Nvidia In the Valley

Nvidia investors have been in the valley before:

A drop in Nvidia's stock price

This chart, though, is not from the last two years, but rather from the beginning of 2017 to the beginning of 2019; here is 2017 to today:

Nvidia's current stock price drop

Three big things happened to Nvidia’s business over the last three years that drove the price to unprecedented heights:

  • The pandemic led to an explosion in PC buying generally and gaming cards specifically, as customers had both the need for new computers and a huge increase in discretional income with nowhere to spend it beyond better game experiences.
  • Machine learning applications, which were trained on Nvidia GPUs, exploded amongst the hyperscalers.
  • The crypto bubble led to skyrocketing demand for Nvidia chips to solve Ethereum proof-of-work equations to earn — i.e. mine — Ether.

Crypto isn’t so much a valley as it is a cliff: Ethereum successfully transitioned to a proof-of-stake model, rendering entire mining operations, built with thousands of Nvidia GPUs, worthless overnight; given that Bitcoin, the other major crypto network to use proof-of-work, is almost exclusively mined on custom-designed chips, all of those old GPUs are flooding the second-hand market. This is particularly bad timing for Nvidia given that the pandemic buying spree ended just as the company’s attempt to catch up on demand for its 3000-series of chips were coming to fruition. Needless to say, too much new inventory plus too much used inventory is terrible for a company’s financial results, particularly when you’re trying to clear the channel for a new series:

Nvidia's gaming revenue drop

Nvidia CEO Jensen Huang told me in a Stratechery Interview last week that the company didn’t see this coming:

I don’t think we could have seen it. I don’t think I would’ve done anything different, but what I did learn from previous examples is that when it finally happens to you, just take the hard medicine and get it behind you…We’ve had two bad quarters and two bad quarters in the context of a company, it’s frustrating for all the investors, it’s difficult on all the employees.

We’ve been here before at Nvidia.

We just have to deal with it and not be overly emotional about it, realize how it happened, keep the company as agile as possible. But when the facts presented itself, we just made cold, hard decisions. We took care of our partners, we took care of our channel, we took care of making sure that everybody had plenty of time. By delaying Ada, we made sure that everybody had plenty of time for and we repriced all the products such that even in the context of Ada, even if Ada were available, the products that after it’s been repriced is actually a really good value. I think we took care of as many things as we could, it resulted in two fairly horrific quarters. But I think in the grand scheme of things, we’ll come right back so I think that was probably the lessons from the past.

This may be a bit generous; analysts like Tae Kim and Doug O’Laughlin forecast the stock plunge earlier this year, although that was probably already too late to avoid this perfect storm of slowing PC sales and Ethereum’s transition, given that Nvidia ordered all of those extra 3000-series GPUs in the middle of the pandemic (Huang also cited the increasing lead times for chips as a big reason why Nvidia got this so wrong).

What is more concerning for Nvidia, though, is that while its inventory and Ethereum issues are the biggest drivers of its “fairly horrific quarters”, that is not the only valley its gaming business is navigating. I’m reminded of John Bunyan’s Pilgrim’s Progress:

Now Christian had not gone far in this Valley of Humiliation before he was severely tested, for he noticed a very foul fiend coming over the field to meet him; his name was Apollyon [Destroyer].

Call Apollyon inventory issues; Christian defeated him, as Nvidia eventually will.

Now at the end of this valley there was another called the Valley of the Shadow of Death; and it was necessary for Christian to pass through it because the way to the Celestial City was in that direction. Now this valley was a very solitary and lonely place. The prophet Jeremiah describes it as, “A wilderness, a land of deserts and of pits, a land of drought and of the shadow of death, a land that no man” (except a Christian) “passes through, and where no man dwells.”

What was striking about Nvidia’s GTC keynote last week was the extent to which this allegory seems to fit Nvidia’s ambitions: the company is setting off on what appears to be a fairly solitary journey to define the future of gaming, and it’s not clear when or if the rest of the industry will come along. Moreover, the company is pursuing a similarly audacious strategy in the data center and with its metaverse ambitions as well: in all three cases the company is pursuing heights even greater than those achieved over the last two years, but the path is surprisingly uncertain.

Gaming in the Valley: Ray-Tracing and AI

The presentation of 3D games has long depended on a series of hacks, particularly in terms of lighting. First, a game determines what you actually see (i.e. there is no use rendering an object that is occluded by another); then the correct texture is applied to the object (i.e. a tree, or grass, or whatever else you might imagine). Finally light is applied based on the position of a pre-determined light source, with a shadow map on top of that. The complete scene is then translated into individual pixels and rendered onto your 2D screen; this process is known as rasterization.

Ray tracing handles light completely differently: instead of starting with a pre-determined light source and applying light and shadow maps, ray tracing starts with your eye (or more precisely, the camera through which you are viewing the scene). It then traces the line of sight to every pixel on the screen, bounces it off that pixel (based on what type of object it represents), and continues following that ray until it either hits a light source (and thus computes the lighting) or discards it. This produces phenomenally realistic lighting, particularly in terms of reflections and shadows. Look closely at these images from PC Magazine:

Let’s see how ray tracing can visually improve a game. I took the following screenshot pairs in Square Enix’s Shadow of the Tomb Raider for PC, which supports ray-traced shadows on Nvidia GeForce RTX graphics cards. Specifically, look at the shadows on the ground.

An image with rasterized lighting
Rasterized shadows
Ray-traced shadows
Ray-traced shadows

[…]The ray-traced shadows are softer and more realistic compared with the harsher rasterized versions. Their darkness varies depending on how much light an object is blocking and even within the shadow itself, while rasterization seems to give every object a hard edge. The rasterized shadows still don’t look bad, but after playing the game with ray traced shadows, it’s tough to go back.

Nvidia first announced API support for ray tracing back in 2009; however, few if any games used it because it is so computationally expensive (ray-tracing is used in movie CGI; however, those scenes can be rendered over hours or even days; games have to be rendered in real-time). That is why Nvidia introduced dedicated ray tracing hardware in its GeForce 2000-series line of cards (which were thus christened “RTX”) which came out in 2018. AMD went a different direction, adding ray-tracing capabilities to its core shader units (which also handle rasterization); this is slower than Nvidia’s pure hardware solution, but it works, and, importantly, since AMD makes graphics cards for the PS5 and Xbox, it means that ray tracing support is now industry-wide. More and more games will support ray tracing going forward, although most applications are still fairly limited because of performance concerns.

Here’s the important thing about ray tracing, though: by virtue of calculating light dynamically, instead of via light and shadow maps, it is something developers can get “for free.” A game or 3D environment that depended completely on ray tracing should be easier and cheaper to develop; more importantly, it means that environments could change in dynamic ways that the developer never anticipated, all while having more realistic lighting than the most labored-over pre-drawn environment.

This is particularly compelling in two emerging contexts: the first is in simulation games like Minecraft. With ray tracing it will be increasingly realistic to have highly detailed 3D-worlds that are constructed on the fly and lit perfectly. Future games could go further: the keynote opened with a game called RacerX where every single part of the game was fully simulated, including objects; the same sort of calculations for light were used for in-game physics as well.

The second context is a future of AI-generated content I discussed in DALL-E, the Metaverse, and Zero Marginal Cost Content. All of those textures I noted above are currently drawn by hand; as graphical capabilities — largely driven by Nvidia — have increased, so has the cost of creating new games, thanks to the need to create high resolution assets. One can imagine a future where asset creation is fully automated and done on the fly, and then lit appropriately via ray tracing.

For now, though Nvidia is already using AI to render images: the company also announced version 3 of its Deep Learning Super Sampling (DLSS) technology, which predicts and pre-renders frames, meaning they don’t have to be computed at all (previous versions of DLSS predicted and pre-rendered individual pixels). Moreover, Nvidia is, as with ray-tracing, backing up DLSS with dedicated hardware to make it much more performant. These new approaches, matched with dedicated cores on Nvidia’s GPUs, make Nvidia very well-placed for an entirely new paradigm in not just gaming but immersive 3D experiences generally (like a metaverse).

Here’s the problem, though: all of that dedicated hardware comes at a cost. Nvidia’s new GPUs are big chips — the top-of-the-line AD102, sold as the RTX 4090, is a fully integrated system-on-a-chip that measures 608.4mm2 on TSMCs N4 process;1 the top-of-the-line Navi 31 chip in AMD’s upcoming RDNA 3 graphics line, in comparison, is a chiplet design with a 308mm2 graphics chip on TSMC’s N5 process,2 plus six 37.5mm2 memory chips on TSMC’s N6 process.3 In short, Nvidia’s chip is much larger (which means much more expensive), and it’s on a slightly more modern process (which likely costs more). Dylan Patel explains the implications at SemiAnalysis:

In short, AMD saves a lot on die costs by forgoing AI and ray tracing fixed function accelerators and moving to smaller dies with advanced packaging. The advanced packaging cost is up significantly with AMD’s RDNA 3 N31 and N32 GPUs, but the small fan-out RDL packages are still very cheap relative to wafer and yield costs. Ultimately, AMD’s increased packaging costs are dwarfed by the savings they get from disaggregating memory controllers/infinity cache, utilizing cheaper N6 instead of N5, and higher yields…Nvidia likely has a worse cost structure in traditional rasterization gaming performance for the first time in nearly a decade.

This is the valley that Nvidia is entering. Gamers were immediately up-in-arms after Nvidia’s keynote because of the 4000-series’ high prices, particularly when the fine print on Nvidia’s website revealed that one of the tier-two chips Nvidia announced was much more akin to a rebranded tier-3 chip, with the suspicion being that Nvidia was playing marketing games to obscure a major price increase. Nvidia’s cards may have the best performance, and are without question the best placed for a future of ray tracing and AI-generated content, but at the cost of being the best values for games as they are played today. Reaching the heights of purely simulated virtual worlds requires making it through a generation of charging for capabilities that most gamers don’t yet care about.

AI in the Valley: Systems, not Chips

One reason to be optimistic about Nvidia’s approach in gaming is that the company made a similar bet on the future when it invented shaders; I explained shaders after last year’s GTC in a Daily Update:

Nvidia first came to prominence with the Riva and TNT line of video cards that were hard-coded to accelerate 3D libraries like Microsoft’s Direct3D:

The GeForce line, though, was fully programmable via a type of computer program called a “shader” (I explained more about shaders in this Daily Update). This meant that a GeForce card could be improved even after it was manufactured, simply by programming new shaders (perhaps to support a new version of Direct3D, for example).

[…]More importantly, shaders didn’t necessarily need to render graphics; any sort of software — ideally programs with simple calculations that could be run in parallel — could be programmed as shaders; the trick was figuring out how to write them, which is where CUDA came in. I explained in 2020’s Nvidia’s Integration Dreams:

This increased level of abstraction meant the underlying graphics processing unit could be much simpler, which meant that a graphics chip could have many more of them. The most advanced versions of Nvidia’s just-announced GeForce RTX 30 Series, for example, has an incredible 10,496 cores.

This level of scalability makes sense for video cards because graphics processing is embarrassingly parallel: a screen can be divided up into an arbitrary number of sections, and each section computed individually, all at the same time. This means that performance scales horizontally, which is to say that every additional core increases performance. It turns out, though, that graphics are not the only embarrassingly parallel problem in computing…

This is why Nvidia transformed itself from a modular component maker to an integrated maker of hardware and software; the former were its video cards, and the latter was a platform called CUDA. The CUDA platform allows programmers to access the parallel processing power of Nvidia’s video cards via a wide number of languages, without needing to understand how to program graphics.

Now the Nvidia “stack” had three levels:

The important thing to understand about CUDA, though, is that it didn’t simply enable external programmers to write programs for Nvidia chips; it enabled Nvidia itself.

Much of this happened out of desperation; Huang explained in a Stratechery interview last spring that introducing shaders, which he saw as essential for the future, almost killed the company:

The disadvantage of programmability is that it’s less efficient. As I mentioned before, a fixed function thing is just more efficient. Anything that’s programmable, anything that could do more than one thing just by definition carries a burden that is not necessary for any particular one task, and so the question is “When do we do it?” Well, there was also an inspiration at the time that everything looks like OpenGL Flight Simulator. Everything was blurry textures and trilinear mipmapped, and there was no life to anything, and we felt that if you didn’t bring life to the medium and you didn’t allow the artist to be able to create different games and different genres and tell different stories, eventually the medium would cease to exist. We were driven by simultaneously this ambition of wanting to create a more programmable palette so that the game and the artist could do something great with it. At the same time, we also were driven to not go out of business someday because it would be commoditized. So somewhere in that kind of soup, we created programmable shaders, so I think the motivation to do it was very clear. The punishment afterwards was what we didn’t expect.

