Ben Thompson from Stratechery on AI ads, the end of SaaS, and the future of media

Ben Thompson, the internet’s premier tech analyst, joins John for a wide-ranging conversation on the mechanics of the internet economy. They discuss the origins of Stratechery and the "1,000 true fans" model, why Taiwan is the most convenient place to live (and the best Uber Eats market), and why the public markets are wrong to think SaaS is "canceled." Ben also explains why the US failure to control the TikTok algorithm is a disaster, why he’s a "crypto defender" in an age of infinite AI content, and gives John some very direct feedback on Stripe’s ACH implementation.

Timestamps

00:00:20 Visiting Taiwan
00:04:59 Aggregation and AI
00:23:53 TikTok/Bytedance
00:29:58 Aggregation and AI redux
00:35:31 Agentic commerce
00:45:08 Is SaaS canceled?
00:52:21 Stratechery
01:03:36 How Ben uses AI
01:06:06 The TSMC break
01:13:53 Rapid fire
01:20:53 Feedback on Stripe

Transcripts

John Collison (00:00:01):

Ben Thompson is the founder and author of Stratechery, the newsletter that everyone in tech reads to make sense of what’s happening. He was also early to the premium newsletter model that’s become very popular in media nowadays. For many years, he ran Stratechery as a solo founder in Taiwan. Cheers. Good to see you.

Ben Thompson (00:00:17):

Cheers.

John Collison (00:00:20):

It feels like people in San Francisco have not properly discovered Taiwan as a tourist destination. Do you agree with that characterization? And what’s your recommendation?

Ben Thompson (00:00:27):

People always ask me about Asia, and the way I always characterize Taiwan is there’s lots of great places to visit in Asia. And I would also put Japan on the list. But I like to think I went to Japan before it was cool.

John Collison (00:00:39):

Yeah. Nothing against Japan.

Ben Thompson (00:00:41):

Well, the whole thing with Japan is going to Japan pre-smartphone was a completely different experience than going there post-smartphone. Like you think, “Oh, the subway system’s amazing, the trains…” Try navigating that with no smartphone and nothing’s in English. Japan used to be very low on English. It’s still lower than places like Taiwan.

John Collison (00:00:59):

It’s surprisingly low.

Ben Thompson (00:01:00):

Yeah. And Japan has the… The way to visit Japan is, you just walk. Don’t go to set destinations. Whereas the way I would talk about this is places to visit, but the best place to live is undoubtedly Taiwan. The one word everyone says for Taiwan sounds not that impressive, but the word is convenient. It is the most convenient place to live. So part of that is actually—

John Collison (00:01:24):

7-Eleven has really good food.

Ben Thompson (00:01:26):

Well, it’s actually downstream from the Japanese because Taiwan was a Japanese colony for the first 50 years of the 20th century. And it’s laid out a lot like… Why is it great to walk around in Tokyo? Because Tokyo is all mixed use. That’s how Taipei is as well. You have these big blocks where the exterior will be commercial and the interior of these big blocks is all residential and the first floor is all like small shops or restaurants, things like that. So wherever you live, you basically have access to everything all around you. But I think the downside as a tourist is it’s kind of an ugly city. Taiwan’s kind of notorious for just these dumpy, dilapidated buildings. Then you go inside and they’re palatial on the inside. Taipei is very, very rich. It’s in the top 10, I think, as the number of billionaires in the world or something like that. All downstream from building out China. It’s a very beautiful country. From Taipei, 30 minutes to the ocean, 30 minutes to the mountains. East Coast is amazing.

John Collison (00:02:25):

But if people listening to this are visiting, I feel like one thing they should do is… It’s a mistake to try and use Yelp or anything like that too much because you should maybe just try and go to a night market and follow your belly and see what looks good. There’s a lot of excellent street food. And so that’d be one thing is don’t try to overplan.

Ben Thompson (00:02:42):

Well, here’s the problem though, where tech has made it worse, I would argue as a tourist. Taiwan is arguably the greatest Uber Eats market ever because there’s just amazing options. It’s all delivered by scooter, so it’s always like 10 minutes to get dinner. I think you were going to ask me about difficulties moving to the States. Not having access to that is definitely one of them. But the problem is that it’s such a huge market now that I think there are fewer and fewer restaurants, in that a lot of these places actually just straight up close their storefronts are just ghost kitchens basically. And all they do is just make Uber Eats orders all day.

John Collison (00:03:19):

I see. So I mean famously, yeah, the restaurant economy and places like Taipei would have been really good, but it’s gotten worse because people are eating in more with Uber Eats and stuff like that.

Ben Thompson (00:03:28):

I think so. As far as walking around and just like stuff on… No, there’s still plenty of places. It’s still great. But there’s a number of restaurants that I used to always take people to, like holes in the wall that I knew had super good beef noodles or something. And I remember a couple times like, “Oh, you can’t actually go eat there anymore, but they’re still on Uber Eats.”

John Collison (00:03:45):

That’s a bummer. It’s like a separate problem. The San Francisco problem at restaurants is that nobody drinks anymore. And so the restaurants can’t… They’ve lost a major revenue source.

Ben Thompson (00:03:52):

It’s so bad you had to get a pub in your own office.

John Collison (00:03:56):

Exactly. We’re trying just firsthand to fix it.

Ben Thompson (00:03:59):

Be the change you want to see in the world.

John Collison (00:04:00):

Should people visit places beyond Taipei?

Ben Thompson (00:04:03):

Oh, for sure. Yeah. Taipei is great. It’s great to walk around. Taipei 101, which is obviously very much in the news these days with the scaling.

John Collison (00:04:12):

But you can go up the elevator on the inside.

Ben Thompson (00:04:13):

You go up there because there’s a massive ball at the top.

John Collison (00:04:17):

The mass damper, yeah.

Ben Thompson (00:04:18):

Yes. Which is amazing. If you’re into engineering, that’s actually a very underrated thing. National Palace Museum is amazing. But the East Coast in particular is incredible. There is a train, but driving, you’re driving on the coast—

John Collison (00:04:33):

It’s like a lost coast of Hawaii, kind of.

Ben Thompson (00:04:34):

Exactly. There’s an incredible gorge called Taroko Gorge. That was really messed up by an earthquake a couple years ago. So I don’t know if it’s even reopened yet. But I used to take people to that all the time because it’s world class.

John Collison (00:04:48):

Yeah. It is impressive that they said, “We’re going to build the tallest skyscraper in the world in a very frequent earthquake region.”

Ben Thompson (00:04:55):

Yeah, it’s a beautiful skyscraper. It worked out well for Netflix.

John Collison (00:04:58):

Yes. So you’ve talked a lot over the years about aggregation theory and really popularized this idea where pre-internet, often power would live with the supply, whereas on the internet, because of the different marginal cost dynamics and things like that, power will rest with the demand aggregators. And so Booking.com is a much bigger company than any hotel chain, something like that.

Ben Thompson (00:05:25):

And Booking.com is a particularly interesting one because they aggregate all the hotels, but they are also aggregated by Google. So they’re like Google’s biggest customer.

John Collison (00:05:34):

Yes.

Ben Thompson (00:05:35):

Even as they’re also on the other side.

John Collison (00:05:36):

I feel like Booking.com is a very underappreciated success story in tech. They’re a European company, much quieter in a lot of ways. But if you invested a dollar in Booking.com and a dollar in Google 20 years ago, you made much more money as a Booking.com shareholder and I think people don’t appreciate that fact. It’s a very well-run business. But where I was going with this is, how does aggregation theory apply to AI? How does one need to update the framework?

Ben Thompson (00:06:04):

TBD, to a certain extent. I mean, this is part of the huge, probably one of the most angsty debates that I have internally generally, which is OpenAI’s sort of welfare going forward. I put forward a few years ago that actually OpenAI could stop making models and be one of the most valuable companies in the world just because of ChatGPT. That’s their most valuable asset. And part of the problem that they have is that was definitely the case in 2023 and 2024, but they never… You have to actually build the business model around that. And I’ve, I think fairly famously, at least based on all the tweets that I got when they announced that they were going to launch ads, I’ve been losing my mind about this fact for a long time. And I think this is interesting. I’d actually be curious to hear your view of this, which is, there’s this mindset in the Valley of this skepticism of advertising and people have sort of internalized that it’s bad and evil. Do you sense that? Do you feel that?

John Collison (00:07:09):

I agree there’s kind of a knee-jerk skepticism of ads. Look, I’m a YouTube Premium subscriber. When I see someone watch a video without Premium—

Ben Thompson (00:07:19):

It’s horrifying.

John Collison (00:07:20):

I gasp. It’s like, “What are you doing with your life?” And so I get the knee-jerk reaction and that Stripe at some level is kind of the anti-advertising company and we’re the opposite form of monetization. But I don’t know, I have no particular issue. I think it’s a very efficient form of monetization. It makes a lot of sense for certain products. And so I think it’s just different strokes.

Ben Thompson (00:07:44):

I think you’re with Stripe, you’re on the skepticism side. I think ads are amazing. And I’m talking about my book a little bit. Stratechery has gotten tremendous traction just by not hating ads, even though I’m not an ad model myself.

John Collison (00:07:59):

Exactly, but you’re a paid model.

Ben Thompson (00:08:00):

Well, it’s funny. I actually think I got a lot of traction over the years by talking about ads when no one else was, despite the fact it’s the most important business model in tech. And I look back and all my early writing about ads was terrible. I had no idea what I…But just by virtue of talking about it, it was helpful. The reality of advertising is, number one, people… If you are making a product in the world, like Stratechery is very fortunate. It is definitely a new model or a new internet native model in that I have subscribers in like 200 countries. Literally, the whole world is my market and Stripe obviously helps make that possible.

John Collison (00:08:40):

A few in the Vatican, they’re following along with that.

Ben Thompson (00:08:42):

I could go check. I mean, I would bet the odds very high that I do, at least one subscriber in the Vatican. But where I benefited from is I was a massive beneficiary of social media, particularly Twitter. And that back in their early days, good days of Twitter, if you want to say whatever it was, there was currency in sharing smart links.

