechery is that it's much cheaper and the market is much larger. Another person worth mentioning, I think he was the first to do this before me, otherwise I would consider myself the first, is Andrew Sullivan, who did paid news.
> 
> John:
> 
> I didn't know he did paid news.
> 
> Ben:
> 
> He did it for about a year. The problem is he did it all wrong. He completely did it wrong. He posted 50 posts a day, right? About countless different things. He totally burned out, and so on. But that fit the advertising model at the time because you would always go back and look, there was always something new, I'm sure he brought huge traffic to The Atlantic, especially while he was there. He later went independent and was quite successful. I think he probably made around a million dollars But that was a paywall full of loopholes, for example, you would encounter a paywall after more than 35 posts, which feels like punishing your most loyal users. And it's easy to bypass. However, he actually inspired me a lot in thinking about this model because he was criticized as a failure for burning out and giving up. But I think he made a million dollars, right? That's pretty good. From the beginning, I thought about this mentality. When I started Stratechery, he had countless ideas to write about, while I limited myself to writing a maximum of twice a week. The reason is that I already had the subscription model in mind, and when I added the paid model, I didn't want to give the impression of "I'm taking something away, now you have to pay." What I thought was, "You love it so much, if you pay, you can get more." So I always wanted you to pay to get more. I think this might be more important at the beginning because this model is new. You know, the metric I looked at was those who visited Stratechery on days I didn't post because they were the ones hoping I would post that day and then leaving disappointed. So in this case, the paywall usually disappoints those who encounter it. And in this case, the paywall would eliminate their disappointment because they can now get what they want. So I thought, if I could capture X% of those visitors, that would be great. Daily goals, weekly goals, monthly goals, none of them were met. As a result, I thought it wouldn't succeed, and I would have to go back to teaching English or something. But it just kept growing, growing, growing. At six months, it actually reached my one-year goal of 1,000 subscribers, 1,000 die-hard fans, you know, that's $100,000 in annual revenue.

John:

How long did it take to get to 1,000 subscribers?

Ben:

Six months. Okay. I posted a little note saying, "Hey, the model worked. My one-year goal was 1,000, and now I've reached it." That was the only time I saw a leap in subscriber numbers. In the next 24 hours, I gained 250 new subscribers. A 25% increase. Who were those people? They were those who wanted to subscribe but didn't believe I could make it, afraid I would go bankrupt and run away with their money. Once they realized I wasn't going to disappear, they all subscribed. So my metrics were actually correct, but I didn't properly account for people's uncertainty and fear of losing money. So I'm grateful now that people subscribe to things anytime and anywhere, right? You know, I might be best known for Stratechery, but I'm also proud of this model, and many people make a living doing this.

John:

How far do you think this model can go? Its defining feature seems to be unbundling; for example, 30 years ago, you might have been writing for a publication, and now it's a direct relationship with subscribers, directly monetized, usually paid. Of course, there may also be some ad-supported components, but mainly it's paid. Clearly, Substack has proven that this has very broad applicability, but how far do you think it can go compared to traditional media bundling?