What was that?

Well, the punishment is all of a sudden, all the things that we expected about programmability and the overhead of unnecessary functionality because the current games don’t need it, you created something for the future, which means that the current applications don’t benefit. Until you have new applications, your chip is just too expensive and the market is competitive.

Nvidia survived because their ability to do direct acceleration was still the best; it thrived in the long run because they took it upon themselves to build the entire CUDA infrastructure to leverage shaders. This is where that data center growth comes from; Huang explained:

On the day that you become processor company, you have to internalize that this processor architecture is brand new. There’s never been a programmable pixel shader or a programmable GPU processor and a programming model like that before, and so we internalize. You have to internalize that this is a brand new programming model and everything that’s associated with being a program processor company or a computing platform company had to be created. So we had to create a compiler team, we have to think about SDKs, we have to think about libraries, we had to reach out to developers and evangelize our architecture and help people realize the benefits of it, and if not, even come close to practically doing it ourselves by creating new libraries that make it easy for them to port their application onto our libraries and get to see the benefits of it.

The first reason to recount this story is to note the parallels between the cost of shader complexity and the cost of ray tracing and AI in terms of current games; the second is to note that Nvidia’s approach to problem-solving has always been to do everything itself. Back then that meant developing CUDA for programming those shaders; today it means building out entire systems for AI.

Huang said during last week’s keynote:

Nvidia is dedicated to advancing science and industry with accelerated computing. The days of no-work performance scaling are over. Unaccelerated software will no longer enjoy performance scaling without a disproportionate increase in costs. With nearly three decades of a singular focus, Nvidia is expert at accelerating software and scaling computer by a 1,000,000x, going well beyond Moore’s Law.

Accelerated computing is a full-stack challenge. It demands deep understanding of the problem domain, optimizing across every layer of computing, and all three chips: CPU, GPU, and DPU. Scaling across multi-GPUs on multi-nodes is a datacenter-scale challenge, and requires treating the network and storage as part of the computing fabric, and developers and customers want to run software in many places, from PCs to super-computing centers, enterprise data centers, cloud, to edge. Different applications want to run in different locations, and in different ways.

Today, we’re going to talk about accelerated computing across the stack. New chips and how they will boost performance, far beyond the number of transistors, new libraries, and how it accelerates critical workloads to science and industry, new domain-specific frameworks, to help develop performant and easily deployable software. And new platforms, to let you deploy your software securely, safely, and with order-of-magnitude gains.

In Huang’s view, simply having fast chips is no longer enough for the workloads of the future: that is why Nvidia is building out entire data centers using all of its own equipment. Here again, though, a future where every company needs accelerated computing generally, and Nvidia to build it for them specifically — Nvidia’s Celestial City — is in contrast to the present where the biggest users of Nvidia chips in the data center are hyperscalers who have their own systems already in place.

A company like Meta, for example, doesn’t need Nvidia’s networking; they invented their own. What they do need are a lot of massively parallelizable chips to train their machine learning algorithms on, which means they have to pay Nvidia and their high margins. Small wonder that Meta, like Google before them, is building its own chip.

This is the course that all of the biggest companies will likely follow: they don’t need an Nvidia system, they need a chip that works in their system for their needs. That is why Nvidia is so invested in the democratization of AI and accelerated computing: the long term key to scale will be in building systems for everyone but the largest players. The trick to making it through the valley will be in seeing that ecosystem develop before Nvidia’s current big customers stop buying Nvidia’s expensive chips. Huang once saw that 3D accelerators would be commoditized and took a leap with shaders; one gets the sense he has the same fear with chips and is thus leaping into systems.

Metaverse in the Valley: Omniverse Nucleus

In the interview last spring I asked Huang if Nvidia would ever build a cloud service;

If we ever do services, we will run it all over the world on the GPUs that are in everybody’s clouds, in addition to building something ourselves, if we have to. One of the rules of our company is to not squander the resources of our company to do something that already exists. If something already exists, for example, an x86 CPU, we’ll just use it. If something already exists, we’ll partner with them, because let’s not squander our rare resources on that. And so if something already exists in the cloud, we just absolutely use that or let them do it, which is even better. However, if there’s something that makes sense for us to do and it doesn’t make for them to do, we even approach them to do it, other people don’t want to do it then we might decide to do it. We try to be very selective about the things that we do, we’re quite determined not to do things that other people do.

It turns out there was something no one else wanted to do, and that was create a universal database for 3D objects for use in what Nvidia is calling the Omniverse. These objects could be highly detailed millimeter-precise objects for use in manufacturing or supply chains, or they could be fantastical objects and buildings generated for virtual worlds; in Huang’s vision they would be available to anyone building on Omniverse Nucleus.

Here the Celestial City is a world of 3D experiences used across industry and entertainment — an Omniverse of metaverses, if you will, all connected to Nvidia’s cloud — and it’s ambitious enough to make Mark Zuckerberg blush! This valley, by the same token, seems even longer and darker: not only do all of these assets and 3D experiences need to be created, but entire markets need to be convinced of their utility and necessity. Building a cloud for a world that doesn’t yet exist is to reach for heights still out of sight.

There certainly is no questioning Huang and Nvidia’s ambition, although some may quibble with the wisdom of navigating three valleys all at once; it’s perhaps appropriate that the stock is in a valley itself, above and beyond that perfect storm in gaming.

What is worth considering, though, is that the number one reason why Nvidia customers — both in the consumer market and the enterprise one — get frustrated with the company is price: Nvidia GPUs are expensive, and the company’s margins — other than the last couple of quarters — are very high. Pricing power in Nvidia’s case, though, is directly downstream from Nvidia’s own innovations, both in terms of sheer performance in established workloads, and also in its investment in the CUDA ecosystem creating the tools for entirely new ones.

In other words, Nvidia has earned the right to be hated by taking the exact sort of risks in the past it is embarking on now. Suppose, for example, the expectation for all games in the future is not just ray tracing but full-on simulation of all particles: Nvidia’s investment in hardware will mean it dominates the era just as it did the rasterized one. Similarly, if AI applications become democratized and accessible to all enterprises, not just the hyperscalers, then it is Nvidia who will be positioned to pick up the entirety of the long tail. And, if we get to a world of metaverses, then Nvidia’s head start on not just infrastructure but on the essential library of objects necessary to make that world real (objects that will be lit by ray-tracing in AI-generated spaces, of course), will make it the most essential infrastructure in the space.

These bets may not all pay off; I do, though, appreciate the audacity of the vision, and won’t begrudge the future margins that may result in the Celestial City if Nvidia makes it through the valley.

  1. TSMC’s 3rd generation 5nm process 

  2. TSMC’s 1st generation 5nm process 

  3. TSMC’s 3rd generation 7nm process 

Sharp Tech and Stratechery Plus

I am excited to announce both a new podcast and a substantial expansion in the value of a Stratechery subscription. We’ll start with the podcast:

Sharp Tech with Ben Thompson

Sharp Tech with Ben Thompson is a new podcast from Andrew Sharp and myself about how technology works, and the ways it is impacting the world. We will publish one free episode weekly, and there are already six episodes in the catalog:

In addition, there will be a weekly subscriber-only episode that will be built on listener questions and feedback; the first paid episode dropped yesterday. You can get Sharp Tech for Apple Podcasts, Overcast, or the podcast player of your choice by loging in at the Sharp Tech website, or search for it in Spotify.

Here is the good news: Sharp Tech Premium is included with a Stratechery subscription.

That leads to today’s second announcement: the Stratechery Daily Update subscription is transforming into Stratechery Plus:

Stratechery Plus

Stratechery Plus is the same $12/month or $120/year price as the Stratechery Update, but it is now expanded to include not just the Stratechery Update and Stratechery Interviews but also Dithering and Sharp Tech.

Stratechery Plus includes the Stratechery Update, Stratechery Interviews, Sharp Tech, and Dithering

The Stratechery Update consists of substantial analysis of the news of the day delivered via three weekly emails or podcasts (including free bi-weekly Stratechery Articles). If you enjoy Stratechery Articles you will love the Stratechery Update.

Stratechery Interviews include interviews with leading public CEOs like Mark Zuckerberg, Jensen Huang, and Satya Nadella; the Founder Series with private company founders like Parker Conrad, Laura Behrens Wu, and Shishir Mehrotra; and discussions with fellow analysts like Eric Seufert, Matthew Ball, and Bill Bishop.

Dithering is a twice-weekly podcast from Daring Fireball’s John Gruber and myself: 15 minutes an episode, not a minute less, not a minute more. Dithering, which costs $5/month, was previously available as a $3 add-on for Stratechery subscribers; now it is available to all Stratechery subscribers. You can get Dithering for Apple Podcasts, Overcast, or the podcast player of your choice by logging in at the Dithering website.

This is, I hope, only the beginning for Stratechery Plus. Right now the content is obviously very Ben-centric, but my hope is to expand the offering over time. For now, I am delighted to be doing my part to make Stratechery more valuable than ever.

The AI Unbundling

My first job was as a paper boy:

Paper boy, by Midjourney

The job was remarkably analog: a bundle of newspapers would be dropped off at my house, I would wrap them in rubber-bands (or plastic bags if it were raining), load them up in a canvas sack, and set off on my bike; once a month my parents would drive me around to collect payment. Little did I appreciate just how essential my role writ large was to the profitability of newspapers generally.

Newspapers liked to think that they made money because people relied on them for news, furnished by their fearless reporters and hard-working editors; not only did people pay newspapers directly, but advertisers were also delighted to pay for the privilege of having their products placed next to the journalists’ peerless prose. The Internet revealed the fatal flaw in this worldview: what newspapers provided was distribution thanks to infrastructure like printing presses and yours truly.

Printing press, by Midjourney

Once the Internet reduced distribution costs to zero, three truths emerged: first, that “news”, once published, retained no economic value. Second, newspapers no longer had geographic monopolies, but were instead in competition with every publication across the globe. Third, advertisers didn’t care about content, but rather about reaching customers.

I illustrated these three truths in 2015’s Popping the Publishing Bubble:

Popping the Publishing Bundle

Editorial and ads used to be a bundle; next, the Internet unbundled editorial and ads, and provided countless options for both; the final step was ads moving to platforms that gave direct access to users, leaving newspapers with massive reach and no way to monetize it.

The Idea Propagation Value Chain

As much as newspapers may rue the Internet, their own business model — and my paper delivery job — were based on an invention that I believe is the only rival for the Internet’s ultimate impact: the printing press. Those two inventions, though, are only two pieces of the idea propagation value chain. That value chain has five parts:

The five parts of the idea propagation value chain: creation, substantiation, duplication, distribution, consumption

The evolution of human communication has been about removing whatever bottleneck is in this value chain. Before humans could write, information could only be conveyed orally; that meant that the creation, vocalization, delivery, and consumption of an idea were all one-and-the-same. Writing, though, unbundled consumption, increasing the number of people who could consume an idea.

Writing unbundled consumption from the rest of the value chain

Now the new bottleneck was duplication: to reach more people whatever was written had to be painstakingly duplicated by hand, which dramatically limited what ideas were recorded and preserved. The printing press removed this bottleneck, dramatically increasing the number of ideas that could be economically distributed:

The printing press unbundled duplication and distribution from creation

The new bottleneck was distribution, which is to say this was the new place to make money; thus the aforementioned profitability of newspapers. That bottleneck, though, was removed by the Internet, which made distribution free and available to anyone.

The Internet unbundled distribution from duplication

What remains is one final bundle: the creation and substantiation of an idea. To use myself as an example, I have plenty of ideas, and thanks to the Internet, the ability to distribute them around the globe; however, I still need to write them down, just as an artist needs to create an image, or a musician needs to write a song. What is becoming increasingly clear, though, is that this too is a bottleneck that is on the verge of being removed.