And so if I was a regular provider of links that people felt made them smart, so they would share them and talk about them and be sort of back and forth. That solved my customer acquisition issue. The reality though is most, because the other thing about content, actually this is the point I’m interested to come back to, is it’s something to talk about. It’s a commonality sort of between us that we can both read the same thing, we can both have opinions on it, react to it. The product that I buy on Instagram is not what I’m, I’m not going to post about it or talk about it, but it can be tremendously beneficial. And they in theory have these small businesses or whatever they might be, or Chinese suppliers or whatever, they have the same opportunity, which is to sell to everyone in the world.

They just need a way to tell people about it. And as someone who buys way too many products off of Instagram. And by the way, one of the great things about moving back to the US is that the Instagram ads are unbelievable. I thought we were pretty good in Taiwan. They’re so much better in the US. They’re like, “Oh, this is the best part of living in the richest country in the world.” It’s amazing. I’m like, “What’s this native content? Give me more ads.” Which by the way, Facebook is very happy to do over the last six months. Lots of ads these days, but I get stuff that I never would have thought of, I didn’t even know about, and it’s great. It’s amazing. And it’s a real benefit to me as a consumer who for sure subscribes to YouTube Premium and looks down on people who don’t, but finding things that I didn’t know about that make my life better.

So as a user, I’m benefiting. As a rich user, I’m benefiting. As the world is six billion people, most of whom do not have the disposable income that I have, much less you have, much less anyone else in San Francisco has, they get the same experience I do.

And something for AI, when you think about it, particularly when it’s so costly to provide, and the free product is so much worse than the paid product, of course it’s a win for them to be able to get access. So to have a mission, a belief that AI makes the world better and to not embrace ads.

John Collison (00:11:14):

I agree with ads being an efficient form of monetization. What do you think is the right way for consumer AI apps to do ads? Like ChatGPT just announced that they’re doing ads.

Ben Thompson (00:11:29):

They’re terrible.

John Collison (00:11:30):

Well, no, they’re doing them as a very separate experience to the answer.

Ben Thompson (00:11:30):

No. This is why it’s so bad. This is why I’m so frustrated with them. So what they’re doing is the bare minimum easiest solution.

John Collison (00:11:36):

It’s like banner ads basically.

Ben Thompson (00:11:37):

It’s banner ads, but it’s based on the context of the conversation. And the problem is that they release their ad principles, right? Which is, our ads do not influence your answer. If you’re using the easiest possible way to target ads, which is based on the context of the conversation, we’re going to show you a roughly relevant ad. Number one, your market’s way smaller because you have to hope someone starts a conversation that matches the inventory you have. Number two, you’re getting into a, my T-shirt answers questions that my T-shirt is raising sort of situation where if the ad is clearly connected to the answer, you’re going to raise suspicion in the user’s minds about what the connection is.

So I would prefer if the ads had nothing to do with the answer. The way you get there is you build a Meta-style understanding of the user and show them stuff that’s relevant to them, like in Instagram. The best Instagram ads don’t have anything to do with the stuff I’m surfing. It’s from Meta’s understanding of me broadly.

John Collison (00:12:34):

So are you saying that AI ads should be more like Facebook ads than Google ads? And right now the focus is on doing targeted ads that are related to the prompt, whereas instead it should all be profiling the user and who this person is and what their interests are?

Ben Thompson (00:12:48):

Yes. I think that would be better. It would present less conflict of interest, less uncertainty amongst the user. And it’s a model that I think… I’m not the world’s biggest fan of search ads precisely for the reason why they work so well. I think there’s a lot of search ads—

John Collison (00:13:02):

Because of the confusion between organic and—

Ben Thompson (00:13:04):

No, because they’re cannibalization. Why is it that I have to buy my own name in search ads, right? Because someone else will go in there and you’re harvesting a click on the ad that would have been there organically, which is fine. It works. The search is providing a lot of value, but the challenge obviously is they only have one space for inventory, which is in ChatGPT.

John Collison (00:13:25):

Well, sorry, isn’t the defense of Google ads that everyone complains about the branded search and yeah, you’re paying for cannibalization, but Google pays so much attention to search quality that the sponsored listings themselves have a relevance ranking of them. And so it’s really just like the Yellow Pages where you need to pay to be listed in the Yellow Pages.

Ben Thompson (00:13:50):

Oh, it’s fine. Again, I’m the ad lover here. I just think that Meta ads are more broadly valuable because they’re showing me stuff I didn’t know that I wanted.

John Collison (00:14:00):

But if the AI apps are to generate a profile of you, does that profile include the content of all your conversations?

Ben Thompson (00:14:08):

Well, so this is the thing. So Demis is out there saying, “Wow, I can’t believe they’re adding ads. We’re not going to do that.” Which is hilarious because the entire Gemini DeepMind apparatus, what is it funded by?

John Collison (00:14:21):

Sure. Yeah. It’s a Google ad machine.

Ben Thompson (00:14:22):

It’s funded by ads and that actually is probably the ideal model. So it’s actually very funny. I was in New York City last year. I was meeting with someone in the office, and in the shared office or across the hallway was a hedge fund or someone and they came over and he’s like, “Oh, longtime reader, you’re responsible for our worst decision ever.” And I’m like, “What?” He’s like, “Putting money in Twitter.” I’m like, “I’ve never said to put money in Twitter.” That’s always been a terrible company. I’ve stopped covering them because it was such a bad business. And I’m like… Oh no, I remember what it was. It was when they bought MoPub. And my theory, my problem with Twitter advertising has always been that, especially it was very textual. I think text doesn’t work. All this applies to the Chat clients.

Text isn’t the best interface for ads. Obviously visuals are generally better and there’s also a posture. If I’m on Twitter, I’m like, “I’m ready to do battle. I’m locked in, or I’m searching for information.” If I’m on Instagram, the whole point of seeing an ad is I don’t really care what I’m seeing right now. I’m wasting time. You’re actually in a much better posture, I think, to absorb. Just like TV, like you’re sort of absorbing, can absorb the ad and Twitter is bad for that. But Twitter, because it’s an interest-based network, at least in theory, it should be able to understand a lot about you above and beyond theoretically having pixels and SDK all over the web. And so my theory with the MoPub acquisition, I thought was a great acquisition because like, oh, they can harness signal from Twitter and manifest it in other apps through this MoPub network.

Now, Twitter was incompetent, so they did nothing with MoPub, gave it to AppLovin who’s now riddenMoPub to the top of the world, but that was sort of my thesis and I think that could apply to AI as well. I think the ideal outcome for Google is they never put ads in Gemini, but they understand so much about you because of what you do in Gemini that they can then manifest that through ads on YouTube, through ads on Google, through ads on their other properties. And the challenge for OpenAI is they only have one place to put inventory, which is in ChatGPT.

John Collison (00:16:29):

Okay. So you’re saying that Google could use Gemini to just improve the targeting of the ads across the Google properties, and then maybe if you want to have ads in Gemini—

Ben Thompson (00:16:40):

I don’t think you need to ever put ads in Gemini.

John Collison (00:16:41):

But just if you did, you would also have the profile that Google has of you from across the web and you can choose—

Ben Thompson (00:16:46):

Yeah. And you don’t need to have ads that are like making the user feel weird because why are you showing me ads about what I’m asking about?

John Collison (00:16:53):

Okay. But in the scenario you’re just describing for Google, wouldn’t that have the same, just like the Meta’s listen to your microphone conspiracy theories where the targeting is too good, people get concerned. Like wouldn’t you get similar issues?

Ben Thompson (00:17:05):

I think that’s a made-up concern.

John Collison (00:17:06):

No, it is sure, but people have it. And wouldn’t you have a similar issue where if you’re using the Gemini data to make better ads, wouldn’t ultimately the targeting be too good and people find it weird?

Ben Thompson (00:17:21):

I mean, I think that that’s a bridge that every tech company would be happy to cross if they came to it.

John Collison (00:17:25):

I see. It’s a thing that people say when you have very good targeting.

Ben Thompson (00:17:27):

I think there’s a lot of …There’s a real stated versus revealed preference about a lot of this stuff. The reality is, you could pay for Facebook or you could show…People would rather see the ads. I think most people don’t care. A lot of tech, and this sort of ties into the skepticism of ads, it’s sort of an elite town, there’s elite regulators, everyone’s thinking about these very theoretical things.

John Collison (00:17:52):

Isn’t that bit of the challenge of banner blindness where Instagram advertising works so well because it’s a picture feed and it’s showing you pictures. And then some of the pictures are like commercials. Whereas with an AI app, you’re looking for an answer and you don’t want to look at the banner.

Ben Thompson (00:18:08):

No, it’s a huge concern. And this is one of the great ironies of Meta/Facebook is the extent to which… Of course, Mark and everyone hates Apple for lots of, I think, very justifiable reasons but Apple saved Facebook from itself. Like back in the day, remember Facebook platform and there’s like Facebook payments and all this sort of thing and Mark has always wanted to build a platform. And if you’re just an app on a phone, you can’t build a platform. And the problem is that I think being an advertising-based model is generally incompatible with being a platform. The whole point of a platform is you’re letting something else shine, something else to bring to the surface. The support structure for something to take over. So an operating system is not about the… Ideally, it’s the application on top of it that you’re using.

When Facebook was forced to not be a platform, but just be an app, suddenly they could be fully leaning into being an advertising thing. And think about a Facebook ad. Even back in the day when it was a feed ad or a story ad, literally your entire device is all an ad. And somehow it’s not a banner, that’s a little thing on the edge. They literally have achieved permission from users to take over your entire device to show you a full screen ad every five seconds. It’s amazing. And they were forced into it by Apple.

John Collison (00:19:35):

Okay this reminds me, and I want to come back to the AI dynamics, but this reminds me of a view I’ve had that I’m curious for your thoughts on, which is often when tech companies become really big, they become really big just because the core idea works better than even the founders could have realized. And so Meta’s a really big company because they have a feed and the feed got really big and they were very smart along the way where they bought Instagram and they’re like incredibly targeted—

Ben Thompson (00:20:07):

No, it’s the feed.

