WWDC Special: Asymco Office Hours - June 2026 - Live Q&A with Horace Dediu: Recording & Transcript

Recording and transcript of the June 9, 2026 Asymco One Office Hours, recorded the day after Apple's WWDC 2026 keynote. Horace Dediu shares his impressions of the event and takes questions from Asymco One subscribers.

Recording

Subscribe to Asymco One to listen to the recording and read the transcript.

https://asymco.com/wp-content/uploads/2026/06/Office-Hours-June-9-Recording-Final.mp3

Transcript

Farshad Nayeri: Okay, we're live. Hi, Horace.

Horace Dediu: All right. Hello. Hi Farshad. Hi everybody. I think we're going to spend most of the time in Q&A, but before we do, I wanted to give some of my impressions. Granted, it's still quite early. It was an interesting event, because at first it seemed very different. There was a lot about safety, a little bit about architecture, and a huge amount of fairly slow demos — not the usual. It felt a bit slow and a bit dense, and I was a little puzzled. I was watching it without a lot of enthusiasm. But after reflection, and after reading some of the documentation that was released, I came to think this is probably one of the most important WWDCs ever.

This isn't taking into consideration some of the great ones, of course. I remember being at the one in '08, when Steve Jobs told everyone there wouldn't be native apps — which actually drew a lot of boos from the audience. We were going to have web apps only on the iPhone. But the other big one was two years ago, in 2024, when Apple Intelligence was first announced. In many ways this felt a little bit like that event all over again. The difference was that they addressed an architectural question.

So I have a note I'm going to publish, and the key element of it is that there's effectively a move toward what might be considered the next generation of operating systems. Not necessarily that there's a new version of iOS, or that iOS gets replaced, but something that moves in the direction of what I call intent computing. We have touch computing, and we have spatial computing in parallel, but somehow we're moving — within touch, or maybe across all these platforms — toward an architecture for intent: computing according to an outcome that's desired, as opposed to a tool or a platform of options available for you to pick and choose from. That's the app model we're familiar with. Moving beyond it, we're looking at what Siri always promised. You want to do something, you make an inquiry, a request, and then it decides how to serve you.

It decides, firstly, by parsing what type of input you're actually communicating through — are you visual, are you audio, are you text, are you capturing an image? Then your intention is parsed or understood, and it's passed to different models that can be used to help you. Those models are assembled, and some of them might end up being enough, but in some cases it might go on to connect with an app that finally completes the request. So the app becomes something of a module that plugs into this intention model. And Siri is the other way of having this experience of getting a result — the Siri model would be exclusive of apps, but the app model is still there.

What's interesting is that the app now changes character. The app becomes a module. In many ways this integrates into the operating system the execution modules and the data models necessary to complete an inquiry. This is why, in many ways, this is moving Apple into a more integrated model, and into becoming far more a combination of Google's search and a set of core applications. It's quite profound. The implications are also profound — architecturally, but also in terms of economics. How does this get paid for? Because there might be a lot of computation needed. And how does it get monetized? How do users pay for some of that cost — presumably more than it actually costs, thus generating a surplus that can be distributed among developers, model suppliers, and Apple itself.

We can get into a lot of this, but I thought that was very profound. If you wanted to see it, it was that concentric set of rings illustrating the user in the middle, surrounded by devices, then surrounded by — I forget the name, but essentially the types of media or input — and then what's called the orchestrator, which connects to the frontier models, and then finally whether the experience is delivered via Siri or apps. This concentric model is very different from the stack — silicon, hardware, operating system, services — that we're familiar with, where the point of integration was that Apple had the hardware, the operating system, and increasingly more system services that it opened up to developers, but the developers still had their apps on top. This model somehow puts the user in the center. By the way, in the stack model the user is at the top, because they're interacting with the apps only. Here we're going to see the user interact not even through their own inputs, but perhaps through their environment, through what they're observing, as an input to the whole system.

It's a fascinating and far more profound architectural development than I had anticipated. I was thinking more along the lines of nice tools, maybe extensions, APIs, and so on. But we saw a lot more than that. We saw a hint toward what I call intent computing. In the piece I'm about to publish, I talk about the orchestrator — I call it the system orchestrator — and then compare and contrast intent computing with spatial computing. What happened to that? I think spatial computing is, if anything, a great foil, a great counterweight to AI-oriented computing. It takes almost the exact opposite point of view. And yet spatial isn't going anywhere; Apple just doubled down on it, in terms of support for the creation of media and applications, bringing Siri and AI and a lot of updates to visionOS.

So it's a very different dichotomy. One is passive, one is ambient; the other is immersive — it's about every pixel that you see. Everything that hits your retina is computer-sourced, computer-generated, or computer-enhanced, versus a more low-bandwidth, passive environment. We debated whether these two things mean we have to choose: that the future is going to take touch into one of these domains, and first, that touch goes away, and second, that it picks one of the two. I argue it's none of those. Touch doesn't go away. There is more intent computing, but it's integrated within a touch environment — in fact integrated into a touchless environment as well, in the case of wearables, and integrated into an immersive experience, which is like a Mac, with even more of a sit-down, commitment-necessary kind of experience.

We won't get too far down this road, because I'm going to eat too much time, and it's going to be written — it's coming out very soon. But these were my initial impressions. Very, very big changes. The market isn't even capturing the improvements in Siri that we're seeing. It's totally missing the point, which is not unexpected — they've always missed the point, as far as I can remember. Anyway — Farshad, any questions we want to deal with right off the bat that have been submitted in advance?

Farshad Nayeri: Bob just asked a question in the chat. Everyone else — hello, and thank you for joining. Please feel free to queue your questions in the meeting chat, and if your mic is working you can ask yourself. So, Bob, please go ahead. Oops.

Bob V: Okay. I was just thinking back to when the iPhone 4S came out, and the rumor mill was saying that Siri was its own operating system that would run alongside iOS. You were talking, Horace, about how the vision you could see was an intent-based OS, and that intent-based apps were coming. It made me hearken back to that original Siri thing and see it as just — what, thirteen years too early?

Horace Dediu: Exactly. No, there's nothing really new here. I think people who have been tooling up these Claude-bot-type configurations have been using this term, intent computing. I'm not introducing anything new. I also think Siri's dream — back when it was a research project over at Stanford — was very much along these lines. In fact, I started my career in information retrieval, which was supposedly a natural-language interface to query computers before we even had search. The idea was: you don't need to know how to program, you don't need to know how to use a database system, you just ask what you need, and we'll give you answers sorted by relevance. All of that pre-existed in the 1980s. This is an old, old area of computer-human interaction research — the so-called natural-language interface. And we've come such a long way.

Now, Siri is just moving right along, but I'm pointing out that you can pay lip service and create a neat chatbot, or create a great interface that fools a lot of people into thinking they're talking to another human being. But the fundamentals of restructuring the entire guts of the OS to enable this — that's what we started to see. I think this is finally a commitment by Apple to say: we need to re-architect everything in iOS, across all of our OSes, to surface this experience.

I'm not as technical as I used to be, but let me put it this way. I came out with an article that said: show me the money. At the end of the day, none of this is interesting unless we see some economic activity. Because if there's no benefit from this, people are going to treat it like another feature that's unrequested, unloved, and unused. This is the bane of product development — the underutilized or unused feature, unused performance in the product, is a symptom of over-service. You're delivering things, and people are not taking them into use, and therefore are not willing to pay for them.

It was Christensen who pointed this out. He said a commodity is a product where, if you make a marginal improvement in it, nobody pays any more for it. Imagine you're dealing with bushels of corn. It might be that Iowa corn is better than Kansas corn — I don't know. But when it comes to the Chicago exchange, where futures are traded, nobody prices corn from different farms differently. It's all a commodity, because although it might be better, it's not worth more. That's a commodity — not by the classical definition. If you look at a product and you're constantly updating it, constantly making it better, and nobody's using it, nobody's paying any more for it.