A flood emerging from a door ajar, by Midjourney

This image, like the first two in this Article, was created by AI (Midjourney, specifically). It is, like those two images, not quite right: I wanted “A door that is slightly open with light flooding through the crack”, but I ended up with a door with a crack of light down the middle and a literal flood of water; my boy on a bicycle, meanwhile, is missing several limbs, and his bike doesn’t have a handlebar, while the intricacies of the printing press make no sense at all.

They do, though, convey the idea I was going for: a boy delivering newspapers, printing presses as infrastructure, and the sense of being overwhelmed by the other side of an opening door — and they were all free.1 To put in terms of this Article, I had the idea, but AI substantiated it for me — the last bottleneck in the idea propagation value chain is being removed.

AI Democratization

What is notable about all of these AI applications it that they go back to language itself; Roon writes with regards to large language models (LLMs) on the Scale blog:

In a previous iteration of the machine learning paradigm, researchers were obsessed with cleaning their datasets and ensuring that every data point seen by their models is pristine, gold-standard, and does not disturb the fragile learning process of billions of parameters finding their home in model space. Many began to realize that data scale trumps most other priorities in the deep learning world; utilizing general methods that allow models to scale in tandem with the complexity of the data is a superior approach. Now, in the era of LLMs, researchers tend to dump whole mountains of barely filtered, mostly unedited scrapes of the Internet into the eager maw of a hungry model.

Roon’s focus is on text as the universal input, and connective tissue.2 Note how this insight fits into the overall development of communication: oral communication was a prerequisite to writing and reading; widespread literacy was a prerequisite to anyone being able to publish on the Internet; the resultant flood of text and images enabled by zero marginal distribution is the prerequisite for models that unbundle the creation of an idea and its substantiation.

This, by extension, hints at an even more surprising takeaway: the widespread assumption — including by yours truly — that AI is fundamentally centralizing may be mistaken. If not just data but clean data was presumed to be a prerequisite, then it seemed obvious that massively centralized platforms with the resources to both harvest and clean data — Google, Facebook, etc. — would have a big advantage. This, I would admit, was also a conclusion I was particularly susceptible to, given my focus on Aggregation Theory and its description of how the Internet, contrary to initial assumptions, leads to centralization.

The initial roll-out of large language models seemed to confirm this point of view: the two most prominent large language models have come from OpenAI and Google; while both describe how their text (GPT and GLaM, respectively) and image (DALL-E and Imagen, respectively) generation models work, you either access them through OpenAI’s controlled API, or in the case of Google don’t access them at all. But then came this summer’s unveiling of the aforementioned Midjourney, which is free to anyone via its Discord bot. An even bigger surprise was the release of Stable Diffusion, which is not only free, but also open source — and the resultant models can be run on your own computer.

There is, as you might expect, a difference in quality; Dall-E, for example, had the most realistic “newspaper delivery boy throwing a newspaper”:

Newspaper boys, by Dall-E

Stable Diffusion was on the other end of the spectrum:

Newspaper delivery boys, by Stable Diffusion

What is important to note, though, is the direction of each project’s path, not where they are in the journey. To the extent that large language models (and I should note that while I’m focusing on image generation, there are a whole host of companies working on text output as well) are dependent not on carefully curated data, but rather on the Internet itself, is the extent to which AI will be democratized, for better or worse.

The Impact on Creators

The worse is easy to envision; Charlie Warzel issued a mea culpa for using an AI image as an illustration in a post about Alex Jones:

I told Bors that what I felt worst about was how mindless my decision to use Midjourney ultimately had been. I was caught up in my own work and life responsibilities and trying to get my newsletter published in a timely fashion. I went to Getty and saw the same handful of photos of Alex Jones, a man who I know enjoys when his photo is plastered everywhere. I didn’t want to use the same photos again, nor did I want to use his exact likeness at all. I also, selfishly, wanted the piece to look different from the 30 pieces that had been published that day about Alex Jones and the Sandy Hook defamation trial. All of that subconsciously overrode all the complicated ethical issues around AI art that I was well apprised of.

What worries me about my scenario is that Midjourney was so easy to use, so readily accessible, and it solved a problem (abstracting Jones’ image in a visually appealing way), that I didn’t have much time or incentive to pause and think it through. I can easily see others falling into this like I did.

For these reasons, I don’t think I’ll be using Midjourney or any similar tool to illustrate my newsletter going forward (an exception would be if I were writing about the technology at a later date and wanted to show examples). Even though the job wouldn’t go to a different, deserving, human artist, I think the optics are shitty, and I do worry about having any role in helping to set any kind of precedent in this direction. Like others, I also have questions about the corpus used to train these art tools and the possibility that they are using a great deal of art from both big-name and lesser-known artists without any compensation or disclosure to those artists. (I reached out to Midjourney to ask some clarifying questions as to how they choose the corpus of data to train the tool, and they didn’t respond.)

I get Warzel’s point, and desire to show solidarity to artists worried about the impact of AI-generated art on their livelihoods. They are, it seems to me, right to worry: I opened this Article discussing the demise of newspapers which, once the connection between duplication and distribution was severed, quickly saw their business models fall apart. If the connection between idea creation and idea substantiation is being severed, it seems reasonable to assume all attendant business models might suffer the same fate.

There are, though, two rejoinders: the first is that abundance has its own reward. I am uniquely biased in this regard, seeing as how I make my living on the Internet as a publisher effectively competing with the New York Times, but I would argue that not just the quantity but, in absolute terms, the quality of content available to every single person in the world is dramatically higher than it was before the distribution bottleneck was removed. It seems obvious that removing the substantiation bottleneck from ideas will result in more good ones as well (along with, by definition, an even greater increase in not so good ones).

The analogy to publishing also point to what will be the long-term trend for any profession affected by these models: relatively undifferentiated creators who depended on the structural bundling of idea creation and substantiation will be reduced to competing with zero marginal cost creators for attention generated and directed from Aggregators; highly differentiated creators, though, who can sustainably deliver both creation and substantiation on their own will be even more valuable. Social media, for example, has been a tremendous boon to differentiated publishers: it gives readers a megaphone to tell everyone how great said publisher is. These AI tools will have a similar effect on highly differentiated creators, who will leverage text-based iteration to make themselves more productive and original than ever before.

The second rejoinder is perhaps more grim: this is going to happen regardless. Warzel may be willing to overlook the obvious improvement in not just convenience but also, for his purposes, quality proffered by his use of Midjourney, but few if any will make the same choice. AI-generated images will, per the image above, soon be a flood, just as publishing on the Internet quickly overwhelmed the old newspaper business model.

Moreover, just as native Internet content is user-generated content, the iterative and collaborative nature of AI-generated content — both in the sense of being a by-product of content already created, and also the fact that every output can be further iterated upon by others — will prove to be much more interesting and scalable than what professional organizations can produce. TikTok, which pulls content from across its network to keep users hooked, is the apotheosis of user-generated content; Metaverses may be the apotheosis of AI-generated content.

I wrote a follow-up to this Article in this Daily Update.

  1. Beyond the $600 annual fee I paid to Midjourney to have access to the fully rights-unencumbered Corporate plan 

  2. In the case of these image applications, noise is added to an known image and then the model is trained on backing out the image from pure noise; the resultant model can then be applied to any arbitrary text applied to pure noise, based on further training of matched text and images 

The Services iPhone

Back when Stratechery first launched there was no bigger day than iPhone announcement day, and there was arguably no bigger year than 2013, when I was just getting started. That was the year Apple was, for the first time, coming out with a non-top-of-the-line iPhone, what ultimately became known as the iPhone 5C. We would find out later that the iPhone 5C was a bit of a one-off: the iPhone 5 with its chamfered edges was simply too expensive to make (and the edges chipped horribly), making it unsuitable for Apple’s “our cheaper phones are our older phones with a lower price” strategy; Apple would one day come out with the iPhone SE to address the lower end, but the real expansion came when the iPhone X was priced at $300 more than the iPhone 8 — up-market, not down.

Still, it’s interesting to look back at how much time I spent on that product launch:

  • Before the event I wrote Thinking about iPhone Pricing, where I guessed that the 5C would cost $450 (but made the case $550 might make more sense).
  • After the event I wrote Two Minutes, Fifty-six Seconds, which referred to how long into the keynote it took me to realize that $550 would be the price; it was clear that Apple was focused on differentiation, justifying its prices, not lowering them.
  • A week later I wrote The $550 iPhone 5C Makes Perfect Sense, which was mostly about subsidies and how they might have impacted Apple’s pricing decisions.
  • A couple of days after that I wrote what I consider one of the seminal Articles of early Stratechery — the ideas in What Clayton Christensen Got Wrong are one of the reasons I wanted to start the site. This was in response to critics who were sure Apple was setting themselves up for disruption with their pricing strategy; my argument was that differentiation in the user experience derived from integration was a sustainable moat that justified higher pricing.
  • Finally, a week after that I wrote Obsoletive, which made the case that thinking about the iPhone as a disruptive product was mistaken; rather, what made it so compelling — including its high price — was the fact it obsoleted so many other products in our lives.

(Oh, and for what it’s worth, I asked a month later: So the 5S is (allegedly) killing the 5C. Why is this bad news?. If customers were actually heavily favoring the more expensive 5S then this actually made all of the points I had written over the previous month.)

I was originally inspired to look up this history as part of an introduction explaining why iPhone events no longer seem Article worthy, but are worth saving an Update slot for; what struck me while reading through these old pieces though, is not only the degree to which they show Apple’s consistency, but also how much the company has changed — and that is Article worthy.

Apple’s Increasing ARPU

CEO Tim Cook is fond of citing Apple’s ability to integrate hardware and software; over the last few years he has taken care to add “and services” as well. What was interesting about his opening in yesterday’s keynote, though, was that he has now moved up to a higher level of abstraction: devices.

Products that are intuitive and easy-to-use, that have a unique integration of hardware and software, and that are incredibly personal. Today we’re here to talk about three products that have become essential in our lives: iPhone, AirPods, and Apple Watch. They’re always with you, whenever and wherever you need them, and are designed to work seamlessly together. On their own, each is industry-leading. Together, they provide a magical experience.

This is an expression of a strategy that became clear several years ago; I wrote in Apple’s Middle Age:

Apple’s growth is almost entirely inwardly-focused when it comes to its user-base: higher prices, more services, and more devices.

Very few people just buy an iPhone: they upgrade to a higher-priced model, they spend money in the App Store and on subscriptions, and they buy an Apple Watch and AirPods that work seamlessly with their phone. The end result is that Apple isn’t making $550 per customer, to go back to the iPhone 5C, or $650 in the case of the 5S: they’re making upwards of $2000 — $1,000+ for a top-of-the-line iPhone, $400+ for a Watch, $200+ for AirPods, and all of that App Store revenue (and this doesn’t even include what is likely a thriving accessories business, AppleCare, or the Google search deal).

It is, to be frank, justified: all of these devices really do work well together, and iPhones remain top-of-class. And, for all of the kvetching about the App Store, and Apple’s arguably anticompetitive actions to maintain its control, it is true that the concept was revolutionary.

The Service Narrative

It was in January 2016 that Apple first articulated the so-called “Services Narrative”; after a quarter in which the company’s new iPhone 6S posted relatively disappointing sales, CFO Luca Maestri made the case on the earnings call that it was a mistake to think of Apple as a hardware company, subject to the vagaries of consumer demand. I wrote at the time (forgive the long excerpt, but it’s directly applicable to the point):

Cook and CFO Luca Maestri went to a lot of effort to sell the narrative that Apple is becoming a services company, and frankly, I think they kind of overdid it.

Specifically, Apple created a new way of evaluating their services called “Installed Base Related Purchases.” This is basically Apple’s services revenue plus the amount they pay to developers from the App Store and to most digital content owners on the iTunes Store. Said payouts don’t appear on Apple’s balance sheet, nor should they: Apple isn’t some sort of middle man for Candy Crush Saga, buying it wholesale and then selling it at a profit. Rather, they are facilitating a transaction between a content creator and a consumer, and taking a 30% tax.

Moreover, one of the benefits of being recognized as a services company is that your revenue is valued more highly with the presumption that it is higher margin; by adding in the 70% Apple pays out they are certainly able to crow about a higher revenue number, but they are dramatically reducing their associated margin by doing so. It’s silly.