John Collison (00:20:08):

But it turns out people spent a lot of time and many people, the P x Q of that with the feed and they monetized it very well, and that’s what got really big. And same with Nvidia, it just turns out that the GPU market got really big and they sell a lot of GPUs. And so maybe founders, because they’re often like high-powered individuals who want to have lots of new ideas, they’re often thinking about the next thing or like what the second act or the third act is and everyone wants to invent an AWS. But I’m curious what you would say to the idea that just generally it’s making the core thing really big and there’s more orders of magnitude at the top than you thought.

Ben Thompson (00:20:43):

Yes. I think that’s always the case. And I think that sometimes people end up making something that they didn’t want to make and they continually push back. I think Meta is the perfect example. My impression is Mark’s not very interested in ads. He’s had very good people along the way that have helped him build these ad products. I think Meta has suffered from that because he has not been front and center fighting for, actually ads are good. They are societal good. They are the driver of all the consumer surplus that tech throws off. The President uses the same search engine as the guy on the street or the same AI or whatever it might be. That’s because of ads and the—

John Collison (00:21:28):

Probably not. The President probably uses a Palantir search engine or something.

Ben Thompson (00:21:31):

Yeah, it’s probably worse. Google has slipped a lot to be fair. There’s so much junk online. But has Donald Trump ever searched? I don’t know. That’s a good question. But he’s not made that case. And I think Meta has suffered because of the failure to make that case. And then you get things like, “We’re going to do the Metaverse. We’re going to do X, Y, Z.” It’s always coming back to be a platform, be a platform. And Meta is an entertainment company. I wrote this years ago about the… It was simultaneously a good call and a bad call. Do you remember that Paul Krugman quote, “The internet’s not going to be very big, or have more impact than the fax machine because people don’t have anything interesting to say.” I actually defend that quote because it’s actually true. Most people don’t have that much interesting things to say. And I brought up that quote around 2015 by saying, “This is a fundamental limiter on Meta’s long-term potential.” As long as they think of themselves as a social media company, they’re going to run into a problem with their feeds becoming insufficiently interesting over time. Now—

John Collison (00:22:41):

The move from kind of peer content to—

Ben Thompson (00:22:44):

Well, so that was, if I might say so myself, a very brilliant insight. The bad insight was my prescription, which was they needed to do more with professional content makers, like more funding of the BuzzFeeds of the world and share revenue, all that.

John Collison (00:22:57):

It was actually user-generated content.

Ben Thompson (00:22:59):

The actual answer is what TikTok did, which is that TikTok’s not a social network at all. It is a harvesting and YouTube, the same sort of idea. What actually matters, and this is a key thing, people get hung up on relative numbers, what matters is absolute numbers. So it is better to have 0.1% of your content is good if your content is in the billions or trillions, as opposed to, “Oh, 10% of our content is good,” but you only have a hundred pieces of content, that’s actually worse, even if you have a better hit rate. And so spurring lots of creation, writing the algorithms to capture the good stuff, put it up there, that actually solves the Paul Krugman fax machine problem. And Facebook was blindsided by that. They were so stuck on their identity of being a social network that they let TikTok take this huge chunk and it was their blind spot.

John Collison (00:23:53):

Speaking of TikTok, I feel like you don’t write about ByteDance that much. And I’m curious just what your thoughts are on ByteDance from here and the TikTok sale and everything.

Ben Thompson (00:24:00):

I mean, what a mess. I had to make a decision a long time ago. I wrote about Chinese companies more previously and I think there’s… Number one, I have to decide what I’m going to be able to cover and what I’m not. I’m not in China. I was in Taiwan. It is a different internet. And there was too much uncertainty and unknowns just in general about a lot of Chinese companies. I would write about them occasionally in the context of US tech companies. So I think I wrote about WeChat and what it meant for the iPhone’s relative competitive position in China, how it’s different from other countries. I think that sort of held up pretty well. Wrote about TikTok in the context, I mean, more about TikTok, I think in the context of Meta. TikTok came up around the same time as Quibi, which was the example of that. Quibi was actually right that there was room for a mobile entertainment product. It was totally wrong about the content acquisition strategy. So even if the hit rate was higher, their total volume was way too small. I follow them, but not super closely. It’s just a hard market to understand and the—

John Collison (00:25:08):

But like TikTok’s very relevant to the US market.

Ben Thompson (00:25:10):

So I wrote the TikTok War, basically making the case that the problem with TikTok, and back then everyone was talking about user data. Who cares? The whole user data thing, people have this view of like the East German Stasi and like folders going through people’s data. These are like vector databases with numbers that no human can parse. It’s really quite anodyne. It’s just really the target ads. And I was very skeptical about that being a forcing function in terms of forcing divester or whatever it might be. The issue I had was the algorithm. And I noticed, I think it was when the Hong Kong protest happened and Daryl Morey, the then GM of the Houston Rockets tweeted, “Free Hong Kong” or something like that. And there’s a huge meltdown of the NBA games being canceled. And I noticed that on TikTok, and this was from, I tested it from Taiwan and via VPN from the US. If you search for every single NBA team, you got NBA clips, except for the Rockets, you got nothing.

John Collison (00:26:18):

Oh, that’s funny. It got demonetized. The Houston Rockets.

Ben Thompson (00:26:19):

There was a thumb on the scale here. And I started talking about it then, and I did support the ban of TikTok or the forced divesture from China because it seems fairly insane to have a primary information source controlled by your chief geopolitical adversary.

John Collison (00:26:39):

Yeah. Same like there’s rules over TV station ownership and it’s not wildly different.

Ben Thompson (00:26:43):

And so everything’s a trade off. Of course, I’m pretty well known for being a pretty stark defender of free speech and against censorship. And my issue wasn’t TikTok per se. The reality of China is the founder of ByteDance is long gone because he got called to the carpet for ByteDance showing a little too much of what people liked, which is mostly like hot girls dancing and insufficiently showing the things that the party wanted. The reality is China has… The price of doing business is they’re somewhere on the control structure, they could tell you what to do and this just seemed like a very foolish thing to tolerate. Unfortunately, the US political process, or fortunately, maybe the reality is the US process and system is such a mess. Can anyone really, truly impact it over time?

The way that it shows up messily is we somehow do pass the law banning TikTok and it didn’t get banned and now it is sold, but China still controls the algorithm. So I think it’s a big disaster. It’s also like, what can I say about it? I said my piece.

We ended up in the worst possible case, which is we violated property rights and we did all this stuff that’s ridiculous and we probably bartered X, Y, Z for ABC and we didn’t get the most important thing, which was control of the algorithm.

John Collison (00:28:10):

Has that not happened as part of the sale?

Ben Thompson (00:28:12):

No. ByteDance still controls the algorithm.

John Collison (00:28:14):

I didn’t know that.

Ben Thompson (00:28:15):

Yeah. Good job by us.

John Collison (00:28:17):

That does seem like it was the point of the spin out.

Ben Thompson (00:28:19):

Well, the data was always the most salient political point.

John Collison (00:28:24):

Yeah.

Ben Thompson (00:28:25):

So when I wrote about it, that was my point. It was like, “I don’t care about the data. The issue is the algorithm.” And unfortunately that did not care. Maybe I should have read about more, but anything like all the politics stuff, there was a period… I mean, thank God for AI. When I wrote Aggregation Theory a couple of weeks later, I wrote something about regulation. I’m like, “This is going to drive a bunch of regulatory issues and antitrust things and all these bits and pieces.” When that actually happened at the late end of the last decade, of course I was writing about it, I was watching congressional hearings, all this sort of thing, and that is the closest I came to quitting and burning out. I think burnout’s not a function of how much work you’re doing, it’s doing work you don’t enjoy. And at one point I’m like, “Oh, either I quit or I stopped covering congressional hearings.” So I decided to stop covering congressional hearings.

I only wrote about antitrust stuff that was super prominent and I’ve been much happier ever since. And maybe that’s part of the price of just not writing about that is maybe I should have pushed on the TikTok thing more.

John Collison (00:29:24):

That’s interesting.

Ben Thompson (00:29:24):

I said my piece.

John Collison (00:29:26):

Is Stratechery very widely read in DC?

Ben Thompson (00:29:28):

It is. Sometimes it’s gratifying. It’s great when you get called and asked for your opinion or you get certain responses or you see impact. It’s less gratifying when you get yelled at and people are mad at you. But fortunately the key thing to succeeding on the internet is something I have in spades, which is a very high level of disagreeableness. So you can yell at me all you want. I’m not going to change my mind.

John Collison (00:29:57):

Okay. But getting back to Aggregation Theory as it pertains to AI. A simplistic view you could have is that the AI apps are the new aggregators and so a huge amount of economic value will accrue to them and that’s it. You could also say that that’s too simplistic in a bunch of ways because like we were saying Booking.com, you expect it to return new hotels that you should book, but you expect a little less of a commercial incentive from the AI apps. And this is like a little more of an abstract technology where it’s actually not trivial to insert all of the commercial incentives in the right way. Anyway, you come up with various objections and so do you think—

Ben Thompson (00:30:38):

Well, I think that the ad model is probably the way to start. Which is what I just talked about before, the lean in versus lean back. Ads are very tied into human psychology and like what you’re sort of tapping into and people’s response to that and how do you make something creative? And in the short term, technology often makes old business models even more powerful before it kills them. So you have something like a newspaper, I used to be limited to my geographic area. Now I can reach the whole world. Oh, and a few years later, everyone can reach the whole world. I mean pure competition, I’m screwed. And that is certainly a concern about this model. If you get to a world of say agentic commerce and the agents are just buying the right thing. And I think this is also something that has driven a lot of tech skepticism of ads. People in tech tend to be fairly nerdy, fairly obsessed. They’re doing a ton of research to find the exact right thing.

John Collison (00:31:39):

Yes. Why didn’t you tell me what to buy when I’ve researched it for two hours?