So the problem with AI is that it has to be more than that. It has to be something really desirable that people are willing to pay for. Those who are paying today clearly are finding some productivity, something valuable about it. But we're talking about consumers right now. Apple is in a consumer device business — consumer products and services. So if Apple's 1.4 billion or so users are just treating this like a nice-to-have, like emoji or Memoji — it's cute, but not life-changing — then we're not going to see an economic impact.

I do believe this is coming together into something useful. A lot of people will begin to use it on a daily basis, maybe more often, and then we'll see how much they're willing to pay for that extra power. Right now they're going to be paying for new devices; they're going to be paying for the fact that they need more hardware to execute a lot of this. But ultimately I think it gets repackaged into new services. We're probably already seeing that — iCloud itself gives you a certain allocation of compute as necessary for these AI services. In the future we might see those price tiers start to change, and we'll see the introduction of more pricing and more price sensitivity. Anyway, sorry, I'm rambling — but let's carry on. Great story about Siri, though. This is nothing really that new.

Farshad Nayeri: Horace, just an addition — I've been watching some of the developer videos, and it feels like this time around the engineers at Apple took it on and started eating their own dog food, so to speak. They seem much more into how you really get AI into your applications. Very interesting videos. Bob, did you have a follow-up?

Bob V: Yes. I researched the deal between Apple and Google for the frontier model they're using. All that was public, that I could find, was that it's a multi-year deal with no known end date. So this requires gazing into your crystal ball fairly far out: what happens after?

Horace Dediu: When that came out, a couple of things happened — people started to jump to conclusions. I talked to Laura Martin about this, and she assumed Google's going to run this, and they're going to get all the data from Apple, and that's where the gold is. So even if they're not getting a lot of money out of Apple, they're going to effectively begin to control the user's data and experience. I was skeptical at the time that Apple would ever release any end-user data; if anything, we're just going to see a technology licensing deal.

These questions have been answered in the follow-up sessions after the keynote. I've seen the tweets and threads from people who were there. People are saying Federighi and his staff have been adamant that Google gets nothing — there's certainly no data. Only one of the many models being accessed through Core AI is running on Google Cloud, which means it's outside of Apple's own private servers; that's the most powerful model, but it's a world model. It is not Google's — they don't have access to anything. So it's very much a technology licensing deal. The other thing people say is that it is not Google's tools in any way. This isn't their image tool — Farshad, what's it called?

Farshad Nayeri: Banana — Banana, whatever.

Horace Dediu: Banana, that's it. So they're not rolling out any of Google's branded or unbranded tools. It's not Gemini, it's not Banana — it's Apple's own model. I think they call them Apple Foundation Models. There are several of them, and one is running with Google in the cloud somehow, but it's private, it's pretty secure. A lot of people assume way more than what's actually happening with Google here. And, like you said, it's probably swappable. At some point Apple may decide to do it themselves — I don't know what the IP issues are, but there's something interesting there.

Now, let's not forget that Google also sends $20 billion-plus to Apple. So now Apple sends back one. What's a billion between friends? Anyway — that's another area where I think people were jumping to a lot of conclusions. Even Apple's response to the initial queries on this topic wasn't very explicit about whose computers these are, whose data this is, whose data centers are being utilized, who's paying whom, for what, and how long. They were very cagey about it, and people jumped to conclusions — "aha, I see you didn't say this or that." I wouldn't worry too much about it. Apple's not going to give away the most important thing they have, which is their relationship with their customer.

Farshad Nayeri: I'd add that some very core aspects of model development involve using Google's tensor stuff. So as far as competition with Nvidia, it might be interesting — a way of seeding the market with Google's way of doing things, as opposed to Nvidia's. But it's layered inside Apple's stuff, so you'd have to be very deep in to get at it.

Horace Dediu: We talked about this on the last Office Hours — whether Apple would create its own silicon. Now it's clear they're not there yet for the server side, server-side compute. But ultimately they probably are working on it, and have been for a year or more, to bring an Nvidia-level chip — the highest level of Nvidia chips — to operate, as are many others, not just Apple.

Farshad Nayeri: Right. The battle is on two grounds, though. It's the Nvidia chips and the API. Is it CUDA? Is it the tensor API? So there's something there.

Horace Dediu: Everybody's going to go there.

Farshad Nayeri: Okay. I think Roger had a question. Roger, would you like to ask yourself? I'll unmute you, see if that works. Thank you, Bob, by the way.

Roger S: Hi, Horace. Yesterday I posted a question to the thread, replying back on the invite. Is the P/E of 38 getting too far ahead of itself? And today the market solved my issue and dropped it down to 35, and I feel better — that's what I posted in the comments today.

Horace Dediu: It only took twenty-four hours.

Roger S: It's crazy. If I was unhappy yesterday, I should have acted on it and bought it back today, but I didn't. What do you see as likely monetization, if there's going to be any besides hardware sales of more capable equipment — phones, Macs, iPads? Do you see a monetization here?

Horace Dediu: Absolutely. I wrote this up, and I'm happy to answer it in advance of it being published. Here's how you go about it. They gave us an architecture. It's about intent, about orchestration, about routing the question through multiple models and multiple input types, and then deciding whether it goes through an app, or back to Siri.

When you look at it strictly from a technical point of view — how do you architect it, how do you implement it, how do you do the right thing for the user, for privacy, for everything else — I was frustrated yesterday, thinking, where's the money? My answer was: follow the compute. If this works without compute, or it works on-device, then the user is happy. They're happy buying the device that does the job, presumably with the memory that's necessary, and Apple provides all of that as part of the purchase price. Everybody's happy. The developers probably still sell their apps, but they're not doing extra compute. They're not selling data services or compute. They're selling: here's my app, it has code in it that solves this puzzle for you.

Now, if there's compute — if you have to do a huge amount of computation to deliver the result — on-device it wouldn't matter that much; that just uses battery. But if it's in the cloud and it's through an app, then you trace who's going to send you the bill. Where's that invoice going to come from for the compute? If it goes through a developer or a model supplier, then it's not free, and they're going to end up saying, "I need you to subscribe to a service to get the answer." Once you do, that could be priced as a fixed fee or it could be variable. We could end up in a situation where the app developer requests a subscription for that app to work, and within that subscription they've worked out the economics of how to get paid, and Apple gets its 30%, or 15%, or somewhere in between. If there are other services that Siri depends on that don't use an app, then Siri will probably charge back, through iCloud, for the compute.

So in many ways, all they did is take the exact same business model — services with subscriptions, primarily — and say: we're going to either tweak these subscriptions to include compute, if that's necessary, or add new subscriptions for just that extra compute. And Apple gets the same distribution commission it's always gotten. Nothing's changing. In fact, the absence of any commentary from Apple on how this is going to have different economics is, to me, a tacit acknowledgment that there's no need for anything different. This is going to work within their existing models. That means the App Store, and if anything economically interesting happens, the App Store will have another kink in its trajectory. We got a kink when it went to subscriptions, after that model became dominant — you can see the kink in the App Store revenue rate — and now we might see another kink, which AI enables.

As far as users are concerned, they don't see anything. Maybe their monthly services bills go up by 10%. Maybe they add a few more subscriptions, or maybe they cut one and add two. Who knows? Will there be some churn? But at the end of the day, it all reflects in services. And that's a 72%-plus gross margin. Apple's hardware will also probably see a nice uptick. It'll take time — I reckon it'll take ten years before this reaches 50% penetration. I haven't written this down, but I think it's a ten-year cycle before half of the Apple population are daily active users of AI services, paying something for them. So hardware is going to take its time, software and developers are going to take their time, and they'll just see more of the same, but bigger.

Roger S: Just to fill in there — I've messed around a little with Private Cloud Compute inside Shortcuts, and very quickly I get cut off, or I exhaust my daily limit for the free tier. I'm an Apple One subscriber, so I may have a little bit more. I've read since yesterday that all users will get a little bit of processing, iCloud+ users get even a bit more, and people who want to process a real project may have to kick into another tier. But I think they're going to wait and see how strong demand is versus what they've contracted to process in Google's cloud, elsewhere, and their own.