To be sure, Apple’s services revenue numbers are impressive (although it should be noted that services revenue on a per-active-user basis actually decreased year-over-year). But it is very clear that the company remains a differentiated hardware company, as evidenced by everyone’s favorite question:

In the past, Apple’s been very known in always having a premium product. With the slowdown in the macro FX and also GDP revision, is Apple’s strategy go-to-market still always at premium product, or is there a need to go to more also a middle market or lower price point to attract more customers?

Ah, it’s the ol’ “Will Apple make a cheaper iPhone?” query. This one, though, was smarter than it appears, thanks to the next sentence:

Just because it seems like growing that installed base and services, as you pointed out, really economically could really help out Apple in the long-term.

As I’ve written innumerable times, services (horizontal) and hardware (vertical) companies have very different strategic priorities: the former ought to maximize their addressable market (by, say, making a cheaper iPhone), while the latter ought to maximize their differentiation. And, Cook’s answer made clear what Apple’s focus remains:

Our strategy is always to make the best products…We have the premium part of our line is the 6s and the 6s Plus. We also have a mid-price point, with the iPhone 6 and the iPhone 6 Plus. And we continue to offer the iPhone 5s in the market and it continues to do quite well. And so we offer all of those and I don’t see us deviating from that approach.

To be clear, I think this is the exact right approach for Apple, and as I noted above, I think these results show that the strategy continues to work. But let’s be honest: that means Apple is not a services company; they have a nice services revenue stream, but the company is rightly judged now and for the foreseeable future on the performance of its hardware.

I revisited this take a few years later, in an Update entitled Apple the Services Company (for Real!), where I noted that Apple’s estimated $44/user per year in services revenue still paled in comparison to Google or Facebook, but was meaningful all the same, particularly once you realized that Apple’s users had to spend a whole lot of money to even enter their ecosystem. I concluded:

Of course only looking at services revenue is not quite right either: that number does not include the cost of the iPhone itself, the price of entry to Apple’s ecosystem (although that payment may not necessarily flow to Apple, as is the case with phones handed down or resold). Apple, though, by de-emphasizing unit sales and focusing on the installed base, is making clear that the number that matters is average revenue per installed device; in other words, to the extent a company is what it measures — or at least reports — Apple is now a services company for real.

This was perhaps a bit generous, given the strategic tension I noted in the earlier excerpt: Apple may have been shifting its metrics to ones that reflected a focus on services, but at the end of the day, the company was still charging a pretty high price for entry.

The Services Company

I thought the products Apple introduced yesterday were pretty impressive:

  • The Apple Watch Ultra looks like the SUV of watches: it will be sold as a tool for extreme athletes, and mostly bought because it is so clearly new and different, clearly distinguishable from all of the other Apple Watches on the market.
  • AirPods Pro was a killer product in its first iteration; the 2nd generation looks set to address the 1st generation’s few flaws while delivering the sort of improvements we might expect from every tech product.
  • The iPhone Pro is increasingly differentiated from the iPhone, not just in terms of having a faster processor — a first — but also with software-driven differentiation like the new Dynamic Island functionality.

The most surprising announcement of all, though, were the prices. Everything stayed the same! This was not what I, or close followers of Apple like John Gruber, expected at all. After all, Apple’s strategy the past several years seemed to be focused on wringing more revenue out of existing customers. More importantly, the last year has seen a big increase in inflation:

U.S. inflation over the last five years

What this means is that in real terms Apple’s products actually got cheaper. Apple did, to be sure, raise prices around the world, but this is better explained by the fact the company runs on the dollar, which is the strongest in years; to put it another way, those foreign prices are derived from the U.S. price, and that price stayed the same, which means the price is lower.

This doesn’t make much sense for the product company Apple has always been thought to be, and doesn’t fully align with the approach I laid out in Apple’s Middle Age. It does, though, make all kinds of sense for a services company, which is focused first-and-foremost on increasing its install base. Indeed, this is the missing piece from that Update I wrote about Apple’s changing metrics. To measure its business based on users, not products, was to measure like a services company; to lower the prices of the products that lead to services revenue is to price like one.

This is, in a weird way, a relief: it has been disconcerting for people who think of Apple as a product company to see the company fight so fiercely for its App Store model, and to see the way it is willing to approach if not cross the line of anticompetitive behavior when it comes to App Tracking Transparency and its clear ambitions in the advertising space. To declare that the company is now clearly driven by Services doesn’t refute these narratives; rather, it at least justifies them, because they are exactly what a Services company ought to do. Here’s hoping that the products that made the company great don’t suffer from what is, at this point, a clear shift in strategy.

I wrote a follow-up to this Article in this Daily Update.

Rights, Laws, and Google

The first and most important takeaway from Kashmir Hill’s excellent article in the New York Times about Mark, the man flagged by Google as a purveyor of Child Sexual Abuse Material (CSAM) for taking pictures of his son’s penis and sending them to their family doctor, and who subsequently lost nearly every aspect of his digital life when Google deleted his account, are the tremendous trade-offs entailed in the indiscriminate scanning of users’ cloud data.

On one hand, it seems like an incredible violation of privacy to have a private corporation effectively looking through every photo you upload, particularly when those uploads happen as part of the expected way in which your smartphone operates (users technically agree to this scanning, but as part of an endless End User License Agreement that is both ridiculously long and, more pertinently, inescapable if you want to use your phone as it was intended). Moreover, Google doesn’t simply scan for CSAM photos that are already known to exist via the PhotoDNA database of photos of exploited children; the company also leverages machine learning to look for new CSAM that hasn’t yet been identified as such.

On the other hand, as horrific as the material in the PhotoDNA database is, much of it has been floating around the Internet for years, which is to say the abuse depicted happened long ago; Google’s approach has the potential to discover abuse as it is happening, making it possible for the authorities to intercede and rescue the child in question. Hill’s story noted that in 2021 the CyberTipline at the National Center for Missing and Exploited Children, the only entity legally allowed to hold CSAM (NCMEC also manages the PhotoDNA database), “alerted authorities to ‘over 4,260 potential new child victims'”. We don’t know how many of those children were subsequently rescued, but a question worth posing to anyone unilaterally opposed to Google’s approach is how big that number would have to be to have made it worthwhile?

But, to return to the original hand, one of those 4,260 potential new child victims was Mark’s son (and another was taken by Cassio, a second father found by Hill caught in the same predicament, for the same reasons): the question for those applauding Google’s approach is how big the number of false positives would have to be to shut the whole thing down?

It was the exploration of these trade-offs that was at the heart of the Update I wrote about Hill’s story last week; as I noted there are no easy answers:

Nearly every aspect of this story is incredibly complex, and I understand and respect arguments on both sides: should there be scanning of cloud-related content? Should machine learning be leveraged to find new photos? Is it reasonable to obliterate someone’s digital life — except for what you give the police — given the possibility that they may be committing horrific crimes? These are incredibly difficult questions, particularly in the absence of data, because the trade-offs are so massive.

However, it seemed to me that one aspect of the case was very clear:

There is, though, one part of the story that is black-and-white. Google is unquestionably wrong to not restore the accounts in question. In fact, I am stunned by the company’s approach in these cases. Even if you grant the arguments that this awesome exercise of surveillance is warranted, given the trade-offs in question, that makes it all the more essential that the utmost care be taken in case the process gets it wrong. Google ought to be terrified it has this power, and be on the highest alert for false positives; instead the company has gone in the opposite direction, setting itself as judge, jury, and executioner, even when the people we have collectively entrusted to lock up criminals ascertain there was no crime. It is beyond arrogant, and gives me great unease about the company generally, and its long-term investments in AI in particular.

Not that it matters, one may argue: Google can do what they want, because they are a private company. That is an argument that may ring familar.

Tech and Liberty

In 2019 I discussed the distinction between public and private restrictions on speech in Tech and Liberty:

Alexander Hamilton was against the Bill of Rights, particularly the First Amendment. This famous xkcd comic explains why:

Free Speech by xkcd

According to Randall Munroe, the author, the “Right to Free Speech” is granted by the First Amendment, which was precisely the outcome Hamilton feared in Federalist No. 84:

I go further, and affirm that bills of rights, in the sense and to the extent in which they are contended for, are not only unnecessary in the proposed Constitution, but would even be dangerous. They would contain various exceptions to powers not granted; and, on this very account, would afford a colorable pretext to claim more than were granted. For why declare that things shall not be done which there is no power to do? Why, for instance, should it be said that the liberty of the press shall not be restrained, when no power is given by which restrictions may be imposed? I will not contend that such a provision would confer a regulating power; but it is evident that it would furnish, to men disposed to usurp, a plausible pretense for claiming that power. They might urge with a semblance of reason, that the Constitution ought not to be charged with the absurdity of providing against the abuse of an authority which was not given, and that the provision against restraining the liberty of the press afforded a clear implication, that a power to prescribe proper regulations concerning it was intended to be vested in the national government. This may serve as a specimen of the numerous handles which would be given to the doctrine of constructive powers, by the indulgence of an injudicious zeal for bills of rights.

Hamilton’s argument is that because the U.S. Constitution was created not as a shield from tyrannical kings and princes, but rather by independent states, all essential liberties were secured by the preamble (emphasis original):

WE, THE PEOPLE of the United States, to secure the blessings of liberty to ourselves and our posterity, do ORDAIN and ESTABLISH this Constitution for the United States of America.

Hamilton added:

Here, in strictness, the people surrender nothing; and as they retain every thing they have no need of particular reservations.

Munroe, though, assumes the opposite: liberty, in this case the freedom of speech, is an artifact of law, only stretching as far as government action, and no further. Pat Kerr, who wrote a critique of this comic on Medium in 2016, argued that this was the exact wrong way to think about free speech:

Coherent definitions of free speech are actually rather hard to come by, but I would personally suggest that it’s something along the lines of “the ability to voluntarily express (and receive) opinions without suffering excessive penalties for doing so”. This is a liberal principle of tolerance towards others. It’s not an absolute, it isn’t comprehensive, it isn’t rigorously defined, and it isn’t a law.

What it is is a culture.

The context of that 2019 Article was the differing decisions between Facebook and Twitter in terms of allowing political ads on their platforms; over the ensuing three years the willingness and length to which these and other large tech platforms have been willing to go to police speech has expanded dramatically, even as the certainty that private censorship is ‘good actually’ has become conventional wisdom. I found this paragraph in a New York Times article about Elon Musk’s attempts to buy Twitter striking:

The plan jibes with Mr. Musk’s, Mr. Dorsey’s and Mr. Agrawal’s beliefs in unfettered free speech. Mr. Musk has criticized Twitter for moderating its platform too restrictively and has said more speech should be allowed. Mr. Dorsey, too, grappled with the decision to boot former President Donald J. Trump off the service last year, saying he did not “celebrate or feel pride” in the move. Mr. Agrawal has said that public conversation provides an inherent good for society. Their positions have increasingly become outliers in a global debate over free speech online, as more people have questioned whether too much free speech has enabled the spread of misinformation and divisive content.

In other words, the culture has changed; the law persists, but it does not and, according to the New York Times, ought not apply to private companies.


The Google case is not about the First Amendment, either legally or culturally. The First Amendment is not absolute, and CSAM is an obvious example. In 1957’s Roth v. United States the Supreme Court held that obscene speech was not protected by the First Amendment; Justice William Brennan Jr. wrote:

All ideas having even the slightest redeeming social importance — unorthodox ideas, controversial ideas, even ideas hateful to the prevailing climate of opinion — have the full protection of the guaranties, unless excludable because they encroach upon the limited area of more important interests. But implicit in the history of the First Amendment is the rejection of obscenity as utterly without redeeming social importance. This rejection for that reason is mirrored in the universal judgment that obscenity should be restrained, reflected in the international agreement of over 50 nations, in the obscenity laws of all of the 48 States, and in the 20 obscenity laws enacted by the Congress from 1842 to 1956.