Ben Thompson (00:31:43):

That’s right. And so, ads have no effect on me. Well, what if that sort of obsessive deep dive approach is now trivially available to everyone because AI is the one actually doing it? Now where do ads function? And I think this is definitely a bit of a “be careful what you wish for” scenario because what this entails is of course more transparency, more details, more understanding sounds good. What it actually entails is sort of perfect competition, which is a very sort of brutal game that can just wipe out entire categories. That’s basically what happened in newspapers in many respects. So that’s number one. Number two, in this sort of world, you’re by definition anchoring on whatever specifications or whatever can be measured, can be put down.

And you had the old Steve Jobs adage about feeds and speeds versus like the feel of something where the intersection of liberal arts and tech. What the fuck does that mean? And it’s just like, well, what it actually means is there’s things that can’t be measured and that don’t go on an Excel spreadsheet. And everyone you talk to acknowledges this. They say yes, there’s things that can’t be measured. And the way it actually plays out in practice is only the things that are measured.. I think a huge problem with sports analytics is a great example of this where—basketball is my favorite sport. There’s a lot that goes into basketball and winning that is somewhat hard to wrap your hands on.

John Collison (00:33:22):

It’s not like baseball which is very measurable.

Ben Thompson (00:33:23):

Baseball’s very measurable. I do think there’s aspects about clutchness and stuff that I don’t know that are properly measured, but around basketball for sure, there’s the interaction and the way teams play together and how your effort on, or your involvement on offense can affect defense or sort of back and forth. And you see it again and again. I like Daryl Morey, I think there’s a reason his teams haven’t won. They’ve over-optimized at the expense of some of these other issues. And if you can’t measure them, they tend to get devalued. And in a world of AI-mediated everything, how many things that can’t get measured fall by the wayside because we end up with very utilitarian goods that have no soul to them. Sort of a silly thing to worry about in some respects, or it sounds silly, but I’m a human and I anticipate liking and preferring the humanity of things of all sorts in the long run.

John Collison (00:34:32):

But you could say that e-commerce aggregators like Amazon and lots of others have led to fairly anonymous manufacturers of lots of everyday goods, the kind of Amazon Basics type stuff, at a much lower price point than they were previously at still perfectly good quality. Isn’t that fine?

Ben Thompson (00:35:02):

So this is where you throw my ad argument in my face. Which is like, actually it brings up the base level for everyone, like your basic consumer, the access of items they have—

John Collison (00:35:12):

There’s no soul in an Amazon Basics power adapter. And that’s fine.

Ben Thompson (00:35:14):

Everyone thinks back to like, “Oh, my washing machine was so much better in the 1960s.” And it’s like, yes, that’s true and also far fewer people had washing machines. And so I’m now making the opposite argument, sort of engaging myself.

John Collison (00:35:26):

I will leave and you can have a one-person play.

Ben Thompson (00:35:27):

I’ll just switch back and forth.

John Collison (00:35:29)

You can change sides of the booth. And you mentioned agentic commerce. We obviously are big into that and had our announcement with OpenAI back in October. Where do you think that goes? How do you see agentic commerce playing out?

Ben Thompson (00:35:43):

I mean, the contrast between your own OpenAI announcement and Google’s announcement I think is pretty interesting and speaks to what the companies are driving for. OpenAI wants to be the place where you do everything. They want to be like the aggregator. I think a critic would say people compare them to Netscape. I think the better analogy if you’re an OpenAI skeptic would be AOL, where they want to be the interface for everything that you might do and it goes through their channels. And Google, just as they were relative to AOL is like, “Actually we want to equip everyone knowing that if everyone is capable, we are the greatest beneficiaries because we still marshal the front-end demand in that regard.” Now, how does that actually manifest in terms of commerce? The funny thing about tech is I don’t think it will manifest in terms of airplane tickets, which is everyone’s example. Everyone can never think of a better example than that, but what is the AI going to buy? What is it going to get? I don’t know. I think I would like to think people will want to have agency in their buying decisions, but then again, we have assistants, whether it be like for work or whatever it might be, and they make buying decisions that we’re necessarily not involved in and that I think is a good predictor precursor of what people will ideally… Do I really need to know… Actually, I have very strong paper towel ideas. That was going to be my….But once that’s set, can that be sort of monitored and done? So I don’t know. I think this is a very unsatisfying answer, other than to say it has big implications on things like advertising and on things like is that going to be a viable business model going forward? What margins are going to be available? Is there going to be perfect competition? Things along those lines.

John Collison (00:37:34):

Okay. Let me try this on you for agentic commerce and I’m curious to have you critique it, which is sort of how I see things playing out. I think some skepticism is triggered by people pitching a very far end state with a lot of agentic autonomy. And so it’s like, “Please book me a honeymoon in Japan and all the activities.” It’s like no one actually does things that way. Whereas actually you should go from the bottom up in some very basic building blocks where step one is just replacing filling out web forms. That’s an activity that sucks. No one likes it. And so imagine you find the winter jacket you like and you copy the URL into ChatGPT and just say, “Please buy this for me.” And that’s a much better experience than going and clicking around a site you’ve never been to before. And so there’s just the agents doing the kind of tool use on your behalf and everyone can create that. Maybe it clarifies there’s multiple colors, which one do you like? But it’s just replacing filling out form fields.

Ben Thompson (00:38:33):

This is, by the way, one thing that I am very…A lot of people are skeptical of this, but I am very optimistic about, which is, I call it ‘just in time UI’.

John Collison (00:38:43):

Exactly. It’s a better UI. Right. Okay. So that’s like level one is a better UI for kind of doing it in action you want to know. Okay. Then level two is better discovery and search. It is crazy that we’ve gotten this far in e-commerce with keyword-based search. Keyword-based search works really well when you’re buying a book that you know the name of. It’s like, I want to go buy this particular title. And for a winter jacket, it’s like, “I don’t know, I want, it’s like a puffer, like what’s it called?” And so instead you want to be able to say, “I’m looking for a jacket, I’m going to this place, it’s going to be this cold, I like these kinds of things,” whatever. And so step two is just better search and the ability to search with parameters that no existing search UI lets you specify the temperature of the place you’re going to actually get a jacket of appropriate warmth, but that’s obviously with a jacket one of the core things. And so better search UI is kind of level two from our point of view.

Ben Thompson (00:39:40):

Right. Which I think is already sort of manifesting.

John Collison (00:39:42):

Exactly. We’re already seeing it and like in the early usage of the ChatGPT buying experience, I think that’s one of those super cool features. And then level three, which we haven’t really seen play out yet, is this idea again of a persistent profile of the user—

Ben Thompson (00:39:57):

That anticipates their needs.

John Collison (00:39:58):

Exactly. It’s like, I want to be able to just pin things I like as I go along. Or maybe if I can share my browser history or maybe if I can just share a Pinterest board of just like, “These are some styles I like, give me a good winter jacket for the cold based on that, here are some photos of me based on this.” And so start—

Ben Thompson (00:40:16):

Oh, I have an even better idea. Imagine if you were using ChatGPT and it’s circa October 1st and there’s an ad for a great winter jacket that is perfectly suited to me because they’ve been understanding my interests. They understand the context of where I am. I’m not searching for winter jackets because I don’t plan well. It’s going to get cold and then I’m searching winter jackets, but what if it could anticipate that and show me an ad at the right time when I need to see it?

John Collison (00:40:42):

Okay. Maybe that’s level four is like the—

Ben Thompson (00:40:44):

That’s what I’ve been wanting them to build! This is my whole bit before. This is why they’re so late. They should be shipping that this year. You’re only shipping that this year if you started your ad product two or three years ago. This is doable today. This is what Meta ads are. You need to watch more reels. I’ve bought more ski equipment this year that I don’t need. Just because it just shows up, I’m moving back to Wisconsin, so I’m buying stuff for the house and I get all those ski hangers and think those would be great. That sounds very useful. They’re still in a box. I haven’t actually put them up.

John Collison (00:41:17):

Yeah. So there’s a limit to what kind of banner ad type experiences you can do. Whereas I think the search thing is very powerful. But yeah, I’m curious what you think of step one, just the very act of checking out and then—or level one—the very act of checking out, level two better search, and then level three, defining your own embedding space of preferences.

Ben Thompson (00:41:41):

No, I completely agree with that approach. I just think you underwrite the extent to which level three has already been built. Actually, one thing that Mark Zuckerberg said on a couple earnings calls ago that I thought was very astute, is we get hung up on technological definitions like, what is an agent? And he’s like, “Actually the largest and most successful agent in the world today is Facebook advertising,” which is exactly right. Facebook advertising, people have it in their head that you go and you put in like demographics and you’re targeting and stuff.

John Collison (00:42:10):

No, it’s very autopilot.

Ben Thompson (00:42:11):

Yeah. What you do is you go in and you say, “Acquiring a customer for this is worth $10 to me. I’ll spend up to $10 and they will deliver you a customer for $10.” Their margin will actually increase because they’ll make sure they deliver it at exactly $10 and they can do it for more and they actually make more money. You get exactly what you asked for. And I think the extent to how powerful this already works, they’re just stuck on “50% of my ads work, I don’t know which ones.” No, on Facebook, they all work.

John Collison (00:42:42):

I feel like a bunch of new, very big successful companies will be created in AI-powered e-commerce. It just feels like a different enough product space.

Ben Thompson (00:42:52):

You’re talking about retailers, merchants, or agents?

John Collison (00:42:56):

I was just talking about discovery and kind of the demand side. Though also probably retailers.