Horace Dediu: Yeah, that's the thing about AI, which I think is a little less disruptive than people imagine, because the more you use it, the more it costs. A lot of the early users are being subsidized — with venture or low-interest money, or whatever you want to call it. There's been a lot of that going on in all of these services over the last fifteen years. Whether it's food delivery, Uber, ride-sharing — they tended to be much, much better when they were new, because they were subsidized.

The difference is that a disruption is something that goes down in price over time. We're seeing the opposite. We're seeing a lot of stuff that actually gets more expensive. You hear this term, enshittification. All that means is that businesses that gave stuff away for free now charge you, and everybody feels like, "Oh, it's gotten worse." Or they charge you in different ways, they make it harder, they slap on fees, they deprioritize service, and so on. All of that happens because you didn't have a transparent business model in the beginning that was actually clearly profitable. You were chasing growth instead of profit, and that causes people to get euphoric, but you're basically giving money away.

For that reason, the problem with AI is actually the bills coming due for a lot of stuff that was subsidized. That's not the way it should be. If you look at the great disruptions, prices go down exponentially, asymptotically to zero. Think of the cost of electricity, the cost of aluminum, the cost of one unit of light, the cost of transportation, even the cost of food. It's all gone very, very quickly lower and lower as you get scale economies. We're not seeing that. Maybe it'll change. Maybe we'll see some oscillation.

But fundamentally, Apple is neutral on this, because they're saying: we are just the pass-through as far as the costs are concerned. We expose things to users, keep it safe for them, and the user is in what they call the walled garden — it's a wonderful place to be. Whether the business models behind it are variable or fixed pricing, or economies of scale, all that's going to pass through. So I believe this is the ultimate sustaining business for Apple. They just get paid to distribute, which has been my thesis from day one.

Farshad Nayeri: Roger, did you have a follow-up? I thought you had two questions.

Roger S: I do have a different question.

Farshad Nayeri: Okay, go ahead.

Roger S: I thought a new wind was blowing in the EU with regard to Apple and gatekeepers. But on Siri AI not being available — meaning Apple's not going to release it and allow app developers full accessh to a user's data — why does the EU think it's okay for these other apps to have access?

Horace Dediu: This is a hardball negotiation underway. At a higher level, the people who are negotiating are not the leadership — they're functionaries, bureaucrats. Whether there's a shift in the wind, that would have to come at the highest levels. There was a realization, maybe a year ago, that Europe was falling far behind in a lot of technology areas, that they regulated themselves into a dead end. That was expressed at the highest levels, and people took it into consideration as much as they could. I'm in Europe now. But before that trickles down and something happens — it's a bureaucratic system. It's got its benefits, but it's slow.

I think Apple's strategy with a lot of this is to induce a sense of FOMO, fear of missing out, among the people who really do make these decisions, whether they're regulators or people in the courts: look, you're doing more damage to your constituents than you think. Part of that came out through that number I put out, the $1.4 trillion, which I increase with Apple's own products and services — the $1.4 trillion is on top, and then there's Apple's own, so together that's reaching $2 trillion. And beyond that, there are things that don't go through apps at all — stuff that's maybe browser-based. I'm sure the whole porn economy sits on top of that and is invisible to what these analysts are trying to quantify. So the overall economy could be even 50% higher than the $2 trillion, making it about $3 trillion. On top of that, you've got Android at probably about half of that. So I'd say the whole mobile economy is three-plus-one-and-a-half — four-and-a-half to five trillion. That is essentially the US companies, Apple and Google, enabling worldwide. You can break it down by continent or country as you wish.

The implicit statement I'd encourage Apple to make is: if you screw this up, you're going to damage your core economy. A five-trillion-dollar economy is beyond a national economy — it's almost a continental economy, depending on your continent. It's not just "you've got a nice thing here, don't mess with it." It's an observation that this isn't only about competition; it's about keeping the lights on overall. The growth story of the whole economy depends on it. But let's not go too far down this road. I think this negotiation will get resolved. They're playing a lot of games right now on both sides. They're even using WWDC public forums — Federighi going out in public saying "Hey, EU, you're not getting this." That's not targeted at the bureaucrats; it's targeted at the users and developers who are being told they can't even test. Those developers and users need to connect with their decision-makers and point out that they're being left behind. The Indians are going to do better than us, the Chinese — well, China is still out of it, but that's a different story; that's easy to flip that switch. To flip the European switch requires some sort of consensus, which is much harder to gain than in an authoritarian regime. Anyway, I really hate doing too much on this, because it's easy to sound clever about it, but that doesn't mean you are.

Roger S: Thank you.

Farshad Nayeri: Okay, thank you, Roger. Dave Emery unfortunately couldn't join — he's an irregular contributor to these meetings, as well as to the site — but he sent two questions for us to ask. Question one: Horace, are we heading to an AI bust, or at least a meltdown of the large hyperscalers? Or, as the advocates claim, is there going to be a continuing increase in spending that eventually produces revenue? And if and when this crash happens, how will Apple adjust?

Horace Dediu: There's revenue, though. There is revenue — there's just not that much profit. There's probably even earnings, but there's not much free cash flow, because everything's being reinvested. This feels a lot more like the early days of Amazon, which was also an infrastructure play. In the early days, Amazon reported almost no earnings, because their CapEx was huge, and there were fairly little margins on retail. So for years and years, their P/E ratio was ridiculous, because earnings were almost nothing — 400, or at least 100. Remember those years. But eventually they did begin to record earnings, retain some, and especially with AWS being the savior. But what I'm saying —

Farshad Nayeri: But with Amazon, they were purposely setting their margins to zero, right? They had a predatory approach. There are ways you can play games.

Horace Dediu: Remember: earnings are an opinion, cash flow is a fact. So yes, their opinion of what their earnings were was fabricated to some degree. They could manipulate it. They could declare all their software development as CapEx, or rather as R&D, as they did. But the question asks: is there a bust? Let's take that on. I think there's a valuation issue, but if you look at some of the hyperscalers, their P/E ratios are fairly modest — lower than Apple's, in the 20s. Microsoft is getting hammered. Google is way below its all-time high, even though a lot of people say it's doing really well. And they're all growing really fast — their top lines are growing. They're just so desperate for cash that they're even doing private placements, meaning they're selling equity. They're diluting their investors to get more.

If you raise equity, I don't worry. My concern, if there were a crash, would be whenever we get into debt fueling a growth or speculative boom. Using free cash flow, or equity sales, is generally non-contagious — if you get hurt, it only hurts that company. When you're dealing with debt, there's a knock-on effect, a contagion. Almost all crises, all crashes, have been on the basis of debt, not equity, being mispriced. The dot-com crash was actually overvaluation, but that was because there wasn't even any fuel — there were no earnings at all, no revenue at all. You had 400 P/E on some names, which made no sense; the top ten companies were all at multiples in the hundreds. That was a little different, and fairly short-lived. But if you have a crisis induced by debt, then things have a long time to unwind, with repercussions across the economy.

So far, my read is that AI hasn't yet become a debt story. There's some debt in there. If it does escalate, you'd have to imagine a scenario where banks, mutual funds, hedge funds, private equity go out and lend huge amounts to build data centers. The usual lending logic is that there's recurring revenue behind it, so you can pay back the interest. Very rarely do you see a hyper-growth or speculative business being backed by debt. The only exception, of course, is anything to do with Elon Musk, where he raised debt to buy X — presumably the people who lent him money thought it was a predictable business, until he wiped out half of it with bad decisions. But they still got out, because he managed to fold it into xAI, and that got folded into SpaceX.