This reasoning is a reminder that laws ultimately stem from culture; still, the law being the law, definitions were needed, which the Supreme Court provided in 1973’s Miller v. California. Obscene works (1) appeal to the prurient interest in sex, (2) portrays in a patently offensive way sexual conduct specifically defined by a relevant law and (3) lack serious literary, artistic, political, or scientific value. The Supreme Court went further in terms of CSAM in 1982’s New York v. Ferber, holding that the harm inflicted on children is sufficient reason to make all forms of CSAM illegal, above and beyond the standards set forth by Miller. Justice Byron White wrote:

Recognizing and classifying child pornography as a category of material outside the protection of the First Amendment is not incompatible with our earlier decisions. “The question whether speech is, or is not, protected by the First Amendment often depends on the content of the speech”…

The test for child pornography is separate from the obscenity standard enunciated in Miller, but may be compared to it for the purpose of clarity. The Miller formulation is adjusted in the following respects: a trier of fact need not find that the material appeals to the prurient interest of the average person; it is not required that sexual conduct portrayed be done so in a patently offensive manner; and the material at issue need not be considered as a whole. We note that the distribution of descriptions or other depictions of sexual conduct, not otherwise obscene, which do not involve live performance or photographic or other visual reproduction of live performances, retains First Amendment protection. As with obscenity laws, criminal responsibility may not be imposed without some element of scienter on the part of the defendant.

“Scienter”, the “knowledge of the nature of one’s act”, is what ties this judicial history back to the original discussion of Google’s actions against Mark. As Hill explained in the New York Times:

I have seen the photos that Mark took of his son. The decision to flag them was understandable: They are explicit photos of a child’s genitalia. But the context matters: They were taken by a parent worried about a sick child.

The problem in this case comes from who is determining scienter.

Google and the Bill of Rights

Quite clearly Mark did not intend for the pictures he took for his son’s telemedicine to be used for pornographic purposes. The San Francisco Police Department, which had been notified by Google after a human reviewer confirmed the machine learning-driven discovery of Mark’s photos of his son, agreed. From Hill’s story:

In December 2021, Mark received a manila envelope in the mail from the San Francisco Police Department. It contained a letter informing him that he had been investigated as well as copies of the search warrants served on Google and his internet service provider. An investigator, whose contact information was provided, had asked for everything in Mark’s Google account: his internet searches, his location history, his messages and any document, photo and video he’d stored with the company.

The search, related to “child exploitation videos,” had taken place in February, within a week of his taking the photos of his son. Mark called the investigator, Nicholas Hillard, who said the case was closed. Mr. Hillard had tried to get in touch with Mark but his phone number and email address hadn’t worked. “I determined that the incident did not meet the elements of a crime and that no crime occurred,” Mr. Hillard wrote in his report. The police had access to all the information Google had on Mark and decided it did not constitute child abuse or exploitation.

Mark asked if Mr. Hillard could tell Google that he was innocent so he could get his account back. “You have to talk to Google,” Mr. Hillard said, according to Mark. “There’s nothing I can do.” Mark appealed his case to Google again, providing the police report, but to no avail…A Google spokeswoman said the company stands by its decisions, even though law enforcement cleared the two men.

In short, the questions about Google’s behavior are not about free speech; they do, though, touch on other Amendments in the Bill of Rights. For example:

  • The Fourth Amendment bars “unreasonable searches and seizures”; while you can make the case that search warrants were justified once the photos in question were discovered, said photos were only discovered because Mark’s photo library was indiscriminately searched in the first place.
  • The Fifth Amendment says no person shall be deprived of life, liberty, or property, without due process of law; Mark lost all of his data, email account, phone number, and everything else Google touched forever with no due process at all.
  • The Sixth Amendment is about the rights to a trial; Mark was not accused of any crime in the real world, but when it came to his digital life Google was, as I noted, “judge, jury, and executioner” (the Seventh Amendment is, relatedly, about the right to a jury trial for all controversies exceeding $20).

Again, Google is not covered by the Bill of Rights; all of these Amendments, just like the First, only apply to the government. The reason why this case is useful, though, is it is a reminder that specific legal definitions are distinct from questions of right or wrong.

Working backwards, Google isn’t legally compelled to give Mark a hearing about his digital life (Sixth Amendment); they are wrong not to. Google isn’t legally compelled to give Mark due process before permanently deleting his digital life (Fifth Amendment); they are wrong not to. Google isn’t legally compelled to not search all of the photographs uploaded to Google (by default, if you click through all of the EULA’s); they are…well, this is where it gets complicated.

I started out this Article discussing the impossible trade-offs presented by questions of CSAM. People can and do make the case that to not search for this vileness, particularly if there is a chance that it can lead to the rescue of an abused child, is its own wrong. Resolving this trade-off in this way, though — that is, to violate the spirit and culture of the Fourth Amendment — makes it all the more essential to honor the spirit and culture of the Fifth and Sixth.

Paper Barriers

James Madison answered Hamilton’s objections in a speech to Congress introducing the Bill of Rights. What is interesting is that while Hamilton took it as a given that people would know and value their rights, Madison assumed the culture would run in the opposite direction, making an articulation of those rights important not just to restrain the government, but to remind the majority to not trample the rights of the minority:

But I confess that I do conceive, that in a Government modified like this of the United States, the great danger lies rather in the abuse of the community than in the Legislative body. The prescriptions in favor of liberty ought to be levelled against that quarter where the greatest danger lies, namely, that which possesses the highest prerogative of power. But this is not found in either the Executive or Legislative departments of Government, but in the body of the people, operating by the majority against the minority.

It may be thought that all paper barriers against the power of the community are too weak to be worthy of attention. I am sensible they are not so strong as to satisfy gentlemen of every description who have seen and examined thoroughly the texture of such a defence; yet, as they have a tendency to impress some degree of respect for them, to establish the public opinion in their favor, and rouse the attention of the whole community, it may be one means to control the majority from those acts to which they might be otherwise inclined.

This Article is a manifestation of Madison’s hope. Start with the reality that it seems quaint in retrospect to think that any of the Bill of Rights would be preserved absent the force of law. This is one of the great lessons of the Internet and the rise of Aggregators: when suppressing speech entailed physically disrupting printing presses or arresting pamphleteers, then restricting government, which retains a monopoly on real world violence, was sufficient to preserve speech. Along the same lines, there was no need to demand due process or a restriction on search and seizure on any entity but the government, because only the government could take your property or send you to jail.

Aggregators, though, make private action much more possible and powerful than ever before: yes, if you are kicked off of Twitter or Facebook, you can still say whatever you want on a street corner; similarly, if you lose all of your data and phone and email, you are still not in prison — and thank goodness that is the case! At the same time, it seems silly to argue that getting banned from a social media platform isn’t an infringement on individual free speech rights, even if it is the corporations’ own free speech rights that enable them to do just that legally, just as it is silly to argue that losing your entire digital life without recourse isn’t a loss of property without due process. The big Internet companies are manifesting Madison’s fears of the majority operating against the minority, and there is nothing the Bill of Rights can do about it.

What remains are those paper barriers, and what respect they might still engender, if it is possible to “rouse the attention of the whole community.” Rights are larger than laws, and Google has violated the former, even if they are not bound by the latter. The company ought not only change its policy with regards to Mark and Cassio, but fundamentally re-evaluate the balance it has struck between its unprecedented power over people’s lives and the processes it has in place to ensure that power is not abused. If it doesn’t, the people ought to, with what power they still conserve, do it for them.

I wrote a follow-up to this Article in this Daily Update.

Instagram, TikTok, and the Three Trends

Back in 2010, during my first year of Business School, I helped give a presentation entitled “Twitter 101”:

The introductory slide from a Twitter 101 presentation in business school

My section was “The Twitter Value Proposition”, and after admitting that yes, you can find out what people are eating for lunch on Twitter, I stated “The truth is you can find anything you want on Twitter, and that’s a good thing.” The Twitter value proposition was that you could “See exactly what you need to see, in real-time, in one place, and nothing more”; I illustrated this by showing people how they could unfollow me:

A slide noting that Twitter is what you make of it

The point was that Twitter required active management of your feed, but if you put in the effort, you could get something uniquely interesting to you that was incredibly valuable.

Most of the audience didn’t take me up on it.

Facebook Versus Instagram

If there is one axiom that governs the consumer Internet — consumer anything, really — it is that convenience matters more than anything. That was the problem with Twitter: it just wasn’t convenient for nearly enough people to figure out how to follow the right people. It was Facebook, which digitized offline relationships, that dominated the social media space.

Facebook’s social graph was the ultimate growth hack: from the moment you created an account Facebook worked assiduously to connect you with everyone you knew or wish you knew from high school, college, your hometown, workplace, you name an offline network and Facebook digitized it. Of course this meant that there were far too many updates and photos to keep track of, so Facebook ranked them, and presented them in a feed that you could scroll endlessly.

Users, famously, hated the News Feed when it was first launched: Facebook had protesters outside their doors in Palo Alto when it was introduced, and far more online; most were, ironically enough, organized on Facebook. CEO Mark Zuckerberg penned an apology:

We really messed this one up. When we launched News Feed and Mini-Feed we were trying to provide you with a stream of information about your social world. Instead, we did a bad job of explaining what the new features were and an even worse job of giving you control of them. I’d like to try to correct those errors now…

The errors to be corrected were better controls over what might be shared; Facebook did not give the users what they claimed to want, which was abolishing the News Feed completely. That’s because the company correctly intuited a significant gap between its users stated preference — no News Feed — and their revealed preference, which was that they liked News Feed quite a bit. The next fifteen years would prove the company right.

It was hard to not think of that non-apology apology while watching Adam Mosseri’s Instagram update three weeks ago; Mosseri was clear that videos were going to be an ever great part of the Instagram experience, along with recommended posts. Zuckerberg reiterated the point on Facebook’s earnings call, noting that recommended posts in both Facebook and Instagram would continue to increase. A day later Mosseri told Casey Newton on Platformer that Instagram would scale back recommended posts, but was clear that the pullback was temporary:

“When you discover something in your feed that you didn’t follow before, there should be a high bar — it should just be great,” Mosseri said. “You should be delighted to see it. And I don’t think that’s happening enough right now. So I think we need to take a step back, in terms of the percentage of feed that are recommendations, get better at ranking and recommendations, and then — if and when we do — we can start to grow again.” (“I’m confident we will,” he added.)

Michael Mignano calls this recommendation media in an article entitled The End of Social Media:

In recommendation media, content is not distributed to networks of connected people as the primary means of distribution. Instead, the main mechanism for the distribution of content is through opaque, platform-defined algorithms that favor maximum attention and engagement from consumers. The exact type of attention these recommendations seek is always defined by the platform and often tailored specifically to the user who is consuming content. For example, if the platform determines that someone loves movies, that person will likely see a lot of movie related content because that’s what captures that person’s attention best. This means platforms can also decide what consumers won’t see, such as problematic or polarizing content.

It’s ultimately up to the platform to decide what type of content gets recommended, not the social graph of the person producing the content. In contrast to social media, recommendation media is not a competition based on popularity; instead, it is a competition based on the absolute best content. Through this lens, it’s no wonder why Kylie Jenner opposes this change; her more than 360 million followers are simply worth less in a version of media dominated by algorithms and not followers.

Sam Lessin, a former Facebook executive, traced this evolution from the analog days to what is next in a Twitter screenshot entitled “The Digital Media ‘Attention’ Food Chain in Progress”:

Lessin’s five steps:

  1. The Pre-Internet ‘People Magazine’ Era
  2. Content from ‘your friends’ kills People Magazine
  3. Kardashians/Professional ‘friends’ kill real friends
  4. Algorithmic everyone kills Kardashians
  5. Next is pure-AI content which beats ‘algorithmic everyone’

This is a meta observation and, to make a cheap play on words, the first reason why it made sense for Facebook to change its name: Facebook the app is eternally stuck on Step 2 in terms of entertainment (the app has evolved to become much more of a utility, with a focus on groups, marketplace, etc.). It’s Instagram that is barreling forward. I wrote last summer about Instagram’s Evolution:

The reality, though, is that this is what Instagram is best at. When Mosseri said that Instagram was no longer a photo-sharing app — particularly a “square photo-sharing app” — he was not making a forward-looking pronouncement, but simply stating what has been true for many years now. More broadly, Instagram from the very beginning — including under former CEO Kevin Systrom — has been marked first and foremost by evolution.