> Ben:
> 
> I think there are a few interesting points here. First, I believe that people, including those in the tech industry, seriously underestimate the enormity of the internet. When I announced my paid product, the biggest opposition I received came from venture capitalists. I won't say who, but "Love you, Ben, but it just doesn't work on the internet." And regarding my previous comments about ponds, I don't know if we've reached the limit of the number of ponds that can be built in the world, and you can occupy these ponds. Another part, a key point, is that AI is actually an important factor here; the key to this model lies in your costs. So, just as technology enables you to reach everyone, you need to leverage technology to keep your costs very, very low. So in the initial years, it was definitely just me. As long as I could support my family, I was fine. This is the problem with traditional media companies. Their cost structure is not the same as the internet's cost structure. They are built on a much higher revenue base. It's interesting to talk about this because a lot of it doesn't really apply. Like I’ve written a lot about advertising before, and I'm not in the advertising business. In this case, I've written a lot about venture capital and high-growth companies, but my actual business is very boutique, small, and artisanal.
> 
> John:
> 
> Exactly.
> 
> Ben:
> 
> In this regard, a very important point is: manage your costs. If you manage costs properly, then the possibilities are endless. But this also means that some things don't fit this model. For example, traditional news investigations that require six months of research. It doesn't get good support under this model. What supports it is the bundling model, where many different writers come together in one publication.
> 
> John:
> 
> The bundling model benefits every participant, but now no one wants to be part of it. Television is a classic example. Why do we have TV bundling? Because in Pennsylvania, there’s a TV station in Philadelphia, and there are towns in the Allegheny Mountains that want to receive the signal from the pay city, but the signal is poor. So they banded together to build a big tower to receive the signal, and from that tower, they actually cabled it to all their houses. That's how cable television started in small towns in rural America, to receive TV signals from big cities. Then Ted Turner came along, thinking he could broadcast directly to those towers, which was great. You got this model, and suddenly, but you had geographical constraints, and you ended up with all these companies that had the best business models in the world. Everyone pays them, whether they watch or not, and they make countless money. What happens once they can do something different? I can stream directly. Business is like that; it’s like being forced into the game, which is optimal from a game theory perspective, and once you can leave, everyone will leave, even if staying is the best option.
> 
> Ben:
> 
> Yes.
> 
> John:
> 
> So I wonder how this applies to your world? Because theoretically, there should be bundling. Substack should become a bundle. You should be able to pay one fee to get everything. But you know, I think Substack made a mistake; to clarify, I’m a huge fan of Substack I was a bit angry before because they positioned themselves as completely writer-friendly. I think that's a mistake because ultimately they can't be writer-friendly, as the most writer-friendly setup is running open-source software on your own server, where no one can do anything to you. And by definition, I thought you would say the most writer-friendly thing is having a thriving consumer business. But the problem is that all their initial terms made bundling impossible, with all individual publishers having their own subscribers, their own Stripe accounts, and so on.
> 
> Ben:
> 
> But wait, isn't there a well-established path in the tech world? OpenTable started out as local software for restaurants, and then they added a consumer discovery layer on top. Shopify initially just provided solutions for merchants, and then they added a consumer-facing network layer like Shop Pay. Yes. But even for those businesses, the Shop Pay part is nice, but it's not the driving force of the business. It's a cool part, I like it, but ultimately the vast majority of Shopify interactions come from me seeing an ad on Instagram, then clicking through and hitting the Shop button, which is incredible. But doing that makes Shopify much more attractive to merchants because you get the Shop Pay network. I suspect that's the driving force of the business; I think it's more of a nice-to-have.
> 
> John:
> 
> But can't Substack add a Substack Prime bundle on top of that? Writers can choose. The problem is that those writers who would make the bundle valuable have no incentive to join because they can earn more by monetizing directly to users. Imagine I'm on Substack; how much extra income does Substack have to give me for me to take my $15 per month from each subscriber and exchange it for a smaller amount from everyone participating in this bundle? So they really have to buy me off to join, and meanwhile, those writers without subscribers would certainly want to join. So that's why the geographical coercive factor seems solvable to me; I feel like it can be solved now. No, sorry, it feels solvable now.
> 
> Ben:
> 
> Yes. I think a counterexample is something like Spotify. Spotify could be said to be the best bundle on the internet. But the reason they can assemble this bundle is that they only need to negotiate with four entities. So interestingly, on one hand, this limits Spotify's upside because these entities can take a big chunk of Spotify's revenue. On the other hand, it's also why Spotify is possible because they only need to negotiate with four. Yes. Yes. And if you try to negotiate with every artist on the planet... of course, I have to deal with Taylor Swift. Good luck with that. All the small fish will sign. Music is especially unique because once a song is out, it becomes part of the catalog, and people actually only listen to old songs in the catalog. So in that regard, it's a particularly unique industry. But that's a formed bundle But I think it's because there are only four players.
> 
> John:
> 
> Okay. How do you use AI when writing for Stratechery?
> 
> Ben:
> 
> I think it has replaced a lot of what I used to do frequently, which is more efficient Google searching. The articles I feel most accomplished about are the ones I write on topics I usually don't cover, and then people from that industry say, "Wow, that was really well written." Because you always have imposter syndrome, thinking, "Oh, I hope I didn't get it wrong."
> 
> John:
> 
> I don't have imposter syndrome, but what's the mechanism when you're at work? When you read something about yourself, do you say, "That's completely wrong," and then you don't trust everything else? Yes.
> 
> Ben:
> 
> I don't want to trigger anyone's memory loss, so I will... People ask me, and I hate that question about books, "What books do you read?" I read a lot of books, but very selectively. I read very quickly, and sometimes when I need to write an article, I know there's a relevant book, and I will read the whole book in the morning to get the background. But overall, I really want to make sure I fully understand the new areas I'm writing about. That's also why I have a significant competitive advantage. I've been thinking about technology since middle school, and I've been writing about tech for 13 years. Yes. So I've done a lot of groundwork, and it's hard for anyone starting from scratch. So I want to dive deep. I'm one of the greatest Google searchers in the world. I think I know all the parameters and how to find information. So I can say quite authoritatively that Google search has gotten worse. I don't think it's Google's fault; I just feel it's harder now. One thing is Google's fault; they've become too biased towards freshness, so you have to be very diligent. But AI is incredible in this regard, for understanding the background, ensuring you grasp the context of a question, how things work. You can query and dig deep. So that's definitely my number one use case at the moment. I don't always use it, but sometimes I let it... I like ChatGPT; I input BBedit as an integration. That's also something that annoys me; I'm very sensitive to fabrications. "Oh, that's great. No, that's not what I asked." I want it to really go in and find things. So I don't use it to generate any exact content.
> 
> John:
> 
> Okay, so it's targeted research and then critique?
> 
> Ben:
> 
> Yes, those are the two main use cases.
> 
> John:
> 
> You've written a lot about the "TSMC bottleneck," arguing that the limiting factor for all AI expansion is essentially the speed of TSMC's capacity expansion, as the vast majority of AI chips are manufactured at TSMC. It seems that everything interesting in the AI field is largely constrained by chips. But that's not inevitable; it could also be limited by power, etc. But currently, it's constrained by chips, and many people want to expand quickly, AI labs, NVIDIA, etc. And TSMC is more conservative in its expansion. Why is that? Why aren't market signals prompting them to build wafer fab capacity faster? Ben:

Because the risks associated with semiconductor fabs are fundamentally greater than anyone else's. You spend billions, hundreds of billions of dollars to build a fab, and if it is not fully utilized, if you end up with excess capacity, first of all, all those costs are locked in. 99.9% of the costs of a fab are depreciation, which you have to pay regardless. So fabs can be extremely profitable. TSMC's profit margins are higher than ever, but they could also quickly find themselves in big trouble. Once built, these fabs can operate for a long time. So excess capacity will depress prices in the coming years. We've seen this in the memory industry. Memory is known for its cyclicality. Believe it or not, in a few years, we will have too much memory capacity because it is currently too scarce. Micron just announced the construction of a huge new fab in Singapore. Everyone will do this. Why does this happen in the memory industry? There are three competitors in memory; if Micron doesn't do it, SK Hynix will; if SK Hynix doesn't do it, Samsung will. So you have this dynamic, a healthy dynamic that the fabs themselves are more aware of, but they can't control themselves, so they take on the risk of building these fabs. The issue we have with logic chips is that TSMC doesn't have that pressure, so they actually behave rationally. TSMC is forgoing potential long-term revenue. Yes. But the downside risk of the fabs is too great; they don't want to take on that risk.

John:

Can't they pass the risk onto the customers? For example, "You will pay for the entire fab"?

Ben:

That might be the goal they need to achieve. So Apple did a lot of this prepayment when TSMC was expanding massively in the 2010s. They might need to be clearer about this. But I think a better solution, a long-term cheaper solution for hyperscale computing companies, is to do what is necessary to make TSMC better. Then you get it for free. Yes, you don't need to prepay. So there is this risk, the risk of overbuilding. Now TSMC has transferred all this risk to hyperscale computing companies, NVIDIA, and Apple. They can do this because the risk is "forgone revenue." Yes. It's money that won't be earned in the future. Worse, it's money that won't be earned four or five years from now. Now how does every company say in their financial reports? "We could have done more, but we didn't have enough supply." If you think it's bad now, why is it bad now? ChatGPT has come out. Every hyperscale computing company has started to invest like crazy. What did TSMC do? They actually reduced capital expenditures year-on-year. For two consecutive years, they did not respond to the market during the ChatGPT moment. Now they increased it to 41 last year, and it might go to 60 this year. But even if it increases to 60, the growth percentage is lower than last year. I think we will face a huge chip shortage in 2029, especially since the computational density for AI is going to be much greater. If an agent is doing things, the amount of computation it performs in a limited time is far beyond what I can compare to my Google search. And we also have a CPU shortage. Intel has shut down some CPU factories, right? So in the entire semiconductor field, I think this is a big problem. For a long time, we've been thinking about how to find alternatives to TSMC geopolitically The fact is, this is somewhat like a bundling issue. It's hard to get companies to buy insurance, especially when the insurance is, first, that everyone wants someone else to do it; who will make the sacrifice? Second, it may not happen; war may not occur. Third, as long as it doesn't happen, going elsewhere is super suboptimal because TSMC is better. It's not just that their fabs are better; their customer service is better, they have all the IP modules you need, they have experience, and you have existing relationships. And they will punish you because they have control; they have such high demand now that they can choose who gets capacity. People are afraid; they don't want to go elsewhere. So how do we solve this problem? I think I even wrote about this on the homepage of Stratechery this week, basically the same thing I wrote in my update, which is that someone needs to intervene. Hyperscale computing companies need to realize that I think a massive tightening is coming. They need to rely on Intel to accelerate, on Samsung to accelerate, to form a credible alternative. Theoretically, you could pay for geopolitical reasons or shortages. No, geopolitical reasons are free. I think there are huge economic reasons to do this, namely that if you don't do it now, you will forfeit all the revenue you will lose by 2029. And then we would be happy to get geopolitical insurance for free.