Ben Thompson (00:43:01):

Yeah. Well, I certainly think, I think the part that would be new, which you were maybe talking about, is this real anticipatory aspect. To go back to Meta ads, is it helps merchants who have a very specialized product find customers that they never would’ve found otherwise. But there’s the inverse of, “I need a very specialized product. How do I find what it is?” Which I think you were referring to before, but to what extent can that not just be an in the moment I need this specific…I remember I needed a server, a piece for a rack to mount this router because I didn’t want to buy a whole new thing or whatever. I had this extra router. And of course there’s some guy in Australia that does 3D prints that perfectly matches this on Etsy or something and it was great. I found this random guy, I’m sure he made a bunch of money selling me a $40 piece that cost him $2 to make. Good for him. But what if an AI should be capable of anticipating that need? So it’s not, “Oh, I have a need, let me go find it.” It’s like, “I know you’re going to need this and let me acquire it.” And that would be very powerful.

John Collison (00:44:12):

In his excellent newsletter, Stratechery, Ben often argues that whoever controls the customer relationship shapes the entire ecosystem. And in mobile apps, there’s always been this interesting tension here with in- app purchases where historically app developers had an intermediated relationship with their customer through app store policies. Over the past year, that layer has started to open up. Mobile developers can choose what they use for in- app payments. Developers now have more freedom, but with that freedom comes new challenges. App store payments were previously handling a whole bunch of different tasks like payments, tax, fraud, disputes, all bundled together. Stripe managed payments is built for this new world, handling all that operational complexity for you with Stripe as the merchant of record. And with our new app to webflow, customers can check out in seconds in an experience that feels native. It converts like in-app payments, but it runs on Stripe.

John Collison (00:44:22):

The public markets indicate as of January 2026 that SaaS is canceled. Are they right?

Ben Thompson (00:44:33):

I think it’s probably a mix. I think one of the brilliance of American business is… Actually, this is one of my theories about why the Europeans are so gung-ho about data privacy and regulation is because they so often interact with European companies. So I was in Paris a couple years ago, and of course going on a tourist trip, going to the Louvre, going to the Museum of Modern Art or whatever it is, just seeing a bunch of museums. They all have their own homegrown registration systems and they’re collecting so much data.

John Collison (00:45:12):

And they’re wildly insecure.

Ben Thompson (00:45:12):

It’s like, yeah, what’s your age? What’s your pet? What color is… It’s like, why do you need to know all this information? They’re all non-standard forms. This is where you need AI to fill all this sort of thing in. And there’s this theoretical idea in their head, “If we capture this data, it could be useful.” So they built these homegrown things in the 2000s that are horribly insecure and I use them. I’m like, “Where’s the regulator? This is ridiculous.” So I get the mindset. US companies don’t do that. US companies are so good. I think one of the big strengths of US business culture is understanding, and I think about this personally, this is when I give life advice. What’s the number one mistake people make when they’re young in particular? They focus on their weaknesses. They’re like, “I have to ameliorate my weakness.” I’m like, no, what you do is you double down on your strength, you get richly rewarded for that, and then you hire someone to take care of your weaknesses.

I’m a big believer in the Getting Things Done system. Great book, Getting Things Done. Even if you don’t use the system, the book is really good, lots of great insights. And there’s this whole thing like tickler files and all these sorts of things. It’s an amazing system. I’m completely incapable of managing the system on my own. So there’s a Mac app called OmniFocus that is completely built around this system that I don’t have a license for, my assistant has a license for. And I text him stuff and his job is to maintain my Getting Things Done file because I can’t do it. What do I do? Actually, my life is very, very optimized around, I write three pieces a week, I do an interview and I do three podcasts and all my focus and energy needs to be on that. And if I do that, that will make a lot of money and I could pay to fix all my problems sort of elsewhere.

And I think American business does this very well. They don’t waste time and energy on stuff they’re not good at. They double down on what they’re good at and they’re focused on the upside, not on their cost centers.

John Collison (00:47:07):

Probably a result of the very large market in the US.

Ben Thompson (00:47:09):

I think so. And just the competition, it would be in a very large common market. So you don’t have your... You go back to newspapers, they have lots of homegrown stuff. If you’re a publication online, if you’re like me on the internet, I get paid to comment on the big tech companies. It’s probably the most competitive market on earth, right? Lots of people have takes on the big tech companies. And so you have to be super focused. Given that, that speaks to the enduring value of just paying someone to manage these business functions from a software perspective. Now, there’s a lot of SaaS applications, not sure they’re all sort of strictly necessary and worth the price. I like to think people talk about tech having a Big Five. I’d say there’s a Big Six. The sixth was Silicon Valley Inc, which is basically this cookie cutter VC goes to this founder addressing this specific business case with the SaaS business model. Everyone likes to, they get to talk about changing the world and it’s actually the most predictable thing yet. That’s why VC returns compressed, but because they’re also very predictable in terms of like this sort of engine going. A big problem there is they’re all seat based. Anyone’s seat based that is somewhat vestigial that there’s going to be probably fewer seats. And then if the replacement is more small scale, ideally there’s lots of…The internet in general has writing or content is a good example. There used to be sort of you wanted to be in the big pond and everyone in the big pond ate. If you had a job at Condé Nast at one of their magazines, like you lived life well if you wrote for magazines… Today if you want to be a writer, I give advice to people that want to be content producers all the time.

And I’m like, look, you don’t want to be in a pond with me. Bill Simmons is like the first internet sports writer and you don’t want to be doing a Bill Simmons impression on the internet because he got there first. And what you want to do is you want to make your own pond. The internet enables the creation of a million different ponds. So you get to define your own pond, be the only biggest fish in that pond, that’s how you succeed. To the extent AI makes that, I think this is the upside case, is AI makes that possible for more than just content, for all sorts of businesses to be lots of smaller-scale individual entrepreneurs or small teams, all of whom don’t really fit in the Salesforce driven, seat-based model for a lot of these companies. So there might be a big return to self-serve, or maybe they’ll just roll their own because their needs aren’t that large. So that’s more a larger structural change. But the problem is it’s fine to say businesses will be okay as they are. If you’re eliminating the growth, that’s the big problem. I think that’s the biggest issue for all the compression.

John Collison (00:50:00):

Via headcount growth?

Ben Thompson (00:50:02):

Just growth in general. If these are just stable businesses with astronomical stock-based compensation that is predicated on we’re going to be very large.

John Collison (00:50:13):

Yes. I can see two critiques you might have of the software space and why everything’s traded down. One is everyone’s just going to use Claude Code to rebuild their own version in house and so the software moat is less. And the second is actually just that many of these products price on a per seat basis. And so if you’re growing headcount less, on the first—

Ben Thompson (00:50:32)

Or shrinking.

John (00:50:33)

Exactly. Yeah. On the first, like Anthropic just installed Workday. So I don’t think we’re Claude Coding—

Ben Thompson (00:50:45):

Systems of record, that’s the category that is definitely safer.

John Collison (00:50:49):

We see this with Stripe Billing as well. I don’t think anyone’s Claude Coding one of those systems of records anytime soon.

Ben Thompson (00:50:52):

Do you use Workday?

John Collison (00: 50:53)

Yeah we use Workday. I don’t know what to make of the second criticism, but again, it just feels like for a very broad and deep system of record, it’s kind of hard to make the argument that the business is somehow impaired versus a year or two ago.

Ben Thompson (00:51:08):

Right. But I think that’s my point though, is people saying they’re going to zero are wrong, but if the assumption is you’re fine, but you’re not going to be growing indefinitely, like that shift from thought of as being a growth company to being a stable… That’s a haircut. And again, it’s combined with these whole compensation structures that are—

John Collison (00:51:32):

Yeah, you’re now valued on EPS rather than revenue or something. So yeah. Can we talk about your business and Stratechery?

Ben Thompson (00:51:37):

Sure.

John Collison (00:51:38):

So you were very early to the, I mean the sovereign writer concept, I think you were one of the first premium newsletters?

Ben Thompson (00:51:52):

I think so.

John Collison (00:51:52):

Yeah.

Ben Thompson (00:51:53):

Well, so there’s two predecessors to talk about. One is just on Wall Street in general, there’s a long history of faxed-out newsletters and things like grants—

John Collison (00:52:03):

All this research and all that stuff. Yeah.

Ben Thompson (00:52:05):

The difference there was that those were very expensive and a very small addressable market. So the difference for Stratechery is it’s much cheaper and the market’s much larger. The other person that deserves a call out, which I think was the first person to do it before me, otherwise I think I was the first, was Andrew Sullivan who did—

John Collison (00:52:26):

I hadn’t realized he had a paid newsletter.

Ben Thompson (00:52:28):

He did for like a year.

John Collison (00:52:29):

Oh.

Ben Thompson (00:52:29):

The problem is he did it all wrong. “You’re doing it wrong.” He would churn out like 50 posts a day. Right? Just about a gazillion different things. He totally burned out and all that sort of stuff. But that happened to be a great fit for the advertising model back in the day, because you would always go back there and there’d always be new stuff. And I’m sure, I’m sure he drove a gazillion impressions for The Atlantic, especially when he was with them. He went independent. He was pretty successful. I think he did around a million dollars or something like that, but it was this very leaky paywall. It was like after like 35 posts, then you’ll hit like a paywall. And there’s a bit where you’re like, you’re punishing your worst users. It’s very easy to get around. But he was actually very inspirational in how I thought about the model, in that he was hailed as a failure because he burnt out and then quit. But I’m like, he made a million dollars. This is pretty good. I wanted, from the beginning, just thinking about the psychology of this. When I started Stratechery, I had a gazillion ideas of things to write about, and I limited myself to writing a max of two times a week. And the reason is I had the subscription model in mind and when I added the model, I didn’t want it to be, I’m taking stuff away and now you have to pay. I’m like, “You like this so much, if you pay, you can get more.” And so I always wanted to be, you’re paying to get more sort of aspect. And I think that probably mattered more at the beginning, especially because the model was new. My metric I looked at was people who visited Stratechery on days I didn’t post because they were people that were going there hoping I had posted that day and they were leaving disappointed.

And so in this case, usually previously a paywall would disappoint people that they hit it. In this case, the paywall would alleviate their disappointment because they could now get what they wanted. And so I’m like, “If I can capture X percentage of these visitors, it’ll be very good.” One day goal, one week goal, one month goal, failed to reach all of them. What happened was, I actually thought it was not going to work. I was going to have to go back to teaching English or something like that, but it sort of grew and grew and grew. And at six months, I hit my one year goal, which was a thousand subscribers, thousand true fans. It was a $100,000 run rate. And I posted a little—

John Collison (00:54:45):

Took you how long to get to a thousand subscribers?