So — I'm not an expert in this, but my observation is that the AI bubble might burst, but only if people stop using it, if businesses run out of money to pay for it, or decide it's not worth it, that it's actually a productivity illusion or mirage, and as a result people say, "This was fun for a while, but we're not doing it anymore." So far, all the indicators I'm privy to are: boy, I wish we had more compute. We're constrained more by the fact that we don't have enough capacity to meet all this demand. That, to me, does not scream that there's a crisis here. Maybe a lot of that is growth-hacking, or they're giving away too much and not pricing. We're not seeing a pricing signal, and prices are important because they tell you how to allocate your resources. If you mess with the price signal — by subsidy, by monopoly behavior, whatever — you don't know what the market's really thinking, and that's dangerous. So for now, let's see how much truth there is in the pricing of AI. But that's a macro issue. There are other factors on the externality side — a lot of stuff floating around even today; we might be seeing some cracks in the foundations. But as far as Apple is concerned, these are externalities.

Farshad Nayeri: Actually, while we were talking about valuations, there were two questions about stock buybacks, past and future, since that seems to be in flux. This is Joe Bond: in your opinion, what has been, and will be going forward, the cumulative impact of Apple stock buybacks — I guess that's been twelve-plus years? And Kenneth Kruger also asked about buybacks in general. I think we discussed this last time — the question was whether JT would change any of that.

Horace Dediu: They signaled that they might no longer be targeting cash zero, or net cash zero, and that was a wake-up call for a lot of people about the buyback story maybe changing. But at least for the next year, that's not going to change, because the allocation has already been made. The board has approved it, so it kind of has to happen. Barring extraordinary circumstances, that's how I see it. So we might be seeing a very gradual transition away from the return-capital-to-shareholders strategy we've seen — partly influenced even by Warren Buffett's philosophy; maybe he rubbed off on Tim years ago.

But fundamentally, this has nothing to do with the business. Whether the business chooses to keep the money or give it back to its owners doesn't change its prospects as a business. It's just what it does with that surplus of capital. There are old adages of finance, like: whether the company is financed using debt or equity doesn't change its prospects. There might be small effects, but fundamentally, what happens in financial terms is not the core of what the business is about. Investors should be thinking, first and foremost: is this a great business? Number one, are they creating customers? Number two, keeping them happy? Number three, is management smart enough not to screw up something good? Are the employees and the people who work in the business motivated and incentivized properly? That would be my hierarchy. Only after that do you ask: now that we've been wildly successful, are we being super efficient with our capital — with taxes, with handing it off to shareholders in the most efficient manner, or with whether we should be reinvesting, or be custodians of it and make it grow through other means.

It's as dangerous to stupidly spend it on pointless adventures or acquisitions, when the owner can just as well buy that business by taking the cash out and buying equity in it themselves. It's the same thing. So why bring it in and overpay? There's a lot you can screw up. These are the priorities as I see them, and managing all that capital is, if anything, maybe the last piece of this short list. We'll revisit this in great detail as anything happens, but for me it's almost a hobby.

Farshad Nayeri: Okay. I think Fred Stein had a question on the topic of the enterprise, and there was also a question from Frederick Hirsch, who couldn't attend. And I have my own addition on the enterprise. Fred, please go ahead — it's your question, and I'll add Frederick's first.

Fred Stein: Sure, thanks. And by the way, Horace, thanks so much for your opening remarks about what's really going on, and how underappreciated it is by the market and analysts. That's worth the price of admission. My question is about MLX. They didn't even mention it during the keynote — maybe it's too in-the-weeds for a general audience — but it seems they've done something almost like a reverse Trojan horse. You write in MLX, you can run on Apple Silicon, and then migrate to the other guy's chip, the Nvidia chip. Most of the time people do it the other way: they make migration tools to go from the other guy's chip to your chip. I thought this was fascinating and wanted to hear more.

Horace Dediu: I'm afraid I'm not an expert on that. But generally speaking, when you look at Apple's history with portability — and I think this speaks to that — they are very, very good at it. Months ago, when we had the Apple-50 celebration, I was watching a couple of videos of the old hands. The Computer History Museum — go to YouTube and you'll see interviews lasting two or three hours with some of these old-timers. I remember watching Avie Tevanian, Mike Markkula, Bertrand Serlet. I think it was Bertrand's interview that described the transition from — well, actually, PowerPC to Intel. This is way before the M series; it was in the early 2000s. When that shipped, people were shocked. How did you get OS X moved to both architectures? Nobody could imagine Windows doing this. Back then it was all about Intel. The idea that they moved lock, stock, and barrel, seamlessly, and even had sandboxes for runtimes that weren't going to run native —

I actually met Bertrand at WWDC. He gave the State of the Union talk after the keynote and talked about this — I think he called the old system a skyscraper, and how they had to move the skyscraper. A beautiful talk. He sat next to me afterward, and I said, "That's really impressive, how you did this." And he was very much, "This is nothing. We've been sweating portability for decades, because of my background and where I came from." He'd been on different OSes back in the day of the Mach kernel, trying to be clever about the portability of Unix, which was the heart of OS X, dealing with that portability issue way back, before he even joined NeXT — where he did a lot of heavy lifting — and then into Apple.

So you'd have to go back decades to see how entrenched this is in their psyche, in their DNA, to avoid becoming dependent on an architecture they don't control. Making stuff portable, being clever about the migration path — that's why they seemingly do this very easily, because it was really heavy lifting back in the day when portability was almost nonexistent and required a huge amount of work. If you remember Y2K, just fixing those things embedded in so many systems made people think we'd have a civilizational crisis. To think that the software engineering at Apple is so strong — the modularity, the attention to portability — it's not surprising they're doing this now at the level of chipsets for AI compute. I'm not surprised, even though I haven't dug into it.

Fred Stein: Great. Maybe one follow-up observation: they also mentioned Device Hub, which basically allows apps to run on any Apple device. Just an observation — it's impossible for any other vendor to do that.

Horace Dediu: You want to hear a funny thing? I don't know where this came up, but I saw it in one of my socials. Somebody was running macOS using an iPad as a display, and they said you can now use touch on the iPad to operate the user-interface elements of macOS. They said this just came out — correct me if I'm wrong, if it was already there, but I'd never heard of it. I sometimes use an extra screen — what's it called? It used to be called Sidekick —

Speaker 6: Is it Sidecar?

Horace Dediu: Sidecar — that's it. So this idea of getting an extra screen — but I never thought I'd control the Mac user interface from it, because it's not known to be touch-friendly; the elements are so small. But it's possible now. Imagine how much engineering went into making the whole user interface — the ability to click, select, right-click — already touch-ready, so that whenever, and likely soon, we have a Mac with a touch UI, it's already there. They probably built it years ago. It's just a matter of turning it on.

Farshad Nayeri: They're very purposeful about portability — as opposed to Microsoft. I think part of the problems Microsoft has, even to this day, is because they always brought the customers along, no matter what.

Horace Dediu: It just wasn't a priority for them. The hunger Apple has is because they went through multiple transitions. They probably bought Unix partly because — I suppose it was Avie who sold it to Jobs — he said, what we get is portability, scale up and down (we can make it a server, we can make it a device), and security, because it was a protected kernel, protected processes. All those came with Unix. When Mach came along, they got a lot of that with it. Back in the day, Microsoft never prioritized any of that, so their OS was a kludge, and it just got patched and patched and patched. That's history, but it's one of the stories that isn't told, because Apple didn't want to highlight what the engineering was all about back then. The engineering was top-notch, excellent.

Farshad Nayeri: Right. Fred had also posted a question, which is the one I intended to ask, so I'll read it now. There are three parts. First, from Fred: the Mac continues to grow rapidly in the enterprise — it'd be nice, Horace, if you could do a graph, real-time or later. Second, a question from Frederick: I'm confused about Apple being a consumer company, yet venturing into the enterprise with Vision Pro. And third, my own: I didn't hear the word "enterprise" once during WWDC this time — maybe I missed it, but it was developers and consumers.