> John:
> 
> But if TSMC is the best, rather than supporting a seemingly difficult Intel, isn't the answer to prepay for building additional fabs? How do we view the ongoing operational costs in the tech sector versus putting some money upfront to permanently solve the problem?
> 
> Ben:
> 
> The market structure is an issue. You're facing a monopolist; it's not that they are a bully; they are very nice, and you could say they haven't raised prices significantly as they should have. But the reality is that this market structure issue will affect hyperscale computing companies. I think they better solve this structural issue; otherwise, the cost of ensuring or overcoming this problem will only get higher.
> 
> John:
> 
> This seems to be the issue you've felt most strongly about in the past year or two.
> 
> Ben:
> 
> I feel strongly about the Apple Vision Pro as well.
> 
> John:
> 
> Well, what are your thoughts on the Apple Vision Pro?
> 
> Ben:
> 
> They finally showcased NBA games, and then they kept switching angles. They applied 2D television production techniques to immersive technology. Just let me watch from the sidelines.
> 
> John:
> 
> I know the TSMC bottleneck seems more important.
> 
> Ben:
> 
> Maybe.
> 
> John:
> 
> I have some quick questions, and they may not be quick; they are just some unrelated questions. Let me tie them together.
> 
> Ben:
> 
> Great.
> 
> John:
> 
> Now that AI exists, how should schools handle assignments?
> 
> Ben:
> 
> I think they should integrate it; they might need to conduct offline exams. I mean, trying to suppress it is foolish. I am very much against those AI detectors; they don't work, and I might be particularly sensitive because obviously, a lot of my work is in those models No one, my thing is I'm not an MD user, but I'm the world's largest semicolon user. Oh, unfortunately, the model doesn't seem to have really absorbed the semicolon. I'm not that influential yet. But yes, you want kids to use it because whoever can use AI most effectively in future jobs will have a significant advantage. So there might be a kind of return, where classroom time becomes more important. This is my general view on content. I believe in a world where not all content, but some content is more valuable than ever because AI is a perfect personalized experience. What you read isn't necessarily what I read. So what we both read is actually appealing, and I'm very interested in how to leverage this for long-term benefits. What can you get from school that you can't get elsewhere? I can read notes, read XYZ, but participating in discussions in class, truly interacting, being pushed to think about these things. This sounds like a beautiful theoretical description of what school could be, which might be far from reality. But identifying things that share common experiences will become increasingly valuable.
> 
> Ben:
> 
> Shared content, shared classroom time, live events, shared experiences, because anything personalized will be completely consumed.
> 
> John:
> 
> Will sports teams be more valuable than in an AI-rich era?
> 
> Ben:
> 
> Everything live becomes more valuable. This is something I think about a lot in my own business. Every day, tens of thousands of people read the same thing. This is actually very powerful. Doing live events brings people together. I've thought a lot about community. I don't think anyone has really solved the community problem around content, like message boards or comments, where there are always very bad dynamics and a few people dominate everything. That's true. But if we're in a group chat, you share an interesting article, and then you all discuss it, that's great. So there are many things around this that I find very interesting and think a lot about.
> 
> John:
> 
> What do you think about the current state of the cryptocurrency field?
> 
> Ben:
> 
> What is cryptocurrency? No, I've always been a defender of cryptocurrency. Just because digital scarcity is fundamentally interesting, in a world of infinite content, it might be even more interesting. We used to think we had infinite content, now we have steroid-infused infinite content. It's not just 6 billion humans typing, but agents are constantly generating things. In that world, I think cryptocurrency as a marker of authenticity will become increasingly important. Ultimately, I want the original, not a copy. I'm optimistic about humanity's ability to create value in places where it seems impossible to exist. I myself am a professional podcaster and content creator and get paid well for it. Imagine explaining this to a farmer worried about automation.
> 
> John:
> 
> Speaking of which, you mentioned that most of the consumption of Stratechery now is in audio form rather than written form.
> 
> Ben:
> 
> As far as I know, I don't do statistics, but over half of the subscribers are subscribed to the audio form. This is interesting. I added podcast functionality, which was actually when I started developing software myself. I urged everyone to support paid podcasts, there are dedicated paid podcasts, there are writings, but no one is doing it So, of course, I had to hire engineers to build it myself, and it turns out that was the right approach. Now everyone is doing it. Anyway, that's probably my fate. But people really like it. Interestingly, I'm not sure if this is beneficial for my business.
> 
> John:
> 
> Why?
> 
> Ben:
> 
> The good news is, I think it promotes retention because people accumulate emails, feel like it's a lot of work, and they say, "I haven't read this in a long time, unsubscribe." And they only spend seven or eight minutes on it. The problem is they don't share. Audio content doesn't get shared. That's right. I listen to it on my way home from work, it's great, and then I never think about it again.
> 
> John:
> 
> That's right.
> 
> Ben:
> 
> But that's great for me because I can write the same content the next day, and you might think, "Wow, insightful commentary," without knowing I just said it yesterday. Yes.
> 
> John:
> 
> If we think about which industries will become important in the future, like AI development, energy will be a big issue, as well as the ability to power new data centers coming online. This could be a more serious constraint than chips.
> 
> Ben:
> 
> I mean, I think any country that can truly build things is in a favorable position. But on the flip side, if I can look at future challenges from an optimistic perspective, somewhat contrary to doomsayers, it’s actually about creating new forms of value, new sources of value, in uniquely human ways. This is by definition an innovation story. It’s about freeing resources from what machines can do to more efficient tasks. It’s a consumer market that can drive this innovation, making it possible to write newsletters or podcasts and get paid for it. And the actual value creation, what makes us human and generates value, I think people in the AI space are skeptical about this, even though 90% of us are in agriculture, and only 1% are in... for some reason this won't repeat. If you want to be optimistic, that's something America has always excelled at.
> 
> John:
> 
> What feedback do you have for us about Stripe?
> 
> Ben:
> 
> Where to start? I obviously find it hard to write about Stripe because I'm not objective; I'm too early in. I think you launched the billing API in 2011, which directly sparked my belief that this business model was viable. So on that point, big thumbs up. You didn't warn me about this; I should have thought of it. Actually, you have a big issue that I just encountered recently.
> 
> John:
> 
> Oh, really?
> 
> Ben:
> 
> Your A plan implementation, if someone comes in, if I try to add someone to an A plan, assuming I have a team plan, this is how you use the A plan, big companies want to add individuals, and if that addition fails, the whole plan gets canceled. So we have to build a lot of logic to handle this issue independently.
> 
> John:
> 
> Okay, that's a very specific problem. We are facing bugs with A plan subscription interactions. Okay, that's good feedback. There’s definitely more; I need to think about it. But I do think we haven't talked about stablecoins in this area **I have always been skeptical about microtransactions because it goes back to the issues of investigative reporting. You can't build something sustainable just by relying on backend monetization. The only way is to have a very large market, and that's YouTube.** YouTube is a bunch of speculative video creators hoping to get enough views for advertisers to pay them. YouTube is so large, and ad monetization is so effective, that it works. For written content or podcast content, there is no such market. People will say, oh, I’ll pay for an article. No, when you pay me, you are paying for my ongoing production. I promise you I will write something every day, and you pay for that promise. You are not paying for the actual content. The content is a byproduct of that promise. The problem with AI and microtransactions is that all these labs are paying people around the world to generate data. If you are a radiologist, you earn $300 an hour. I saw an article talking about all these floating ratios, and everyone is repeating work. This is because everyone feels there is differentiation. We clearly need some kind of market mechanism for data generation that will replace what we get from news enterprises, which are more doomed to fail than ever. So, how do we generate a market where you pay directly for content, and then AI can access it, and you can build a large market like YouTube where people will speculate, believing they will be rewarded because the market is large enough? This needs to happen. Like many things, there is a huge gap; how do we get from here to there? But we will see. I know Cloudflare is pushing this forward. We will see.