Ben Thompson (00:54:47):

Six months. And I posted a little note saying, “Hey, model works. My goal is a thousand for a year. I’ve already reached it.” And this is the only step change in subscribers I’ve ever had. In the next 24 hours, I got 250 new subscribers, a 25% increase. What they were was I had identified those people who wanted to be subscribers, they just didn’t trust that it was going to work and I was going to go out of business and take their money. And so once they realized I wasn’t going anywhere, then they all signed up. And those people signed…So I actually had my, my metrics were right, but I didn’t properly calculate the uncertainty, people’s fear of losing their money. So I’m very grateful that now people just sign up for stuff all the time. Of course, I’ll probably go to my grave being most well known for Stratechery, but I am equally proud of the model and that lots of people make a living doing this.

John Collison (00:55:42):

How far do you think this model can go? Again, the defining characteristic to me seems like the unbundling, like maybe 30 years ago you’d have been writing for a publication, whereas now it’s unbundling, it’s the direct relationship with your subscribers, it’s direct monetization and generally paid. I mean, there might be some ad supporting components as well as paid. Obviously, Substack has proven that this is very broad applicability, but how far do you think this goes versus traditional media bundling?

Ben Thompson (00:56:14):

I think there’s a couple interesting angles to this. Number one, I think people, including people in tech, seriously underrate how large the internet is. And like some of the biggest pushback I got when I announced like the Stratechery paid product was from VCs, I won’t say who. It’s like, “Love you Ben, just not going to work on the internet.” And my bit about ponds before, I don’t know that we’ve scratched the limit of how many ponds can be built in the world and you can sort of occupy it.

And the other part of this, the critical piece of this, and AI is actually an important factor here, is the key to the model is your costs. So just as technology enables you to reach everyone, you need to leverage technology to keep your costs very, very low. And so for the first several years, for Stratechery, it was just me. So as long as I could feed my family, I was fine. And this is the problem for the traditional media companies. Their cost structures were not internet cost structures. They were predicated on much higher revenue. And it’s interesting to think about and talk about this because a lot of this is not really applicable. It’s like before. I write about ads a lot and I’m not an ad business. In this case, I write about VC, high scalable companies, but my actual business is very sort of boutique and small and artisan in that regard. And this is a super important point is managing your costs. If you manage your costs appropriately, then the possibilities, but that also means there’s some things that don’t work with this model. Like your traditional classic investigative journalism, six month sort of piece, it’s not well supported by this. What did support that was the bundle. Having lots of different writers in one publication altogether. And the thing I worry about, I wonder about is bundles are good for everyone involved and no one wants to be a part of it. So TV is the classic example. Why did we have a TV bundle? Because you have like…So in, I think it started in Pennsylvania. So you have a television station in Philadelphia and you have the Allegheny Mountains and you have a bunch of towns there that want to get the signal from Philadelphia, but they can’t get a good signal. So they band together, they put up a big tower to get the signal, they run actual cable from that tower to all their houses. And all of cable television started in small town rural America to get TV from the big cities.

And Ted Turner comes along and is like, “I could just broadcast directly to these towers. This would be amazing.” And you get the model and suddenly, but you had a geographic forcing function and you ended up with all these companies with the best business model in the world. Everyone paid them whether they watched you or not, and they made a ton of money. And what happened the moment they could do something different? I could also go directly. I could stream directly. And there’s just something about business because it’s almost like you have to be forced into the game, the optimal game from a game theory perspective. And the moment you can desert, everyone always deserts, even if it’s the best thing. So I do wonder, can bundles—

John Collison (00:59:29):

How does this apply to your world?

Ben Thompson (00:59:30):

Because in theory, there should be bundles. Substack should be a bundle. You should be able to have one fee and get everything, but they started…I think a mistake Substack made, and I’m a huge Substack fan, just to be clear. I’ve made this disagreement before and they get mad about it, but they characterize themselves as being totally writer-friendly. And I think that was a mistake because it’s impossible for them to be ultimately writer-friendly because the most writer-friendly setup is running an open source software on your own server. Then no one can do anything to you.

John Collison (01:00:00):

I thought you were going to say the most writer friendly thing is to have a humming consumer business.

Ben Thompson (01:00:05):

It would be. The problem is all their initial terms made the bundles impossible. And all the individual publishers owning their own subscribers, having their own Stripe account and all these sorts of bits and pieces.

John Collison (01:00:17):

But hang on. I feel like there’s a well-trodden path in tech here where OpenTable started as entirely, it started as on-prem, purely software for restaurants, and then they added the customer discovery layer on top of it. Shopify started as a solution just for merchants and then they added the Shop Pay kind of network file layer for consumers on top.

Ben Thompson (01:00:38):

Even with those businesses though, the Shop Pay bit is nice. It’s not the driver of the business.

John Collison (01:00:43):

It’s a pretty cool part of the business.

Ben Thompson (01:00:45):

It’s cool. I like it. But at the end of the day, the vast majority of shop interactions are I see an ad on Instagram and I go there and I get the Shop button, which is incredible, one of the greatest things.

John Collison (01:00:54):

No, but yeah, but then it makes the Shopify offering to merchants so much more compelling because you get the Shop Pay network. And it’s where I’m going with this—

Ben Thompson (01:01:01):

I’m skeptical that’s the driver of the business. I think it’s a nice to have.

John Collison (01:01:05):

But can Substack just add a Substack Prime bundle on top of it and merchants can—

Ben Thompson (01:01:11):

The problem is the merchants who will make that bundle valuable have no incentive to join the bundle because they could make more just monetizing users directly. So imagine I’m on Substack. How much more revenue does Substack have to give me for me to trade $15 a month from my subscribers for a smaller amount from whoever’s part of this? And so the problem is they have to really pay me off to be a part of it. Meanwhile, everyone who doesn’t have any subscribers, of course they love to be in the bundle. So this is why the geographics—

John Collison (01:01:47):

This feels solvable to me.

Ben Thompson (01:01:48):

I think it’s solvable at the beginning.

John Collison (01:01:51):

No, sorry. It feels solvable now.

Ben Thompson (01:01:53):

I think a counter example is something like Spotify. Spotify is arguably the best bundle on the internet. But the reason why they were able to assemble the bundle is because they only need to negotiate with four entities. And so it’s interesting because on one hand that limits Spotify’s upside because those entities are able to negotiate such a large share of Spotify’s revenue. On the other hand, that’s also why Spotify was possible because they only need to negotiate with four.

Ben Thompson (01:02:21):

If you’re trying to get every artist on earth, well, of course I got to get Taylor Swift. Okay, good luck with that. All the small fry will sign up. But the music’s unique in particular because music, the moment a song comes out, it’s now part of the back catalog. And actually people only ever listen to back catalogs. So it’s a particularly unique industry in that regard. But that is a bundle that formed, but I think it’s because there’s only four players.

John Collison (01:02:48):

Okay. And how do you use AI in writing Stratechery these days?

Ben Thompson (01:02:55):

I think it probably replaces what I used to do a lot of on… It’s much more efficient Googling. The most gratifying articles I write are when I write about a topic that I usually don’t, and then someone from that industry is like, “Wow, that was good.” Because you’re always worried about—

John Collison (01:03:12):

You have this imposter syndrome.

Ben Thompson (01:03:12):

No. I mean, fortunately I don’t really have imposter syndrome.

John Collison (01:03:20):

That’s why it all works.

Ben Thompson (01:03:21):

What’s the mechanism where you’re reading something about your area of expertise?

John Collison (01:03:26):

Gell-Mann amnesia.

Ben Thompson (01:03:27):

Yeah, that’s right. It’s like, “This is totally wrong.” And then you trust everything else. I don’t want to trigger Gell-Mann amnesia amongst anyone. So if people ask me, I hate the book question, like what books do you read? I read a lot of books, but they’re very targeted. I’m a very, very fast reader. So sometimes I’ll write an article and I know there’s a pertinent book and I will just read the whole book in the morning. But in general, I really want to make sure I fully understand a space, particularly if it’s new that I’m writing about. This is partly why I have a big competitive advantage. I’ve been thinking about tech since I was in junior high school and I’ve been writing about tech for 13 years. So I’ve already done so much preparatory work that anyone starting from scratch, it’s hard. But something I want to dive into. I’m one of the world’s greatest Googlers. I’d like to think. I know every sort of parameter and how to find… So I think I can say it pretty authoritatively. Google has gotten worse. And I don’t think it’s Google’s fault. I just think that it’s harder. One of Google’s faults is they got so biased towards recency and so you have to be super diligent, but AI is so incredible for this. Just sort of getting background, making sure you understand an issue, the ins and outs of it, how things work. You can query stuff, dive deeper. So that is by far my number one use case. I do, not always, but I will sometimes ask it to—this is where I like ChatGPT. I type in BBEdit as an integration. This is also why I’m very annoyed by and am very sensitive to the cloying nature. “Oh, this is really great, but no, that’s not what I’m asking for. I want you to actually go in and find stuff.” So I do not use it to actually generate any exact content, but—

John Collison (01:05:14):

Okay. So targeted research and then critique.

Ben Thompson (01:05:16):

Yes.

John Collison (01:05:17):

Yeah.

Ben Thompson (01:05:17):

Those would be the two biggest use cases.

John Collison (01:05:20):

You’ve written a lot about the TSMC break, this idea that the limiting factor on all AI expansion is basically the rate of TSMC capacity expansion because all AI chips are fabbed at TSMC. It seems like as you look at the AI space and everything interesting going on, so for mostly chip constrained right now, which would not have to be the case, you could be power constrained and stuff. But if you’re chip constrained, there’s a population of people who want to expand very quickly, the AI labs, NVIDIA, people like that. And then yeah, famously, which TSMC, which is more conservative in how it expands, why is that? Why does the market signal not cause them to build out fab capacity faster?