To put this in Lessin’s framework, Instagram started out as a utility for adding filters to photos put on other social networks, then it developed into a social network in its own right. What always made Instagram different than Facebook, though, is the fact that its content was default-public; this gave the space for the rise of brands, meme and highlight accounts, and the Instagram influencer. Sure, some number of people continued to use Instagram primarily as a social network, but Meta, more than anyone, had an understanding of how Instagram usage had evolved over time.

Kylie Jenner and Kim Kardashian asking Instagram to be Instagram

In other words, when Kylie Jenner posts a petition demanding that Meta “Make Instagram Instagram again”, the honest answer is that changing Instagram is the most Instagram-like behavior possible.

Three Trends

Still, it’s understandable why Instagram did back off, at least for now: the company is attempting to navigate three distinct trends, all at the same time.

The first trend is the shift towards ever more immersive mediums. Facebook, for example, started with text but exploded with the addition of photos. Instagram started with photos and expanded into video. Gaming was the first to make this progression, and is well into the 3D era. The next step is full immersion — virtual reality — and while the format has yet to penetrate the mainstream this progression in mediums is perhaps the most obvious reason to be bullish about the possibility.

The trend in mediums online

The second trend is the increase in artificial intelligence. I’m using the term colloquially to refer to the overall trend of computers getting smarter and more useful, even if those smarts are a function of simple algorithms, machine learning, or, perhaps someday, something approaching general intelligence. To go back to Facebook, the original site didn’t have any smarts at all: it was just a collection of profile pages. Twitter came along and had the timeline, but the only smarts there was the ability to read a time stamp: all of the content was presented in chronological order. What made Facebook’s News Feed work was the application of ranking: from the very beginning the company tried to present users the content from their network that it thought you might be most interested in, mostly using simple signals and weights. Over time this ranking algorithm has evolved into a machine-learning driven model that is constantly iterating based on every click and linger, but on the limited set of content constrained by who you follow. Recommendations is the step beyond ranking: now the pool is not who you follow but all of the content on the entire network; it is a computation challenge that is many orders of magnitude beyond mere ranking (and AI-created content another massive step-change beyond that).

The trend in AI and content online

The third trend is the change in interaction models from user-directed to computer-controlled. The first version of Facebook relied on users clicking on links to visit different profiles; the News Feed changed the interaction model to scrolling. Stories reduced that to tapping, and Reels/TikTok is about swiping. YouTube has gone further than anyone here: Autoplay simply plays the next video without any interaction required at all.

The trend in UI online

One of the reasons Instagram got itself in trouble over the last few months is by introducing changes along all of these vectors at the same time. The company introduced more video into the feed (Trend 1), increased the percentage of recommended posts (Trend 2), and rolled out a new version of the app that was effectively a re-skinned TikTok to a limited set of users (Trend 3). It stands to reason that the company would have been better off doing one at a time.

That, though, would only be a temporary solution: it seems likely that all of these trends are inextricably intertwined.

Medium, Computing, and Interaction Models

Start with medium: text is easy, which is why it was the original medium of the Internet; effectively anyone can create it. The first implication is that there is far more text on the Internet than anything else; it also follows that the amount of high quality text is correspondingly high as well (a small fraction of a large number is still very large). The second implication has to do with AI: it is easier to process and glean insight from text. Text, meanwhile, takes focus and the application of an acquired skill for humans to interpret, not dissimilar to the deliberate movement of a mouse to interact with a link.

Photos used to be more difficult: digital cameras came along around the same time as the web, but even then you needed to have a dedicated device, move those photos to your computer, then upload them to a network. What is striking about the impact of smartphones is that not only did they make the device used to take pictures the same device used to upload and consume them, but they actually made it easier to take a picture than to write text. Still, it took time for AI to catch up: at first photos were ranked using the metadata surrounding them; only over the last few years has it become possible for services to understand what the photo actually is. The most reliable indicator of quality — beyond a like — remains the photo that you stop at while scrolling.

The ease of making a video followed a similar path to photos, but more extreme: making and uploading your own videos before the smartphone was even more difficult than photos; today the mechanics are just as easy, and it’s arguably even easier to make something interesting, given the amount of information conveyed by a video relative to photos, much less a text. Still, videos require more of a commitment than text or photos, because consuming them takes time; this is where the user interaction layer really matters. Lessin again, in another Twitter screenshot:

I saw someone recently complaining that Facebook was recommending to them…a very crass but probably pretty hilarious video. Their indignant response [was that] “the ranking must be broken.” Here is the thing: the ranking probably isn’t broken. He probably would love that video, but the fact that in order to engage with it he would have to go proactively click makes him feel bad. He doesn’t want to see himself as the type of person that clicks on things like that, even if he would enjoy it.

This is the brilliance of Tiktok and Facebook/Instagram’s challenge: TikTok’s interface eliminates the key problem of what people want to view themselves as wanting to follow/see versus what they actually want to see…it isn’t really about some big algorithm upgrade, it is about relesing emotional inner tension for people who show up to be entertained.

This is the same tension between stated and revealed preference that Facebook encountered so many years ago, and its exactly why I fully expect the company to, after this pullback, continue to push forward with all three of the Instagram changes it is exploring.

Instagram’s Risk

Still, there is considerably more risk this time around: when Facebook pushed forward with the News Feed it was the young upstart moving aside incumbents like MySpace; it’s not as if its userbase was going to go backwards. This case is the opposite: Instagram is clearly aping TikTok, which is the young upstart in the space. It’s possible its users decide that if they must experience TikTok, they might as well go for the genuine thing.

This also highlights why TikTok is a much more serious challenge than Snapchat was: in that case Instagram’s network was the sword used to cut Snapchat off at the knees. I wrote in The Audacity of Copying Well:

For all of Snapchat’s explosive growth, Instagram is still more than double the size, with far more penetration across multiple demographics and international users. Rather than launch a “Stories” app without the network that is the most fundamental feature of any app built on sharing, Facebook is leveraging one of their most valuable assets: Instagram’s 500 million users…Instagram and Facebook are smart enough to know that Instagram Stories are not going to displace Snapchat’s place in its users lives. What Instagram Stories can do, though, is remove the motivation for the hundreds of millions of users on Instagram to even give Snapchat a shot.

Instagram has no such power over TikTok, beyond inertia; in fact, the competitive situation is the opposite: if the goal is not to rank content from your network, but to recommend videos from the best creators anywhere, then it follows that TikTok is in the stronger relative position. Indeed, this is why Mosseri spent so much time talking about “small creators” with Newton:

I think one of the most important things is that we help new talent find an audience. I care a lot about large creators; I would like to do better than we have historically by smaller creators. I think we’ve done pretty well by large creators overall — I’m sure some people will disagree, but in general, that’s what the data suggests. I don’t think we’ve done nearly as well helping new talent break. And I think that’s super important. If we want to be a place where people push culture forward, to help realize the promise of the internet, which was to push power into the hands of more people, I think that we need to get better at that.

There is the old Internet AMA question as to whether you would rather fight a horse-sized duck or 100 duck-sized horses. The analogy here is that in a world of ranking a horse-sized duck that everyone follows is valuable; in a world of recommendations 100 duck-sized horses are much more valuable, and Instagram is willing to sacrifice the former for the latter.

Meta’s Reward

The payoff, though, will not be “power” for these small creators: the implication of entertainment being dictated by recommendations and AI instead of reputation and ranking is that all of the power accrues to the platform doing the recommending. Indeed, this is where the potential reward comes in: this power isn’t only based on the sort of Aggregator dynamics underpinning dominant platforms today, but also absolutely massive amounts of investment in the computing necessary to power the AI that makes all of this possible.

In fact, you can make the case that if Meta survives the TikTok challenge, it will be on its way to the sort of moat enjoyed by the likes of Apple, Amazon, Google, and Microsoft, all of which have real world aspects to their differentiation. There is lots of talk about the $10 billion the company is spending on the Metaverse, but that is R&D; the more important number for this moat is the $30 billion this year in capital expditures, most of which is going to servers for AI. That AI is doing recommendations now, but Meta’s moat will only deepen if Lessin is right about a future where creators can be taken out of the equation entirely, in favor of artificially-generated content.

What is noteworty is that AI content will be an essential part of any sort of Metaverse future; I wrote earlier this year in DALL-E, the Metaverse, and Zero Marginal Content:

What is fascinating about DALL-E is that it points to a future where these three trends can be combined. DALL-E, at the end of the day, is ultimately a product of human-generated content, just like its GPT-3 cousin. The latter, of course, is about text, while DALL-E is about images. Notice, though, that progression from text to images; it follows that machine learning-generated video is next. This will likely take several years, of course; video is a much more difficult problem, and responsive 3D environments more difficult yet, but this is a path the industry has trod before:

  • Game developers pushed the limits on text, then images, then video, then 3D
  • Social media drives content creation costs to zero first on text, then images, then video
  • Machine learning models can now create text and images for zero marginal cost

In the very long run this points to a metaverse vision that is much less deterministic than your typical video game, yet much richer than what is generated on social media. Imagine environments that are not drawn by artists but rather created by AI: this not only increases the possibilities, but crucially, decreases the costs.

These AI challenges, I would add, apply to monetization as well: one of the outcomes of Apple’s App Tracking Transparency changes is that advertising needs to shift from a deterministic model to a probabilistic one; the companies with the most data and the greatest amount of computing resources are going to make that shift more quickly and effectively, and I expect Meta to be top of the list.

None of this matters, though, without engagement. Instagram is following the medium trend to video, and Meta’s resources give it the long-term advantage in AI; the big question is which service users choose to interact with. To put it another way, Facebook’s next two decades are coming into sharper focus than ever; it is how well it navigates the TikTok minefield over the next two years that will determine if that long-term vision becomes a reality.

I wrote a follow-up to this Article in this Daily Update.

Political Chips

Last Friday AMD surpassed Intel in market capitalization:

Intel vs AMD market caps

This was the second time in history this happened — the first was earlier this year — and it may stick this time; AMD, in stark contrast to Intel, had stellar quarterly results. Both stocks are down in the face of a PC slump, but that is much worse news for Intel, given that they make worse chips.

It’s also not a fair comparison: AMD, thirteen years on from its spinout of Global Foundries, only designs chips; Intel both designs and manufactures them. It’s when you include AMD’s current manufacturing partner, TSMC, that Intel’s relative decline becomes particularly apparent:

Intel, AMD, and TSMC market caps

Of course an Intel partisan might argue that this comparison is unfair as well, because TSMC manufactures chips for a whole host of companies beyond AMD. That, though, is precisely Intel’s problem.

Intel’s Stumble

The late Clay Christensen, in his 2004 book Seeing What’s Next, predicted trouble for Intel:

Intel’s well-honed processes — which are almost unassailable competitive strengths in fights for undershot customers hungering for performance increases — might inhibit its ability to fight for customers clamoring for customized products. Its exacting manufacturing process could hamper its ability to deliver customized products. Its sales force could have difficulty adapting to a very different sales cycle. It would have to radically alter its marketing process. The VCE model predicts that operating “fast fabs” will be an attractively profitable point in the value chain in the future. The good news for IDMs such as IBM and Intel is that they own fabs. The bad news is that their fabs aren’t fast. Entrants without legacy processes could quite conceivably develop better proprietary processes that can rapidly deliver custom processors.

This sounds an awful lot like what happened over the ensuing years: one of TSMC’s big advantages is its customer service. Given the fact that the company was built as a pure play foundry it has developed processes and off-the-shelf building blocks that make it easy for partners to build custom chips. This was tremendously valuable, even if the resultant chips were slower than Intel’s.

What Christensen didn’t foresee was that Intel would lose the performance crown; rather, he assumed that performance would cease to be an important differentiator:

If history is any guide, motivated innovators will continue to do the seemingly impossible and find unanticipated ways to extend the life of Moore’s Law. Although there is much consternation that at some point Moore’s Law will run into intractable physical limits, the only thing we can predict for certain is that innovators will be motivated to figure out solutions.

But this does not address whether meeting Moore’s Law will continue to be paramount to success. Everyone always hopes for the emergence of new, unimagined applications. But the weight of history suggests the unimagined often remains just that; ultimately ever more demanding applications will stop appearing or will emerge much more slowly than anticipated. But even if new, high-end applications emerge, rocketing toward the technological frontier almost always leaves customers behind. And it is in those overshot tiers that disruptions take root.