Horace Dediu: I don't think I ever heard them talk about enterprise. The only time I hear about it is during the earnings calls, where they give a little example — Delta Airlines was one, some company in Europe, an insurance company — somebody switching to Macs or Apple devices across the board, and how they're getting wins in these large Fortune 500 companies. But fundamentally, from a messaging point of view, the two are opposite ends. Marketing to enterprise puts consumers off, and consumer marketing puts enterprises off. The idea that they're only getting consumer-grade technology is not attractive to many a CTO or CIO — though that may be changing. And consumers find enterprises to be jargon-filled and without any heart, any beauty or sensitivity. Besides that — you can do both, but with separate events for each. I think they're conflicted on this. It's great that they're getting wins.

By the way, the Vision Pro right now is getting traction with a lot of enterprises, especially in healthcare. They're careful not to pigeonhole it, because they don't want people to assume it's only for that. There's no real publicity; the direct salespeople out there talking to enterprise have their list of wins, but it's not something Apple talks about publicly. So I'm not surprised they didn't mention it at WWDC. It's opportunistic if they do. Maybe there are non-keynote sessions where it comes up.

And let's not forget, they spent almost fifteen minutes — maybe I'm exaggerating — on child safety. Ben Evans made a note of this. I saw him comment that it's uncharacteristic of them to talk about safety in such a lengthy, laborious, and dull way. But I think it's a political move. It's there to signal to authorities, who have probably been pressuring Apple in all kinds of ways — as usual, to do the wrong thing: to have a good ambition or intention with the wrong approach. That's very typical of regulators, because they're not engineers. Engineers work with constraints; regulators work with outcomes, and they don't think about how to get there. So my observation, like Ben's, is that this was very carefully calculated. They felt it was really important to dive into it and expose it at this venue, which was one of three tent poles.

By the way, remember, no hardware was launched. How many stories before this event were like, "We're going to see this device, that device"? No. It was architecture, then safety, then AI. Actually, the first thing was a general update of the OS — there was something about Liquid Glass, a few things like that. Normally, in the past, they'd do a segment for each OS: here's what's coming in iOS, here's what's new in macOS, in iPadOS, and so on, segmented by OS. And maybe something special, like, "Hey, here's SwiftUI," or something new with HealthKit, or Metal — the old developer stuff, a lot of software. This one was very different. A lot of architectural tidbits, very light on the new features in the OS, and suddenly this big lump of safety in the middle. So you have to figure out what's going on.

Farshad Nayeri: Well, first, a lot of developers have children, and developers are at the forefront of explaining things to the public. I also thought the format had to change a little, because they had these live-recorded sessions. The pacing seemed odd to me — they'd go from their usual rhythm to these places where people were typing things, and they had to live-record them so they didn't look fake.

Horace Dediu: That's why the tempo is a bit off. But a lot of people complimented that, those in the know. Even Neil Cybart, even Gurman, said the way it came across was that this was real, not contrived, and as a result it had more sincerity. But, as somebody pointed out, the editors of this video would have been pulling their hair out — it was so bad from an editing standpoint. An editor wants to keep things snappy; they know exactly how long you should hold a shot, and nothing should last more than a second. If there's no change in one second, you're doing it wrong. Here we were, dragging with five, six, ten seconds per shot. It was dreadfully bad as a movie. But again, they said: we need to be sincere about this, we need to demonstrate that this is real, to fight against the perception that we haven't got it done yet. Which is why I'm irritated by Gene Munster's comment that this looks to be vapor, or unfinished. It's being tested right now by the hands of millions.

Farshad Nayeri: Yeah. Greg, did you want to bring up your comment?

Greg: Sure, can you hear me? First, thanks for doing this — it was great. My comment: my sense of the safety-and-kids part was partly about championing the walled garden and the structure it provides. Apple can do things other companies can't, and even if those other companies wanted to make the same safety choices, their business plans are predicated on using people's data and selling it. So it seemed like the segment was targeted as much at Facebook as it was at politics.

Horace Dediu: Absolutely — this is super important. But the word I'd use is hygiene. I forget that I used it online, but here's the term: I remember this from when I was at Nokia. There are features considered hygiene, meaning if you have them, most people don't care, but if you don't have them, everybody complains, because it's unusable — how could you not? It's also a bit like plumbing. Plumbing is only noticed when it doesn't work; when it works, nobody thinks or cares about it. For that reason, safety — even automotive safety — is never a marketable thing. Volvo, maybe, but they're a super niche. Mostly, people don't want to hear about negative things. We had discussions, in public, watched by millions, about child abuse. We're talking about unpleasant things, things everybody wants to avoid. So from a marketer's point of view — as I said, the editors were cringing about how Siri AI was portrayed, but a marketer at Apple would have been suicidal about how that segment could have been used to create an enormous amount of euphoria on behalf of the consumer.

Of course, it wasn't a consumer audience primarily — it was about developers, but my point is it actually targeted the policymakers. This was designed to induce a sense of: we are custodians doing our jobs really, really well, and we're making this very public. It's in your face; you can't ignore it. And: we're burning our cycles and our goodwill in front of millions of people to tell you that we're doing our jobs. That's a weird thing to do. There must be something behind the scenes going on there.

Farshad Nayeri: Okay, next question. This was a little bit addressed already, but in case folks came late: will agentic AI replace the vibrant app economy? Would that be a consolidation of revenue to Apple and to the agent providers? Or will AI enable a new set of apps? That was from Dave Emery.

Horace Dediu: This is in the post that's in the pipeline, but I'm happy to expand, because you never get the full story when you try to write it down. The premise of intent as an operating system — designing in intent compute, as opposed to touch UI, touch compute — is that you assemble the solution to a complex problem using a multitude of modules. Those modules include models, and they include data — private data, awareness of the context of the user — and multiple types of inputs, which aren't just single taps or swipes; it's pictures, it's audio. So you have a much more complex set of inputs and components that you're assembling. The orchestrator assembles the solution and decides whether to farm out, or hire, a solution or module provided by an app.

In the post, I say that most people who assume intent computing assume there's a post-app future. But Apple doesn't. On the outer layer, they put either Siri or apps to create the full experience. By "experience," I mean it might even assemble a little mini-app — a set of questions you need to answer in order to get to the final thing you need — and then it assembles a solution to the job being asked, the intention being asked.

My further observation is that within apps lie a lot of the permissions we entrust. You have a banking app with your banking credentials; you have various other things you protect inside apps. The reason it's done that way is that you trust the institution behind it, and vice versa. These places where you've entrusted something have earned your trust. The Claude theory goes: just harvest all of these permissions — find out what the bank access code is, the social-security number, the mother's maiden name — give me all the access as an orchestrator, and I'll go do all those things that apps, or websites, were hired to do. I'll become your agent to execute all of these.

Now Apple says no. What they're hoping to do, because they're so concerned about trust, is to not have those things that have earned your trust be disintermediated. We're not going to fire all the things you've already hired because of the trust you've given them. That relationship is very delicate. Apple doesn't want to betray that trust, especially if things go wrong. So if you get hacked, it's: well, it wasn't my OS that lost your keys — be careful you don't get hacked in your Bank of America account, but that's on Bank of America to preserve. That's why they're going to eventually hand off to the institution with the credentials necessary to assemble the final solution, rather than try to take over completely.

So what apps used to do is provide interfaces, plus data, plus various other things you've established as a relationship with the provider. What they're possibly taking away from apps is the interface — and that was the costliest part of building an app anyway. Millions of developers are recreating the same basic thing: "Okay, we need a drop-down menu, we need a bunch of steps to authenticate," and so on. So what Apple might do is allow that interface to dissolve a little, or maybe just pop you into it. I think they're going to try to move away from it. That's the intent. Yeah, go ahead, sorry.

Farshad Nayeri: No, I'm sorry — it's like a reverse hub strategy. Instead of the Claude model, where Claude is the hub doing everything, Apple is distributing it to each of the apps. And in a sense, they're all sandboxed already, so that gives you a layer of protection. And now that everything's running on Swift — something they brought up is that they've open-sourced these tools, and you can run their AI stuff on Linux — eventually they could say, "Oh, this workflow that used to go through the app, you click a button and we'll run it for you on our Private Cloud Compute."