Ben Thompson (01:06:15):

Because the risks for fabs are basically larger than for anyone else. You’re spending billions and billions of dollars on a fab that if it’s not fully utilized, if you end up with too much capacity, number one, all your costs are locked in. Like basically 99.9% of the cost for a fab is depreciation, which you’re paying the depreciation, like you already showed you paid in cash flow obviously, but it’s on your accounting statement no matter what. So the fabs can be extremely profitable. TSMC’s margins are higher than ever, but they can very quickly tip over into having a huge problem. And then once it’s already built, these fabs can run for a long time. So that excess capacity depresses prices for years to… I mean, we see this in memory all the time. Memory famously goes through these cycles.

Like what’s going to happen, believe it or not, we are going to have too much memory capacity in a few years because we have such a shortage right now. Micron just announced they’re a huge new fab in Singapore and everyone’s going to do that. But why does it happen to memory? There’s three competitors in memory. If Micron doesn’t do it, SK Hynix will. If SK Hynix doesn’t do it, Samsung will. And so you have a dynamic where—a healthy dynamic—which is the fabs know better, but they can’t help themselves. And so they take on the risk and they build these fabs. The problem we have with logic is that TSMC doesn’t have that pressure. And so they’re actually behaving rationally. TSMC is giving up potential long-term revenue, but the downside for fab in particular is so large that they don’t want to realize that downside.

John Collison (01:08:04):

Can they not pass the risk onto the customer where it’s like, “You are going to pay for the entire fab.”

Ben Thompson (01:08:08):

That’s probably what they need to get to. And so Apple famously did a lot of this sort of prepaid and particularly when TSMC was expanding hugely in the 2010s and they maybe need to get even more explicit about that. But I think the better solution and the cheaper solution for the hyperscalers in the long run would be to do what is necessary for TSMC to competitors. Then you get it for free.

Ben Thompson (01:08:38):

You don’t need to prepay it. So there’s this risk that’s out there, this risk of overbuilding. Right now TSMC is shifting all that risk to the hyperscalers, to NVIDIA, to Apple. And the way it met, and the reason why they get away with it is because the risk is foregone revenue. It’s money you don’t make. And worse than that, it’s money you don’t make four or five years down the road. And everyone, like what does every company say on their earnings call right now? We could have made more, but we don’t have enough supply. And if you think it’s bad, why is it bad right now? ChatGPT comes out. Every hyperscaler starts investing like crazy. What does TSMC do? They actually decreased their CapEx year over year, two years in a row. There was no market response from TSMC to the ChatGPT moment. Now they increased to 41 last year. They’re going up to like 60 this year, but even that increase to 60 is a less percentage increase than last year. I think we’re looking at a massive shortness in chips in 2029 or so.

Ben Thompson (01:09:41):

Particularly as the other thing, the compute density of AI is so much larger, right? If you have an agent out doing stuff, it’s doing so many more computations in a limited amount of time than me and my Googling is even humanly possible to do. And all these lookups. And so we have a CPU shortage too. And Intel shut down some of their CPU plans, right?

John Collison (01:10:04):

Yes.

Ben Thompson (01:10:05):

So the whole semiconductor, I just think it’s a big problem and we’re shifting to, for a long time it’s like, how can we get an alternative to TSMC for geopolitical reasons? And the truth is this is kind of like the bundling thing. It’s really hard to get companies to buy insurance, particularly when the insurance is… Number one, you have like everyone else wants someone else to do it, right? Who’s going to be the one to go and make the sacrifice? But also it might not happen. China might not attack Taiwan. And also, as long as it doesn’t happen, it’s super suboptimal to go somewhere else because TSMC is better.

And their customers, it’s not just their fabs are better, their customer service is better and they have all the IP blocks you need and they’ve done this before and you have an existing relationship and they’ll punish you because they have limited…They have control because they’re not going to fulfill all their orders right now because there’s so much demand. And so they can pick and choose sort of who… And so people are scared. They don’t want to go anywhere else. And so how are we going to solve this problem? And I think I actually wrote on the front page of Stratechery this week, which is basically the same thing I wrote in an update, but this was a… The hyperscalers in particular need to appreciate, I think a massive crunch is coming and it’s now on them to get Intel up to speed, to get Samsung up to speed, to get a credible alternative. Yes, in theory, you could pay the—

John Collison (01:11:33):

For geopolitical reasons? Or for shortage reasons?

Ben Thompson (01:11:35):

No, we’ll get the geopolitical reasons for free.

John Collison (01:11:37):

Okay.

Ben Thompson (01:11:37):

I think there’s massive economic reasons to do so, which is all the revenue you’re going to be foregoing in 2029 if you don’t do it now, and then we’ll happily get geopolitical insurance for free.

John Collison (01:11:49):

But if TSMC are the best, rather than like standup Intel, which seems hard, isn’t the answer to just again, prepay for an extra fab build out for—

Ben Thompson (01:11:58):

But this is like, how do we feel in tech about ongoing operational costs as opposed to putting in some money up front and fixing the problem permanently? The market structure is a problem. You’re dealing with a monopolist and not like a mean monopolist.

John Collison (01:12:11):

Yeah, exactly.

Ben Thompson (01:12:13):

They’re very nice, right? And they actually have not arguably not raised prices nearly as much as they should have, but the reality is there’s this market structure problem that is going to impact the hyperscalers and it behooves them, I think, to fix the structure. Otherwise, the costs of ensuring that or overcoming that are just going to be larger and larger.

John Collison (01:12:35):

This seems like the topic you have felt strongest about in the past year or two.

Ben Thompson (01:12:39):

I felt pretty strong with Apple Vision Pro.

John Collison (01:12:41):

Okay, fair. What was your take with Apple Vision Pro?

Ben Thompson (01:12:43):

They finally showed an NBA game, and they kept changing cameras. They’re applying 2D television production techniques to an immersive technology. Just let me stay court side.

John Collison (01:12:53):

The TSMC break seems like a bigger deal.

Ben Thompson (01:12:55):

Oh, probably.

John Collison (01:12:56):

I have some rapid fire questions, or I’m not going to say rapid fire necessarily, but more a collection of disconnected questions for you.

Ben Thompson (01:13:06):

I’ll connect them. That’s what I do.

John Collison (01:13:07):

Great. How should schools do homework now that AI exists?

Ben Thompson (01:13:13):

I think they should incorporate it and they should probably do in-person exams.Yeah, I mean it’s silly to try to crush it out. I’m very opposed to these AI detectors because they don’t work. I mean, probably I’m particularly sensitive to it because obviously a lot of my prose is in these models. My thing was I wasn’t an em dash user, but I’m the world’s biggest semicolon user.

John Collison (01:13:43):

I was a big em dash user all along.

Ben Thompson (01:13:44):

Oh, fortunately the models don’t seem to have really incorporated the semicolon. I haven’t been that influential, but yeah, no, you want kids to use it because whoever can use AI most effectively in their jobs going forward is going to have a big advantage. So there’s probably some return to more in class being more important. I think this is my view on content generally. I think there’s a world in which not all content, but some content is more valuable than ever because AI is a perfectly individualized experience. What you read is not necessarily what I read. So stuff that we both read is actually compelling and I’m very interested in figuring out how to leverage that to be sort of beneficial to people in the long run and what can you get from school that you can’t get elsewhere, right? I can read the notes, I can read X, Y, Z, but there’s being in class, having a discussion about it, like actually interacting, being pushed on these sorts of things. All this is a sort of a beautiful theoretical depiction of what school might be that is probably very far removed from the reality. But identifying things that are common experiences are going to be more and more valuable, common content, common classroom time, live events, like shared experiences, because anything that’s individualized is just going to be completely swallowed.

John Collison (01:15:10):

Yes.

John Collison (01:15:12):

Do sports teams become more valuable in an AI abundance future?

Ben Thompson (01:15:15):

Of course. Everything live becomes more valuable.

Ben Thompson (01:15:17):

That’s something I’m thinking a lot about as far as my business. There’s some aspects of tens of thousands of people reading the same thing every day. That is actually really powerful. There’s something interesting there, the possibility of doing live events where people can come together. I think a lot about community. I think no one’s going to really solve community around content, like a message board or comments or not. You actually get very bad dynamics. There’s a few people that dominate it.

John Collison (01:15:41):

Totally. Yeah.

Ben Thompson (01:15:42):

But what is great is if we’re in a group chat and you share an interesting article and you have a discussion about that. So there’s a lot of stuff around that that I think is really interesting and that I’m thinking a lot about.

John Collison (01:15:52):

What do you think of what’s going on in crypto these days?

Ben Thompson (01:15:55):

What’s crypto? No. I’ve always been a crypto defender. Just because digital scarcity is fundamentally interesting. It’s probably even more interesting to this point in a world of infinite content, which you thought we had infinite content before, now we have infinite content on steroids. Not just a six billion humans typing away, but agents generating stuff sort of constantly. And in that world, I think crypto as an identifier of authenticity is going to be more and more important. At the end of the day, I want the original, I don’t want a reproduction. And I’m optimistic about humans’ ability to create value where it seemed impossible to ever exist. I’m literally a professional podcaster and content creator and get paid a lot of money to do it. Imagine explaining that to someone on the farm worried about automation.

John Collison (01:16:51):

Speaking of that, you mentioned that a majority of Stratechery consumption is now in the audio form rather than the written form.

Ben Thompson (01:17:00):

As far as I can tell. I don’t do, but well more than half my subscribers are subscribed to—

John Collison (01:17:05):

I consume it in the audio form.

Ben Thompson (01:17:05):

Yeah. It’s quite interesting. This is actually where I started building my own software. I was begging everyone to support paid podcasts. There was dedicated paid podcasts and there was like writing ones and no one would do it. So of course I had to just hire engineers and build it myself, at which point it obviously was the right thing to do. Now everyone does it. Whatever. That’s my fate in life, I guess. But the…Yeah, people love it. The interesting thing is, I’m not sure it’s been good for my business.