How can we tell if customers are overshot? One signal is customers not using all of a product’s functionality. Can we see this? There are ever-growing populations of users who couldn’t care less about increases in processing power. The vast majority of consumers use their computers for word processing and e-mail. For this majority, high-end microprocessors such as Intel’s Itanium and Pentium 4 and AMD’s Athlon are clearly overkill. Windows XP runs just fine on a Pentium III microprocessor, which is roughly half as fast as the Pentium 4. This is a sign that customers may be overshot.

Obviously Christensen was wrong about a Pentium III being good enough, and not just because web pages suck; rather, the infinite malleability of software really has made it possible to not just create new kinds of applications but to also substantially rework previous analog solutions. Moreover, the need for more performance is actually accelerating with the rise of machine-learning based artificial intelligence.

Intel, despite being a chip manufacturer, understood the importance of software better than anyone. I explained in a Daily Update earlier this year about how Pat Gelsinger, then a graduate student at Stanford, convinced Intel to stick with a CISC architecture design because that gave the company a software advantage; from an oral history at the Computer Museum:

Gelsinger: We had a mutual friend that found out that we had Mr. CISC working as a student of Mr. RISC, the commercial versus the university, the old versus the new, teacher versus student. We had public debates of John and Pat. And Bear Stearns had a big investor conference, a couple thousand people in the audience, and there was a public debate of RISC versus CISC at the time, of John versus Pat.

And I start laying out the dogma of instruction set compatibility, architectural coherence, how software always becomes the determinant of any computer architecture being developed. “Software follows instruction set. Instruction set follows Moore’s Law. And unless you’re 10X better and John, you’re not 10X better, you’re lucky if you’re 2X better, Moore’s Law will just swamp you over time because architectural compatibility becomes so dominant in the adoption of any new computer platform.” And this is when x86– there was no server x86. There’s no clouds at this point in time. And John and I got into this big public debate and it was so popular.

Brock: So the claim wasn’t that the CISC could beat the RISC or keep up to what exactly but the other overwhelming factors would make it the winner in the end.

Gelsinger: Exactly. The argument was based on three fundamental tenets. One is that the gap was dramatically overstated and it wasn’t an asymptotic gap. There was a complexity gap associated with it but you’re going to make it leap up and that the CISC architecture could continue to benefit from Moore’s Law. And that Moore’s Law would continue to carry that forward based on simple ones, number of transistors to attack the CISC problems, frequency of transistors. You’ve got performance for free. And if that gap was in a reasonable frame, you know, if it’s less than 2x, hey, in a Moore’s Law’s term that’s less than a process generation. And the process generation is two years long. So how long does it take you to develop new software, porting operating systems, creating optimized compilers? If it’s less than five years you’re doing extraordinary in building new software systems. So if that gap is less than five years I’m going to crush you John because you cannot possibly establish a new architectural framework for which I’m not going to beat you just based on Moore’s Law, and the natural aggregation of the computer architecture benefits that I can bring in a compatible machine. And, of course, I was right and he was wrong.

That last sentence needs a caveat: Gelsinger was right when it came to computers and servers, but not smartphones. There performance wasn’t free, because manufacturers had to be cognizant of power consumption. More than cognizant, in fact — power usage was the overriding concern. Tony Fadell, who created the iPod and led the development of the first three generations of the iPhone, told me in an interview earlier this year:

You have to have that point of view of that every nanocoulomb is sacred and compatibility doesn’t matter, we’re going to use the best bits, but we’re not going to make sure it has to be the same look and feel. It doesn’t have to have the same principles that is designed for a laptop or a standalone desktop computer, and then bring those down to something that’s smaller form factor, and works within a certain envelope. You have to rethink all the principles. You might use the bits around, and put them together in different ways and use them differently. That’s okay. But your top concept has to be very, very different about what you’re building, why you’re building it, what you’re solving, and the needs of that new environment, which is mobile, and mobile at least for a day or longer for that battery life.

The key phrase there is “compatibility doesn’t matter”; Gelsinger’s argument for CISC over RISC rested on the idea that by the time you remade all of the software created for CISC, Intel would have long since overcome the performance delta between different architectures via its superior manufacturing, which would allow compatibility to trump the competition. Smartphones, though, provided a reason to build up the software layer from scratch, with efficiency, not performance, as the paramount goal.1

All of this still fit in Christensen’s paradigm, I would note: foundries like TSMC and Samsung could accommodate new chip designs that prioritized efficiency over performance, just as Christensen predicted. What he didn’t foresee in 2004 was just how large the smartphone market would be. While there are a host of reasons why TSMC took the performance crown from Intel over the last five years, a major factor is scale: TSMC was making so many chips that it had the money and motivation to invest in Moore’s Law.

The most important decision was shifting to extreme ultraviolet lithography at a time when Intel thought it was much too expensive and difficult to implement; TSMC, backed by Apple’s commitment to buy the best chips it could make, committed to EUV in 2014, and delivered the first EUV-derived chips in 2019 for the iPhone.

Those EUV machines are made by one company — ASML. They’re worth more than Intel too (and Intel is a customer):

Intel, AMD, TSMC, and ASML market caps

The Dutch company, to an even greater degree than TSMC, is the only lithography maker that can afford to invest in the absolute cutting edge.

From Technology to Economics

In 2021’s Internet 3.0 and the Beginning of (Tech) History, I posited that the first era of the Internet was defined by technology, i.e. figuring out what was possible. Much of this technology, including standards like TCP/IP, DNS, HTTP, etc. was developed decades ago; this era culminated in the dot com bubble.

The second era of the Internet was about economics, specifically the unprecedented scale possible in a world of zero distribution costs.

Unlike the assumptions that undergird Internet 1.0, it turned out that the Internet does not disperse economic power but in fact centralizes it. This is what undergirds Aggregation Theory: when services compete without the constraints of geography or marginal costs, dominance is achieved by controlling demand, not supply, and winners take most.

Aggregators like Google and Facebook weren’t the only winners though; the smartphone market was so large that it could sustain a duopoly of two platforms with multi-sided networks of developers, users, and OEMs (in the case of Android; Apple was both OEM and platform provider for iOS). Meanwhile, public cloud providers could provide back-end servers for companies of all types, with scale economics that not only lowered costs and increased flexibility, but which also justified far more investments in R&D that were immediately deployable by said companies.

Chip manufacturing obviously has marginal costs, but the fixed costs are so much larger that the economics are not that dissimilar to software (indeed, this is why the venture capital industry, which originated to support semiconductor startups, so seamlessly transitioned to software); today TSMC et al invest billions of dollars into a single fab that generates millions of chips for decades.

That increase in scale is why a modular value chain ultimately outcompeted Intel’s integrated approach, and it’s why TSMC’s position seems so impregnable: sure, a chip designer like MediaTek might announce a partnership with Intel to maybe produce some lower-end chips at some point in the future, but there is a reason it is not a firm commitment and not for the leading edge. TSMC, for at least the next several years, will make the best chips, and because of that will have the most money to invest in what comes next.

Scale, though, is not the end of the story. Again from Internet 3.0 and the Beginning of (Tech) History:

This is why I suspect that Internet 2.0, despite its economic logic predicated on the technology undergirding the Internet, is not the end-state…After decades of developing the Internet and realizing its economic potential, the entire world is waking up to the reality that the Internet is not simply a new medium, but a new maker of reality…

To the extent the Internet is as meaningful a shift [as the printing press] — and I think it is! — is inversely correlated to how far along we are in the transformation that will follow — which is to say we have only gotten started. And, after last week, the world is awake to the stakes; politics — not economics — will decide, and be decided by, the Internet.

Time will tell if my contention that an increasing number of nations will push back against American Internet hegemony by developing their own less efficient but independent technological capabilities is correct; one could absolutely make the case that the U.S.’s head start is so overwhelming that attempts to undo Silicon Valley centralization won’t pan out anywhere other than China, where U.S. Internet companies have been blocked for a generation.

Chips, though, are very much entering the political era.

Politics and the End-State

Taiwan President Tsai Ing-wen shared, as one does, some pictures from lunch on social media:

Taiwan President Tsai Ing-wen's Facebook post featuring TSMC founder Morris Chang

The man with glasses and the red tie in the first picture is Morris Chang, the founder of TSMC; behind him is Mark Liu, TSMC’s chairman. They were the first guests listed in President Tsai’s write-up of the lunch with House Speaker Nancy Pelosi, which begins:


Taiwan and the United States not only share the values ​​of democracy, freedom and human rights, but also continue to work together on economic development and democratic supply chains.

That sentence captures why Taiwan looms so large, not only on the occasion of Pelosi’s visit, but to world events for years to come. Yes, the United States supports Taiwan because of democracy, freedom and human rights; the biggest reason why that support may one day entail aircraft carriers is because of chips and TSMC. I wrote two years ago in Chips and Geopolitics:

The international status of Taiwan is, as they say, complicated. So, for that matter, are U.S.-China relations. These two things can and do overlap to make entirely new, even more complicated complications.

Geography is much more straightforward:

A map of the Pacific

Taiwan, you will note, is just off the coast of China. South Korea, home to Samsung, which also makes the highest end chips, although mostly for its own use, is just as close. The United States, meanwhile, is on the other side of the Pacific Ocean. There are advanced foundries in Oregon, New Mexico, and Arizona, but they are operated by Intel, and Intel makes chips for its own integrated use cases only.

The reason this matters is because chips matter for many use cases outside of PCs and servers — Intel’s focus — which is to say that TSMC matters. Nearly every piece of equipment these days, military or otherwise, has a processor inside. Some of these don’t require particularly high performance, and can be manufactured by fabs built years ago all over the U.S. and across the world; others, though, require the most advanced processes, which means they must be manufactured in Taiwan by TSMC.

This is a big problem if you are a U.S. military planner. Your job is not to figure out if there will ever be a war between the U.S. and China, but to plan for an eventuality you hope never occurs. And in that planning the fact that TSMC’s foundries — and Samsung’s — are within easy reach of Chinese missiles is a major issue.

China, meanwhile, is investing heavily in catching up, although Semiconductor Manufacturing International Corporation (SMIC), its Shanghai-based champion, only just started manufacturing on a 14nm process, years after TSMC, Samsung, and Intel. In the long run, though, the U.S. faced a scenario where China had its own chip supplier, even as it threatened the U.S.’s chip supply chain.

This reality is why I ultimately came down in support of the CHIPS Act, which passed Congress last week. I wrote in a Daily Update:

This is why Intel’s shift to being not simply an integrated device manufacturer but also a foundry is important: yes, it’s the right thing to do for Intel’s business, but it’s also good for the West if Intel can pull it off. That, by extension, is why I’m fine with the CHIPS bill favoring Intel…AMD, Qualcomm, Nvidia, et al, are doing just fine under the current system; they are drivers and beneficiaries of TSMC’s dominance in particular. The system is working! Which, to the point above, is precisely why Intel being helped disproportionately is in fact not a flaw but a feature: the goal should be to counteract the fundamental forces pushing manufacturing to geopolitically risky regions, and Intel is the only real conduit available to do that.

Time will tell if the CHIPS Act achieves its intended goals; the final version did, as I hoped, explicitly limit investment by recipients in China, which is already leading chip makers to rethink their investments. That this is warping the chip market is, in fact, the point: the structure of technology drives inexorably towards the most economically efficient outcomes, but the ultimate end state will increasingly be a matter of politics.

I wrote a follow-up to this Article in this Daily Update.

  1. As an example of how efficiency trumped performance, the first iPhone’s processor was actually underclocked — better battery life was more of a priority than faster performance. 

Big Ten Blame

Penn State has always been a usurper, at least to me.

In 1984 the Supreme Court ruled that the NCAA’s attempt to control individual universities’ college football TV rights was an illegal restraint on trade; while the lawsuit was instigated by the University of Oklahoma, it was conferences that were the biggest beneficiary, as it simply made more sense to negotiate TV rights as a collective of similarly situated schools. This was a problem for independent Penn State; its traditional rivals like Pitt, Syracuse, and Boston College joined the basketball-focused Big East, leaving the football power to make overtures to the Big Ten.

The 1990 announcement that Penn State was joining the conference was a controversial one within the Big Ten itself. A fair number of the conference’s athletic directors were opposed to the move, and most of the coaches (it is university presidents that make the decision, and even there Penn State only received the minimum 7 votes in favor); I was a 10 year-old sports fan of a then-decrepit football team in Wisconsin that had nothing going for it other than Big Ten pride, and resented the idea that Penn State was going to come in and potentially dominate the conference.