Horace Dediu: So once again, they're bootstrapping to some later model of this — guide rails and scaffolding to make stuff happen behind the scenes. The thing about Claude, which was super, super exciting, was that it did all this stuff the orchestrator is supposed to do. When I first heard about it, I thought, man, this would be great if it were an OS-level thing inside iOS. I thought iOS, or macOS, should be that thing you currently have to cobble together — why not have that be super, super secure? But what you see appearing now is, yeah, we're going to do that, but we're not going to take on the liability, as Claude did, to lose all your stuff and be compromised by attacks. Those were notoriously bad at security. The warning I always gave people was: you can experiment with that, but be very careful to keep anything sensitive and private out of it, and keep anything with money out of it. But people love to play with it.

So for Apple to do this properly, they understand inherently that it's about managing liability and trust issues. It's not just, "Oh, I'll take it from here, thank you, we're going to destroy the app economy." Yes, there's money in the app economy, but there's much more than that.

Farshad Nayeri: And Horace, one other addition to what you said about apps — that was a really wonderful explanation. You were talking about what developers spend a lot of time on. As a developer, another thing you spend a lot of time on is integration into content, and loading all the sample data. You know how Apple stores always have all the beautiful pictures. And not just that — you want end users to be able to do those things efficiently: importing all their photos and so on. I just watched a developer video about creating an app very similar to Duolingo — this tutorial is about making an app to learn Chinese, which is near and dear to your heart. The developer uses AI models running locally to speak the Chinese words; she's taking pictures of things and saying, "This is a piece of wood," and the model figures out the description in Chinese. Then later she says, "Okay, now I'm on the Mac, I want to load my entire photo album from this trip" — a completely different workflow from what you do on the iPhone. As a developer, you have to cover all these workflows if you want your apps to be used. Now they're a few minutes of work instead of a few days. That's very exciting. I'll send a link — it's a little technical.

Horace Dediu: Yeah, app development is going to get a little easier, a little different. I use this phrase, modularity, a lot, and the idea is that apps become more like modules that can be assembled into something else. Disruption theory suggests that when a value network consolidates, or integrates, or modularizes, there's a transition in the point of profit capture — what Clay calls evaporation and condensation at different points in the value chain. As the value chain evolves, there's this transition or transfer between modules of value capture.

As I said, the Apple stack — the historical compute stack, from silicon to applications — was such a value network. What Apple did is control the interface between the operating system and the hardware, and between the operating system and the apps. Microsoft didn't get into hardware; they got a few core apps, actually got more into apps. So there's a difference between the two stacks when we look at PC versus Mac. The Mac did very well in profitability because of that integration, but not in scale. With phones, they did well in scale as well.

But the point I was making — and this is in the article — is that if the module at the top of the stack is the app, but now we're starting our journey of the interface through Siri, into the middle, into the orchestrator, and then it goes to the app and pulls from it, suddenly the app is a callable API. That seems like a demotion — like it's less important. But that's why I put the caveat: they're going to hold on to critical data, to critical elements. So they might even deliver the ultimate experience — they'll get stuff populated and then still deliver it, so the app opens and you go to the right place inside it and still interact with the app. Apple is going to invoke the app, but keep it there, because they don't want to take over all of it. Nor could they, because there are millions of them. I think they'll offer something so that app developers are incentivized to continue developing and innovating.

And the old way of doing things — invoke the app and do everything outside of intent computing — that's going to live on, no question. In fact, it can probably live on fifty years. What I concluded with this dichotomy question is: no, we're going to have both. We're going to have intent, and touch, and vision, and all these things. It's like a tree that branches off into many different branches — like evolutionary trees. Just because something evolves into something else and forks, doesn't mean it has to go extinct. It just goes on in parallel. That's why there are still apes and humans. Anyway — stay tuned, it's coming out.

Farshad Nayeri: All right, thank you for that. This is a question from Salvatore Saiva — I hope I got that right. Apple felt it was important to own their map technology. Is it similarly important for them to develop IP for foundation models and so on? Well, I think they are.

Horace Dediu: This is the thing. Again, just reading the commentary afterward from people who were there listening to the deep dives — there are four or five foundation models that Apple is going to deploy, or make available, whatever the word is. And only one has anything to do with Google.

Farshad Nayeri: Google technology — I don't know precisely. I think they use Google the way they used Mach. They started with it. Mach was open source, but their next step wasn't open source; they started with it and built on top of it whatever they needed.

Horace Dediu: They took just something from Google, because obviously they'd have wanted to do it all themselves, but they wanted something from Google. And I think they built a lot of stuff: there's on-device, there are world models, there's stuff in the cloud, there's Private Cloud Compute — and they're controlling all of it. The question is: can they swap out what they use Google for today? Or is their relationship indefinite, hard to break? Or can they actually replicate the whole thing and eventually do it themselves? So: can they swap, or can they do it themselves? That's the question of how important Google is to this story.

I think they're keeping their options open — it's always a relationship. When people said, "Oh, they've thrown in the towel, it's all Google now," I said: the way they deal with suppliers is as a very tough negotiator. This is legend. I don't see them treating a software supplier, or a technology supplier, or even a cloud-service supplier, any differently than they treat their hardware-component suppliers. They're a tough negotiator. They probably actually pissed off OpenAI, which wanted to sue, because OpenAI wanted to be in that position. Apple effectively derailed them: they said, "We're going to bring you in," way back in 2024, and OpenAI somehow had reason to believe they'd get huge distribution — that with Siri you'd actually do ChatGPT. Not much came of it, because they dreamt of riches but nothing came, and then they got kicked to the curb and replaced by Gemini, their nemesis. That's very typical, though. Maybe Google will be kicked out too. I think even Microsoft was in the running at some point. Everybody wanted to be in that position of being the supplier of some models to Apple.

Farshad Nayeri: Yeah. But for OpenAI, it feels like the ChatGPT integration was more like the ROKR situation —

Horace Dediu: Yeah, they got used. But from Apple's point of view, it's not malicious. It's: let's try it; doesn't work; we'll move on. People interpret it differently, obviously — they feel they've been betrayed somehow, or deceived.

Farshad Nayeri: Yep. Okay, moving on. Related to this, there was a question: evidently Apple was number 12 on the Wall Street Journal's "best companies for the future," and Nvidia was at the top, of course. I also saw some list about innovation, and it was funny — I guess that's confusion about the word "innovation." What's going on with Apple's brand image here?

Horace Dediu: First of all, Nvidia. If you look at their history — Farshad, you collected this — there used to be these graphs showing the revenues, the FAANG group, gross margins, profit, and so on, comparing Apple versus the usual suspects. The new suspect is Nvidia. But if you look at their history beyond the last few years, it was literally flat, and then the last few years it's not. Usually you think of growth as a nice hockey stick — which doesn't really look like a hockey stick to me, but it's a sort of exponential, a beautiful curve. Nvidia's is not that. It's a line, and then another line at a very steep angle.

Here we go — yeah, there it is. I'll reduce the screen so you can see. Compared to everyone else. Now, Tesla doesn't belong here, but you have to include it, because it's one of the seven. They haven't had growth for years, they're tiny, and they have no profits. But look at the others — beautiful, right? Meta looks nice, somewhat linear but with a bit of a curve. Microsoft looks brilliant. Google looks brilliant. Apple, of course, because it's a consumer company, has seasonality, so it's very spiky — and so is Amazon, because of the holidays. Amazon is also a low-margin business, because sales are huge but retail sucks as a margin business. But then look at Nvidia. It's zero to one. It's not even a curve. It is a slope.

Farshad Nayeri: Yeah, it's one of those overnight successes thirty years in the making, right? They started around 2008.