John Collison (01:17:36):

Why? Oh, because people don’t share.

Ben Thompson (01:17:36):

The good news is, I think it drives retention because people would build up emails because it feels like a lot of work and they say, “Oh, I haven’t read this in ages. I can unsubscribe.” Whereas they just consume seven minutes or eight minutes. The problem is they don’t share. Audio content is not shared.

John Collison (01:17:51):

Totally. I listened to it in the car on the way home from work and that’s great. And then I—

Ben Thompson (01:17:57):

Never think about it ever again. But it’s great for me because I can say the same thing the next day and you’re like, oh, that was a very insightful comment. You didn’t even know I said it yesterday.

John Collison (01:18:06):

Yeah. If we reason about what sectors are going to be important down the road, for the AI build out, energy is going to be a big deal and the ability to actually power the data centers that are coming online. That may be a bigger constraint going forward than even chips. Robotics are clearly going to be a big thing. It seems like China is doing better on energy and better on robotics and is catching up on chips, doing okay on the AI models, but does that mean China’s potentially very well positioned for the coming wave of tech trends?

Ben Thompson (01:18:44):

I mean, I think any country that is capable of actually building things is well positioned. But then again, the counter argument, if I could sort of put a silver lining on it, is the challenge, the trick going forward and to sort of defy the doomers as it were, is actually creating new sorts of value, new sources of value in a way that humans are uniquely capable of. And that is by definition a sort of an innovation story. It’s a freeing up resources from things that can be done by machines to more productive…It’s having a consumer market that pulls out that sort of innovation, that makes it possible to write a newsletter or a podcast and actually pay for it. And so there’s a scenario where China is well positioned to win the total commodification of everything, which doesn’t have much margin and the actual value creation and what makes humans humans and generates the value that I think people in AI are skeptical can be created, despite the fact 90% of us used to work in agriculture and like 1% do. For some reason, that’s not going to repeat. If you want to be optimistic, that’s the sort of thing that America has always done well.

John Collison (01:20:08):

What’s your Stripe feedback for us?

Ben Thompson (01:20:10):

Oh, where to start? I mean, I’m obviously, it’s hard for me to write about Stripe because I’m not biased because I was very early. I think you introduced the billing API in 2011, which is a direct spur for wanting to do Stratechery and thinking this was a business model that was possible. So very, very big thumbs up on that. Oh, you didn’t warn me about this. I should have thought about this. Actually, you have one huge issue that I was just dealing with. Oh yeah, ACH. Your ACH implementation is…Someone can go in and if I try to add on to an ACH plan, so say I have a team, because that’s where you use ACH, like large companies, and they want to add someone on. If that add-on fails, the entire plan gets canceled. So we have to build a bunch of logic to handle that independently.

John Collison (01:20:56):

Okay.

Ben Thompson (01:20:57):

That’s a very detailed specific problem that we’re facing.

John Collison (01:21:00):

Buggy, ACH subscription directions. Okay. That’s a good one.

Ben Thompson (01:21:04):

There’s definitely more. I’d have to go back and think about it. But I mean, I do think the… We didn’t talk about stablecoins in this sort of area of… I’ve always been a big skeptic of microtransactions because the problem is, it goes to the investigative reporting thing. You can’t build something sustainable if you’re only monetizing on the back end. And the only way to do that is to have a very large market, which is what YouTube is. YouTube is a bunch of speculative video makers hoping that they’ll get enough views that the ads will pay for it. And there’s such a large scale and they monetize their ads so effectively that it works. There’s no market like this for written content or like podcast content. And you can’t…People are like, “Oh, let me pay for one article.” I’m like, no, what you’re paying for when you pay for me is you’re paying for my ongoing production. I’m making a promise to you I’m going to write something every day and you’re paying for that promise. You’re not paying for the actual content. The content is a byproduct of that. The question for AI and microtransaction is you have all these labs paying people all over the world to generate data. And if you’re like a radiologist, you get paid $350 an hour. I saw some article about it and all these sliding scales and they’re all duplicating work because everyone feels the sense that I can get differentiation. What we clearly need is some sort of market mechanism for data generation that in the long run will replace what we’re getting from journalistic enterprises, which are even more doomed than ever before. So how do you generate? We’re paying directly for content and then AIs can get it and can…You can build a large market like YouTube that people will speculatively do it, trusting that they’ll get paid because the market is large enough. That’s what needs to happen. Like a lot of things, there’s this massive value of how do we get from here to there, but we’ll see. I know Cloudflare is trying to push on that, so we’ll see what happens. Yeah.

John Collison (01:23:00):

Last question. How would you rate the execution of the major tech companies?

Ben Thompson (01:23:04):

Like the Big Five?

John Collison (01:23:05):

Yeah, sure.

Ben Thompson (01:23:07):

Apple, traditionally very strong. Their manufacturing obviously remains amazing. Like the iPhone Air that just the alarm went off and made sure I turned the snooze off. The greatest smartphone ever made.

John Collison (01:23:13):

Really? Oh yeah. I’ve never even seen one. It’s thin. Is the battery life good?

Ben Thompson (01:23:24):

Good enough.

John Collison (01:23:26):

It’s bad.

Ben Thompson (01:23:27):

What’s that?

John Collison (01:23:27):

It sounds like it’s bad then.

Ben Thompson (01:23:28):

No, it’s fine. I mean, I actually forgot my external battery and it’s doing okay now. And now that I’m back in Wisconsin I have to wear jeans because it’s cold and it slides right in. Actually, I love it.

John Collison (01:23:38):

Okay.

Ben Thompson (01:23:38):

I’m very devastated to hear they might not make it regularly. Obviously, Apple’s software has gotten pretty rough. Their relationship with the… I mean, Apple is so interesting because the reality is when it comes to platforms, you have to build…The price of becoming a platform is making a great product. So Apple gets platforms because they make great products and they’re terrible stewards of the platforms. Microsoft is a great platform steward, but they can’t make good products so they never get the permission to sort of have big platforms, which is sort of a tragedy there.

But Apple, it’s an old company driven by managers, not founders, and maybe they …The AI Siri got as bad as it was, is obviously really bad, but at the end of the day, we still need devices. They’re still better than anybody else, so they’ll probably be okay. Google, I’ve had the hardest time understanding Google, in part because I think Google does a lot of stuff suboptimally. Almost everything I feel like they do suboptimally, but I think that lack of… Apple can be super optimized, but I think it’s their lack of optimization that actually makes them maybe the most resilient of all the tech companies because they never get so exactly doing what they should do and they have all this extra fluff and doing things and gazillion science projects, but because their core business model is so good and throws off so much cash, they can just sort of like be sort of very flexible and I’ve come to appreciate that about them.

Everything that frustrates me in analyzing them actually has this hidden benefit of resiliency and strength and adaptability. And they’re like the …What’s the slime that just …And if they’re coming in your direction, like you’re actually in big— it might take them a really long time to get there, but when they get there, you’re doomed. So Microsoft, I’ve gotten a lot of mileage writing about Microsoft. Everyone, especially in the SaaS era, all these companies are like, “Oh, Microsoft sucks. We’re going to make the best of breed product.” And guess what? Startups in Silicon Valley, they want to buy all the best of breed products and they have the abilities to come together. Joe in managing the tire shop doesn’t care about that. He just wants this crap to work and to work together. And if it’s all mediocre, but it kind of works together, that’s better than best of breed.

And Microsoft is just squashing these companies that grow and boom, just hit that Microsoft wall again and again. Is that going to persist in an AI world? It’s probably tied to the SaaS question before in some respects. Their distribution and power there remains substantial. Meta’s probably, in my experience, been the best execution. I mean, you just see stuff like interacting with PR or executives. They just run such a tight show.

John Collison (01:26:38):

They’re really on it, yeah.

Ben Thompson (01:26:38):

That’s always been very impressive to me. I think their ad model is underrated. The trick with them is keeping engagement. That’s what makes the whole thing go. They’ve done a decent job of that. Hours spent in ChatGPT are hours not spent on Instagram or not spent…And I think that that’s an underrated area. And I think they’re kind of betting that, look, that’s all fine and well today, but in the long run, this is an infrastructure game. We have cashflow to fund it and OpenAI doesn’t. I think OpenAI might be a bigger threat to Facebook than Google, something worth considering, but Facebook is obviously clearly spending to meet it. So Amazon. Amazon, there’s a lot of fab capacity and power being spent on Trainium that one wonders could be better spent on other chips, but we’ll see what happens.

John Collison (01:27:34):

Aren’t people happy with the Trainium chips?

Ben Thompson (01:27:37):

The degree to which Amazon optimized cloud computing, I think is underappreciated. When you’re operating in a commodity market…So there’s two ways to succeed, right? You could have a differentiated product where you can charge a high margin, or you can have a lower cost structure in a commodity market where the price floor is the market price, but your cost structure is lower than your competitors, so that’s where you make your margin. That was how Amazon dominated the cloud. Their cloud was way more optimized than anyone else’s. The whole Nitro architecture, just the way they architect everything, doing a lot of their own chips, shifting to Graviton. I think the thing with Graviton, their arm CPU is they could…Who’s the number one customer for Graviton? Amazon itself. And so they can move all their loads to that, optimize it, build all the software libraries, and then start offering it on a cost-cut basis to others. That’s the playbook that they’re trying to run with Trainium, where the number one customer of Trainium in the long run is Amazon, but then they develop all the capabilities around it for other people for it to be attracted to other people at lower prices, and they have that structurally smaller cost structure.The problem is that works when you’ve sort of leveled off in performance, right? Amazon executed this model between 2005 and 2025. Of course, processors got faster in that time, but it wasn’t like the ‘80s or ‘90s when every leap was massive. Does that work in a relatively new market when there’s massive leaps being made generation on generation? And they have NVIDIA servers. Do they have as many as they could because they’re on this strategy? Probably not.

John Collison (01: 29:17)

Ben, thank you.

Ben Thompson (01: 29: 18)

Thank you.