> John:
> 
> Last question. How do you evaluate the execution of large tech companies, like the big five?
> 
> Ben:
> 
> Okay. Apple has traditionally been very strong. Their manufacturing is still amazing. The newly released iPhone Air, I turned off the alarm and made sure I turned off the snooze. The best smartphone ever.
> 
> John:
> 
> Really? I've never seen it.
> 
> Ben:
> 
> Oh, it's thin.
> 
> John:
> 
> Is the battery life good?
> 
> Ben:
> 
> Good enough.
> 
> John:
> 
> That sounds like it's not good.
> 
> Ben:
> 
> No, it's fine. I forgot to bring my external battery today, and it’s still holding up. In fact, I really like it. I was shocked to hear they might not be producing it regularly. Obviously, Apple's software has become quite rough. Their relationship with developers, the interesting thing about Apple is that the reality is, when it comes to platforms, you have to pay the price to be a platform, which is to make great products. So Apple understands platforms, they make great products, but they are terrible platform managers. Microsoft is a good platform manager, but they can't make good products, so they never got the license to own a large platform. It's a bit tragic. But Apple is an old company driven by managers rather than founders. Maybe Siri becoming so bad is obviously very bad. But at the end of the day, we still need devices. They are still better than anyone else So they might be okay. Google, I've always found it the hardest to understand, partly because I think Google has done many suboptimal things. Almost everything I feel they do is suboptimal. But I think the lack of optimization like Apple actually makes them the most resilient among all tech companies. Because they have never done precisely what they should be doing, they have all these extra embellishments, doing various scientific projects. But because their core business model is so good, generating so much cash, they can be very flexible. I’ve started to appreciate this. All the things that frustrated me when analyzing them actually have this hidden benefit: resilience, strength, and adaptability. They are like amoebas; if they come your way, you could be in big trouble. It may take them a long time to get there, but once they do, you're done for. Microsoft, I’ve always written that Microsoft got it right. Everyone, especially during the SAS era, all the companies said, "Microsoft is terrible, we want to make the best-in-class products." Guess what? Silicon Valley startups, they want to buy all the best-in-class products and have the ability to integrate them. Joe running a tire shop doesn’t care; he just wants things to work and work together. If everything is mediocre but works together, that’s better than best-in-class. Time and again, these companies grow and then hit the wall that is Microsoft. Will this continue in the AI era? This may relate to the previous SAS issues. Their distribution and power there are still immense. Meta might be the best executor I’ve experienced. You usually see interactions with PR or executives, and they operate so tightly, which always impresses me. I think, again, their advertising model is underestimated. Their trick is to keep users engaged, which is the key to the whole operation. The time spent on ChatGPT is time not spent on Instagram. I think this is an underestimated area. I think they are somewhat gambling, well, everything is fine today, but in the long run, this is an infrastructure game, and we have the cash flow to fund it, while OpenAI does not. I think OpenAI might pose a greater threat to Facebook than to Google, which is worth considering. But Facebook is clearly spending money to respond. Now, Amazon. **There is a lot of wafer capacity and power spent on Tranium; one might wonder if it would be better spent on other chips. But we shall see.**
> 
> John:
> 
> Aren't people satisfied with the Tranium chips?
> 
> Ben:
> 
> **The extent to which Amazon optimizes cloud computing, I think, is underestimated.** When you operate in a commodity market, there are two ways to succeed, right? You can have differentiated products and charge high margins; or you can have a lower cost structure in the commodity market, where the price floor is the market price, but your cost structure is lower than your competitors. That’s where you make your profits. This is how Amazon dominates the cloud. Their cloud is optimized far more than anyone else's. The whole Nitro architecture, the way they build everything, making many chips themselves, shifting to Graviton I think regarding Graviton, their ARM CPU, who is Graviton's number one customer? Amazon itself. So they can move all the loads over, optimize it, build all the software libraries, and then offer it to others on a cost basis. This is the script they are trying to play with Tranium; in the long run, Tranium's number one customer is Amazon, and then they develop all the capabilities for others to attract them at a lower price, giving them a structural cost advantage. The question is, this approach only works when performance levels off. Amazon executed this model from 2005 to 2025. Of course, processors became faster during that time, but not like the huge leaps every generation in the 80s and 90s. In a relatively new market, when there are huge leaps between each generation, and they have NVIDIA servers, does this approach still work? Can they have as many as NVIDIA? Because they execute this strategy.
> 
> John:
> 
> Ben, thank you.
> 
> Ben:
> 
> Thank you