Horace Dediu: They were very lucky, because the type of chip they'd been making for a long time turned out to be very useful — first for crypto, which got them that first little bump; you see that wobble. They were a graphics-chip company, and that was extremely commoditized, because even major players decided to put graphics on the main system-on-a-chip; you didn't want a separate chipset. It was a messy way of doing graphics — it was an accelerator. Remember, back in the day, instead of the CPU doing the heavy lifting on graphics, you'd use a graphics accelerator, and that's been their legacy. But it turns out that to do graphics fast, you needed the same kind of circuitry — a digital signal processing, Fourier-transform type of thing — that does one thing, just one thing, scaled up. That turned out to be super useful for graphics, then for crypto, and then suddenly super useful for AI, because it's doing nothing more than giant matrix multiplications, matrix linear algebra. So they suddenly won the lottery.

Now, they're smart people — I don't want to take anything away. They took the advantage and weren't stupid about it. They're developing ecosystems and platforms on top of it. All that is true; I'm not against Nvidia in any way. But to suggest, as the Wall Street Journal does, that a lottery winner — even a smart one — is the best company of the future because they won the lottery, doesn't say much about their sustainability, their processes, their priorities, or their market scale. Because, again: how many customers does Nvidia have? Five, maybe, because they're the big hyperscalers. And how many chips does it sell? As we discussed when Ben Bajarin was on the call, tens of millions, maybe — each costing tens of thousands. So that's a big number, very profitable. But tens of millions, versus Apple shipping probably 400 million A-series, plus now more and more other series of chips. The amount of silicon Apple is going to output and put into people's hands is orders of magnitude higher — not a single order, multiple orders of magnitude.

So if Apple's roadmap includes the same functionality Nvidia provides — these types of computations, which, by the way, they already had on the chip for machine learning — and now they move toward exactly what AI needs, baking that into their chipset going forward for decades, where's Nvidia? It's great that they've done a great job with what they were given. But maybe they'll be in the PC business. Maybe they'll be the next Intel. How many customers do they have, at the end of the day? They're trying to wiggle their way, after winning the lottery, into other things: "Hey, I won the lottery, I'm going to get into real estate. I'm going to buy a liquor store. I'm going to become a golf pro." That's great, you're going to diversify — but it's a different story about distribution, market creation, customer satisfaction, dealing with end users.

So the premise that this is a company with an unassailable advantage — that's how you should judge a company's potential, right? What Buffett calls the moat, what the business-theory people call sustainable competitive advantage. It's a mouthful: sustainable competitive advantage. Meaning you have an advantage, and it stays — somebody doesn't leapfrog and copy you. I'm not convinced Nvidia is that kind of company. IBM, in its heyday, had far more diversity in its business opportunities and business logic, even though it wasn't in consumer — well, maybe the IBM PC was their first, or the PCjr, God forbid. The great thing about those graphs is they give you a perspective of about ten or twenty years. And yeah, Nvidia is, to me, a decent, very good company — but a very lucky one. Comparable, I'd say, to Qualcomm. Qualcomm is a great company too, but building communications chips is not a foundation for becoming the world's greatest company forever.

Farshad Nayeri: Actually, in one of the talks, Jensen brought up Qualcomm as something similar — they're very similar companies.

Horace Dediu: The narrowness is what bothers me most. And every one of those five customers it has is dying to get rid of them. They're looking every which way to fire their supplier.

Farshad Nayeri: Right. Although, lately, I think they've become — we'd need Ben and Jay here to explain — more of a Cisco. We used to think of them as an Intel for GPUs; now they're kind of a Cisco for networks, helping the build-up.

Horace Dediu: All that's consuming them right now is: how do we keep this ball rolling? And how do we build contingencies for the fact that everyone is painting a target on our back? That's the question of sustainable advantage.

Farshad Nayeri: Yeah, related to that, Horace — there are a lot of comparisons between Nvidia and the late-dot-com stuff about dark fiber, building capacity. But these are GPUs and machines that have to be amortized — in five years they'll have to be replaced — whereas dark fiber is going to stick in the ground and stay there forever. So I don't understand that comparison at all.

Horace Dediu: It's true, and that's one of the hidden liabilities of data-center, compute capital. In the old days, capital was defined as machinery used to manufacture things, and the idea of a machine used to make things was that it compounded value. At the heart of capitalism is this notion that you can compound value — it's a magical thing for wealth-building, and that's what created the tension with labor, because labor can't do that. Labor is a linear process: more bodies, more people, increase output linearly, whereas machines can grow exponentially. That created the tension between capital and labor that fueled the industrial revolution and every revolution thereafter, socially speaking.

But the machines we're dealing with now — these new pieces of capital, these new looms and mills, replacing the old lathes and tooling for manufacturing — these new machines are far more fragile. In the old days you'd say, "These will run for a couple of years until they wear out," because they had consumables, parts to replace, and then obsolescence: a better machine improves productivity significantly, so we'll obsolete it. But those were measured in many years. The challenge with new compute capital is that its lifespan is unpredictable. It also has externalities — cooling costs, energy costs — and we're not sure what the impact of those externalities is.

With industrialization, we had pollution. One of the unaccounted-for, invisible liabilities of those old machines was that pollution was generated, and societal tension — two things factory planners had no idea about. Now we have data-center planners suddenly facing a lot of popular opposition, again an unrealized liability, plus energy, water, and other consumables that are probably also finite resources. The economics of all that — even for capital back in the industrial revolution — took decades, if not a century, to see the implications. We ran through the 19th century trying to figure it out. Now we're trying to figure this out in three years.

Farshad Nayeri: We don't have to go that far back. From the dot-com to the dot-bomb — I put up a picture of the leading hosting companies. Exodus Communications: their peak capitalization was $32 billion, and they sold for $750 million, which was their initial capitalization. Above.net was $1.2 billion at peak — actually four and a half billion at peak — and they went to zero. I invested in these companies and got out at a pretty good time. These were all companies that invested in this kind of infrastructure. We think of the fiber, but when you see somebody like Oracle doing that now, it just feels like this is where they're going, possibly.

Horace Dediu: Yeah, there's a lot going on.

Farshad Nayeri: Okay. We have a couple more questions, and we're about half an hour beyond time, so let's wrap up. You recently talked about Apple's CAGR of an impressive 26.6% over the last twenty years. Can you talk about your forward idea of Apple — without, obviously, giving stock advice, which we don't do?

Horace Dediu: So, forward CAGR. Well, first of all, the point of my article was that more than half of analysts had a hold, a sell, or no opinion on Apple, and I found that surprising. By chance, I went through the process of trying to aggregate all the known ratings and targets. I know Philip Elmer-DeWitt does a great job with that, but I wanted my own copy of the current opinion out there. You see an upgrade or a downgrade once in a while; I just like to see them all together. At first I had a collection of 20, then it grew, and I found a site that aggregates them and suggested it was more than 60. But when I looked them all up, I found about 10 had actually stopped covering — leftovers in the list. So it settled to about 50 actively covering. Of those, the split was about half yes, half no.

I thought: wait a minute. If you have an opinion of "hold" — and I'm not giving any recommendation — that suggests the stock is fairly priced. That's okay. But it also suggests that if the company is growing at 20% a year, it's probably going to double its revenues in about three to four years. Rule of 72: divide 72 by the growth number, and that's how many years it takes to double. At 25%, you're talking three years or so. So if that's the growth rate Apple would have in the future, then why would you not buy? That's a funny thing to suggest.

So I thought: what is the historic return? I'd never calculated this before. I calculated the CAGR. The trouble is, it depends on when you buy. So I calculated it from every year — assuming you buy on, say, January 1st, then compute from that year to the present — and did that for about twenty years. The curves are very volatile at the beginning, then they settle around 20–25%. You could do it for every day, every millisecond; it doesn't really matter. My experiment showed they all aggregate. Now, with a drop of 10–20%, maybe that 26% becomes like 22% — there's huge volatility here. But the point, back to these analysts, is: why would they do that? Then I realized they're making recommendations for one year ahead, very short-term, whereas all these CAGR calculations are multi-year returns. You're going to see a lot of volatility short-term, but a lot of steady growth long-term.

I became a little disappointed that nobody is doing this on a more-than-one-year basis, because it would make a lot of sense to have, say, 20 analyst ratings of a five-year CAGR for Apple. So the question was, essentially: what's my expectation for the five-year CAGR? My answer is, I'm only dumb enough to assume that twenty years of history can be projected forward. Now, I know all about the fact that history does not represent the future. But we could have had this same conversation every single day for the last fifteen years, and we'd have seen even faster growth, actually, in the first five years of those twenty — and still, it just kept growing. So it's actually extraordinary to assume that this massive flywheel, with all this momentum, just keeps spinning — and then somehow is not going to continue. That's much harder to argue than to say, well, it's continuing.

It's like the old story: what's the stock price tomorrow? The most probable price of any stock tomorrow is exactly what it is today — random-walk theory. It could go up or down with 50-50 probability, so the most probable thing is to be where it is now. So with a consistent growth rate, knowing quite a lot about the business and what's in the pipeline, you'd say, "I'm just going to be dumb and guess." By the way, you can do this with GDP. Look at the history of country GDPs, world GDP, city GDPs — they're like 2–3% compounding for 200 years. And what's the future going to look like? Most people say catastrophe, end of the world — or hypergrowth, utopia. Dystopia, utopia. No — the most likely outcome is 2% compounding for the next 10, 15, 20, 50 years. It's a natural thing to extrapolate a long-term trend out one, two, three, four years. That's not an extraordinary claim. In fact, it's an extraordinary claim to suggest it doesn't continue.

So why are half the people making extraordinary claims about a business they follow? The only excuse is that they're looking one year out and completely swamp their thinking with near-term, non-fundamental analysis. They're saying there's a lot of crap going on and it's all going to go wrong in the next few months. Otherwise — and this is the other point I made — what stands up to scrutiny? That they're going to lose the core business of being the user interface for most devices in the world? That customer satisfaction is going to evaporate, that they'll get swapped out by this or that? All the usual suspects, all the threats we've heard a hundred times before. I've heard them before. They never happened. So why should I believe you now?

Now, had we done this in 2008 or '09, it was a completely different story, because the history wasn't the iPhone — it was the iPod, or the Mac. It was a lot more volatile, but growth was far faster. The iPhone came out of the gate with triple-digit growth. It was a very dynamic period, and that probably conditioned me not to think about CAGR in the long term, because things were just very dynamic. Now we're in a period where things are far more predictable, simply because there's so much more stability, consistency, scale, size — what they call the flywheel. Yes, there are problems with that, and there's less dynamism. But you're an analyst, and you're going to tell me it's all coming to an end tomorrow? I don't know why you wouldn't just make basic assumptions. Does that make sense?

Farshad Nayeri: Yeah. Oh, we have Philip joining at the last minute — he missed the shout-out. Hi, Philip, wherever you are.

Horace Dediu: We have this recorded, right? So we're good.

Farshad Nayeri: Yes, we can release this. So, one last question, from Lalit: does Apple TV — some future version of it — morph into a data center for the family, or for a small business, but definitely the family? And they did talk about Home a lot, I thought, with AI.

Horace Dediu: I haven't written a post about Home yet, because I already had three others in mind first. But the Home thing was something I'd anticipated for years, and I finally thought maybe this is the year. Home needed a big upgrade in terms of its intelligence, its reach and power. I felt that had been idling — literally just spinning its engine — and what it needed was AI. Because it's all configuration. Anybody who tries to set up their smart home thinks, "Oh my God, I have to be a programmer." It feels like the old days: what are events? What are triggers? What are scenes? You get all this language about setting everything up, and then it doesn't work. I don't want to tell it every single thing it needs to do, every trigger. I need it to be far more self-aware of my needs. So the missing ingredient was some sort of inference, some intelligence.

Where it got interesting was the rumors of a new device type — some sort of screen-based device. Think of it as a HomePod with a screen. The old HomePod had a tiny screen, but this would be a display that follows you, creates eye contact. I saw a startup do this once — pretty clever. The idea is something more anthropomorphic, more like the Pixar lamp — articulating, moving, something more like a pet.

Now, Apple TV is currently the hub, and that's by convenience: first, it has a display that happens to be a TV; second, it has enough horsepower to act as a gateway, and it does the integration work effectively, with the necessary security. The question is what's going on with that — will it remain the hub? The answer might be that there's a new hub coming, another device. Or maybe we get this idea of distributing between the two: some people may choose to enable their Apple TV.

The challenge is this: with AI, you're going to need to upgrade more often, because the hardware will probably get obsoleted faster. One thing I notice about my own use of Apple TV is that I have stuff that's really old, and it doesn't compel me to upgrade, because I don't change my TV often. I actually have a TV from 2008. I'm embarrassed to say it's so obsolete that in Finland, where I am now, the TV signal — it's digital — is no longer processed by this TV. We have to dispose of it, because it's non-functional as a TV. But it still displays my HD output. There are very few reasons for you to replace it. Unless you upgrade to a bigger screen, very few people will need to. It's an appliance — it's like a refrigerator, which doesn't get replaced unless it's broken. So the thing attached to it, the Apple TV, also tends to wither. You're not rushing around going, "Honey, can we get a new Apple TV this year? It's super cool." Hardly ever. It's an appliance.

For that reason, I'm a little worried that whatever they build has to have some reason to be a software object. Right now, their appliances tend to feel more like hardware that doesn't dynamically evolve. I used to say, years ago, that the problem with smart homes is that they're the most expensive things we own. The second is the car. The most expensive thing is the house — and it's the least intelligent thing. We put much more intelligence in our wristwatch than in our homes. A wristwatch is $300; a home is possibly $3 million, or $300,000 at least. Somewhere between those two there's no brainpower at all — no software, no network, nothing. It's been like that for decades, as long as we've had computers; we just never bothered. Cars are the second-worst offender — stupid, unintelligent by every metric — and yet they resist the injection of software, for many reasons. Many people actually don't want them to be that smart. They're holdouts. Those who try feel quixotic; they're tilting at windmills, fighting against the tide trying to bring intelligence into the home, and users feel betrayed every time, because it's hard to use.

So the struggle continues. Home and health are huge problems to be solved, and I think they're going to take another run at it. With AI, it's a huge opportunity to reinvent these holdouts — or reinvent the attacks on these holdouts that are still unintelligent and unpenetrated by software. So, to the question about Apple TV specifically: I think the idea of a home hub is very likely. Not necessarily as a storage device, but as a secure gateway, and a place where you might interact more — especially if it gains an always-on interface. The problem with Apple TV is that the TV doesn't stay on all the time. But if you had this thing on a countertop or a shelf, and it greeted you every time you came home, it'd be better than a faceless speaker and a TV that's hired to do other things. "Hey, I'm your home assistant. I'm your home orchestrator." A home hub.

By the way, the dreams of the utopians with robots are an interesting observation — the idea of, let's say, Optimus, but there are hundreds of others, especially in China: these robotic assistants, humanoid robots. This completely baffles me, because I don't think people want that. Number one, they'd be freaked out by them; they're not comfortable objects to have. But more importantly — never mind the cost — I think we like to buy appliances. We like things in our homes that are single-purpose, very approachable, and simple. The idea of a speaker with a tiny screen that moves is far more attractive than a robot, which would feel scary and probably not do the job as well as that little thing could in managing your home. If you think labor is what you need, that's a different conversation — hiring labor for your home is a completely different psychological problem. But we could get into that next time.

Farshad Nayeri: Okay, thank you, Horace. I think we've exhausted all the questions, but it took us two hours. Thank you to everyone still here — we appreciate your support and your joining us, the Asymco One subscribers who fund this, and your guests. We'll be doing the usual transcript and recording, and you'll see this in other forms in a few days. Meanwhile, if you feel there are things we can improve, please reach out — everyone has my email, or just send a note to our contact address. With that, we'll close.

Speaker 6: Thank you. Bye-bye, everybody.

Sign up and latest updates at asymco.com/live.