### Related Stocks

- [TSM.US - Taiwan Semiconductor](https://longbridge.com/en/quote/TSM.US.md)
- [META.US - Meta Platforms](https://longbridge.com/en/quote/META.US.md)
- [GOOGL.US - Alphabet](https://longbridge.com/en/quote/GOOGL.US.md)
- [GOOG.US - Alphabet - C](https://longbridge.com/en/quote/GOOG.US.md)
- [AMZN.US - Amazon](https://longbridge.com/en/quote/AMZN.US.md)
- [AMZW.US - Roundhill AMZN WeeklyPay ETF](https://longbridge.com/en/quote/AMZW.US.md)
- [FBL.US - GraniteShares 2x Long META Daily ETF](https://longbridge.com/en/quote/FBL.US.md)
- [AMZZ.US - GraniteShares 2x Long AMZN Daily ETF](https://longbridge.com/en/quote/AMZZ.US.md)
- [SOXX.US - iShares Semiconductor ETF](https://longbridge.com/en/quote/SOXX.US.md)
- [METD.US - Direxion Daily META Bear 1X Shares](https://longbridge.com/en/quote/METD.US.md)

## Related News & Research

| Title | Description | URL |
|-------|-------------|-----|
| AI 债务泡沫初现端倪？分析师警告：Alphabet 发行百年期债券或为市场 “见顶” 信号 | Alphabet 首次发行 10 亿英镑百年期债券，策略师警示，在信贷利差历史低位、AI 需求不确定及技术变革加速背景下，此举是市场过度乐观及 AI 泡沫积聚的典型信号。本次发行是其 200 亿美元多币种融资计划一部分。 | [Link](https://longbridge.com/en/news/275752138.md) |
| 春节周重磅前瞻：美联储最爱通胀指标，DeepSeek V4 或发布 | “史上最长” 春节假期开启，A 股及多国市场休市。宏观方面，美联储会议纪要、12 月 PCE 及四季度 GDP 修正值将指引降息路径，日 本举行首相指名选举。AI 领域，首届印度 AI 峰会召开，黄仁勋等出席；国内 DeepSeek V4  | [Link](https://longbridge.com/en/news/275992825.md) |
| 姚顺宇谷歌首秀，Gemini 新模型刷爆 SOTA：人类仅剩 7 人捍卫碳基编程 | 谷歌推出 Gemini 3 Deep Think 模型，取得 3455 Elo 分数，位列全球第 8，超越以往最高分 2727。该模型在 ARC-AGI-2 基准测试中得分 84.6%，刷新 SOTA，远超 Claude Opus 4.6  | [Link](https://longbridge.com/en/news/275891427.md) |
| 科技 IPO 预期升温，但华尔街的主战场已转向债市？ | 尽管 SpaceX、OpenAI 等明星公司 IPO 前景备受关注，美国科技资本市场的重心已实质性转向债务融资。为支撑 AI 基建，全球科技债发行规模预计今年逼近 1 万亿美元，Alphabet、Oracle 等巨头密集发债，Meta、亚马 | [Link](https://longbridge.com/en/news/275770702.md) |
| 台湾表示其公司应自行决定对美国的投资承诺 | 台湾承诺在美国投资 2500 亿美元，具体支出由各公司决定。赖清德总统强调，尽管台积电在亚利桑那州投资 1650 亿美元，但台湾最大的生产能力将仍然留在本土。最近的美国贸易协议包括对来自台湾的进口征收 15% 的关税，并需要在台湾获得立法批 | [Link](https://longbridge.com/en/news/275867858.md) |

---

> **Disclaimer**: This article is for reference only and does not constitute any investment advice.