It all feels quite quaint here in 2022, and not just because Wisconsin has had more football success than Penn State; the Big Ten added Nebraska in 2011, and Maryland and Rutgers in 2014, in both cases setting off seismic shifts in the college landscape. Both expansions made sense on the edges, both figuratively and literally: Nebraska was a traditional football power neighboring Iowa that, more importantly, gave the conference the 12 teams necessary to stage a lucrative conference championship game. Maryland and Rutgers bordered Pennsylvania — Penn State was a well-established member of the Big Ten by this point, in practice if not in my mind — and, more importantly, brought the Washington D.C. and New York City markets to the Big Ten’s groundbreaking cable TV network.

It is the latest expansion announcement, though, that blows apart the entire concept of a regional conference founded on geographic rivalries: UCLA and USC will join the Big Ten in 2024. They are not, needless to say, in a Big Ten border state:

That, though, very much fits the reality of 2022: geography doesn’t matter, but attention does, and after the SEC grabbed Texas and Oklahoma last year, the two Los Angeles universities represented the two biggest programs not yet in the two biggest conferences. As for the timing, it’s all about TV: the Big Ten is in the middle of negotiating new rights packages this summer, while the Pac-12’s rights package ends in 2024. One thing is for sure: everyone in college sports, particularly those who, like 10-year-old me, value tradition, know exactly who to blame:

This appears to be true as far as the mechanics of the Big Ten’s expansion: Fox reportedly instigated the UCLA and USC talks with the Big Ten. Figuring out why it was Fox, though — and ESPN with the SEC — exculpates both networks from ultimate responsibility.

A Brief History of TV

TV’s origins are, unsurprisingly, in radio: thanks to the magic of over-the-air broadcasting a radio station could deliver audio to anyone in its geographic area with a compatible listening device. Originally all of said audio was generated at the radio station, but given that most people wanted to listen to similar things, it made sense to link stations together and broadcast the same content all at once. In 1928 NBC became the first coast-to-coast radio network, linking together radio stations with phone lines; those stations weren’t owned by NBC, but were rather affiliates: local owners would actually operate the stations and sell local ads, and pay a fee to NBC for content that, because it was funded by stations across the country, was far more compelling than anything that could be produced locally.

TV followed the same path: thanks to the increased cost of producing video relative to audio, the economic logic of centralized content production was even more compelling (as were the proceeds from selling ads nationally). The content was more compelling as well, leading to further innovation in distribution, specifically the advent of cable that I wrote about earlier this year. That new distribution led to further innovation in content: new networks were created specifically for cable, even though cable had originally been created to help people receive over-the-air broadcasts.

In fact, new cable networks were so compelling that local broadcast stations (particularly those unaffiliated with the national networks) were worried about losing carriage, leading the FCC to institute “must-carry” rules that compelled cable networks to carry local broadcast networks for free. However, in the 1980s must-carry rules were ruled by a federal Appeals Court to be an infringement on the cable carriers’ First Amendment rights, threatening local broadcast networks that depended on the rules for access to an increasingly large percentage of homes.

This ruling was an impetus for the Cable Television Consumer Protection and Competition Act of 1992, which among other provisions, gave local broadcast stations the choice between must-carry status or charging retransmission fees; if the station chose the latter, then they forewent the former. This led to a clear bifurcation in broadcast channels: cheap and mostly local programming chose must-carry, while stations with the most desirable programming — which per the aforementioned point, were affiliated with national networks — chose the latter. Those national networks took notice: a significant part of the increase in cable fees over the last thirty years has been driven by national networks increasing the fees they charge affiliates for programming; those affiliates recoup those fees by increasing their retransmission fees (cable companies, for their part, continue to break out these fees on bills — my “Broadcast TV Surcharge” in Wisconsin is $21/month).

Local stations originally pushed back against this shift, de-affiliating with networks that pushed too hard. The problem, though, came back to content: networks had everything from popular sitcoms and dramas to national news to late night talk shows. The most important bit of content, though, was sports. Moreover, the importance of sports has only increased as those other content offerings have been unbundled by the Internet: streaming services have sitcoms and dramas, websites have all of the news you could ever want to consume, and social media provides all kinds of comedy and commentary; the one exception to Internet disruption are live games between teams you care about.

Fox’s Contrarian Bet

The importance of sports to ESPN is self-explanatory: the entire point of the network is to show games, particularly as its SportsCenter and talk show franchises have suffered from competition with the Internet. ESPN’s parent company, Disney, has jumped into this competition with both feet, launching Disney+ and taking a controlling stake in Hulu.

That controlling stake came from a deal that Disney announced in 2017: the company acquired the majority of 21st Century Fox, including: its film and television studios, most of its cable TV networks (including FX), a controlling stake in National Geographic, Star India, and the aforementioned stake in Hulu. Disney’s rationale was that it needed to beef up its content offerings to compete in streaming, which was clearly the future; traditional TV was, by implication, the past.

That left a newly spun-out Fox Corporation, which included the Fox broadcast network, Fox News, and Fox Sports (including the Fox Sports cable channels and Fox’s share in the Big Ten network); the nature of these channels signaled a completely different strategy than streaming. I wrote at the time in a Daily Update:

In that last sentence I actually put forth two distinct strategies: selling direct to consumers, and charging distributors significantly more. Both do depend on having differentiated content, which was the point of that Daily Update, but the similarities end there. To explain what I mean, this deal actually offers two great examples: if it goes through, that means Fox is pursuing the second strategy, and Disney the first.

Start with Fox: its news and sports divisions — particularly the former — are highly differentiated. That gives Fox significant pricing power with shrinking-but-still-very-large TV distributors. Moreover, given that both news and sports are heavily biased towards live viewing, they are also a good fit for advertising, which again, matches up with traditional TV distribution. What Fox would accomplish with this deal, then, is shedding a huge amount of that detritus I mentioned earlier: sure, more was better when there was only one distributor, but now that there is competition for viewer’s attention, filler is a drag on the content that actually gives negotiating leverage. Fox could come out of this deal with the same pricing power it has today but a vastly streamlined corporate structure and cost basis.

It’s a bet that has paid off: while Disney, like other streaming companies, enjoyed a huge run-up during the pandemic, it is Fox that today has the superior returns in its two years as an independent company:

Fox's stock price relative to Disney since the 21st Century Fox acquisition

Still, I think my original analysis was incomplete; Fox isn’t simply wringing more money out of a dying business model with a leaner corporate structure. Rather, it is driving the paid-TV business model to its logical endpoint: nothing but sports and news. That doesn’t mean, though, they are the biggest winner.

Sports Concentration

The foundation of Fox’s offering is the NFL, but it is an expensive offering: $2.025 billion per year for Sunday afternoon regular season games, playoffs, and the Super Bowl every four years. The NFL understands its position in the sports landscape very well — it has a monopoly on the professional version of America’s favorite sport — and it prices its rights accordingly, even as it is careful to spread its games across multiple networks:1

The NFL monopoly

College football is America’s second favorite sport; it has also traditionally been a much more profitable one for the networks. CBS, for example, currently pays the SEC $55 million per year to broadcast the conference’s top games, which average over 6 million viewers per telecast; that is about a third of what CBS averages for NFL games, but at a fraction of the cost for rights. The relative cheapness of that deal is explained by the fact it was negotiated a decade ago, when the college football landscape was considerably more diffused:2

College diffusion

The SEC’s new deal looks a lot different: ESPN has exclusive rights to SEC football3 for $300 million per year. That increase is driven by the SEC’s dominance of college football, and corresponding national interest; that interest will be that much greater thanks to the addition of Texas and Oklahoma (which will almost certainly lead to an increase in rights fees). In short, sports are the biggest driver of pay-TV, which means it is essential to have the sports the most people want to watch; the SEC figures prominently in that regard.

Still, it is the Big Ten, based in the sports-obsessed Midwest, and filled with massive public universities churning out interested alumni who live all over the country, that is the most attractive of all; even before this expansion the conference was rumored to be seeking a deal for $1.1 billion/year. Add in the Los Angeles market and UCLA and USC fan bases and that number could end up even higher.

Blame Games

Put all of these pieces together, and the question of who exactly is responsible for college football’s conference upheavel gets a bit more complicated:

  • Local TV stations charge ever higher retransmission fees to pay-TV operators because they have compelling content that subscribers demand.
  • Networks charge ever higher affiliate fees to TV stations for that compelling content, extracting most (if not all) of those retransmission fees.
  • The most compelling content is sports, especially as alternative content loses out to the Internet, and the most popular sport (the NFL) is governed by a single entity, allowing it to extract the greatest fees.

All of this extraction is a function of relative bargaining power that is ultimately derived from what fans want to see:

The NFL has the most bargaining power

Given this, the logic of the Big Ten’s expansion into California is obvious: the more of an audience that the Big Ten can command, the more of the money flowing through that value chain it can extract. Sure, it’s not quite to the level of the NFL, but it’s the next closest thing. This is also the downside to Fox’s bet on live: while the company owns the content it produces on channels like Fox News, it has to buy sports rights, and it is the Big Ten that is determined to take its share, even if that means an expansion that otherwise makes no sense at all.

In other words, I think the tweet above has it backwards: Fox and ESPN are not “grandmasters calling the shots behind the scenes”; they are essential but ultimately replaceable parts in the movement of money from consumers to the entities that provide the content those consumers want.

The Big Ten is accruing NFL-like bargaining power

Still, the tweet is instructive: perhaps the most essential role Fox and ESPN play for universities is taking the blame as the latter make more money than ever.

I wrote a follow-up to this Article in this Daily Update.

  1. The NFL is also very careful about cultivating local fans: the league has always insisted that all games are available over-the-air in local markets, whether they are on broadcast TV or not 

  2. The SEC also sold secondary rights to ESPN 

  3. And basketball, but that doesn’t drive rights fees to nearly the same extent 

Spotify, Netflix, and Aggregation

When Spotify filed for its direct listing in 2018, it was popular to compare the streaming music service to Netflix, the streaming video service; after all, both were quickly growing subscription-based services that gave consumers media on demand. This comparison was intended as a bullish one for Spotify, given that Netflix’s stock had increased by 1,113% over the preceding five years:

Netflix's stock from 2013-2018

The problem with the comparison is that Spotify clearly was a different kind of business than Netflix, thanks to its very different relationship to its content providers; whereas Netflix had always acquired content on a wholesale basis — first through licensing deals with content owners, and later by making its own content — Spotify licensed content on a revenue share basis. This latter point is why I was skeptical of Spotify’s profit-making potential; I wrote at the time in Lessons From Spotify, in a section entitled Spotify’s Missing Profit Potential:

That, though, is precisely the problem: Spotify’s margins are completely at the mercy of the record labels, and even after the rate change, the company is not just unprofitable, its losses are growing, at least in absolute euro terms:

Spotify Gross and Net Profit

Moreover, it seems highly unlikely Spotify’s Cost of Revenue will improve much in the short-term: those record deals are locked in until at least next year, and they include “most-favored nation” provisions, which means that Spotify has to get Universal Music Group, Sony Music Entertainment, Warner Music Group, and Merlin (the representative for many independent labels), which own 85% of the music on Spotify as measured by streams, to all agree to reduce rates collectively. Making matters worse, the U.S. Copyright Royalty Board just increased the amount to be paid out to songwriters; Spotify said the change isn’t material, but it certainly isn’t in the right direction either.

I compared this (unfavorably) to Netflix in a follow-up:

Netflix has licensed content, not agreed-to royalty agreements. That means that Netflix’s costs are fixed, which is exactly the sort of cost structure you want if you are a growing Internet company. Spotify, on the other hand, pays the labels according to a formula that has revenue as the variable, which means that Spotify’s marginal costs rise in-line with their top-line revenue.

Both companies proceeded to do quite well in the markets over the next three-and-a-half years, with a decided edge to Netflix; however, the last six months undid all of those gains, with Netflix getting hit harder than Spotify:

Netflix and Spotify stock since 2018

Setting aside any analysis of absolute value, I think the relative trend is justified: I still maintain, as I did in 2018, that Spotify and Netflix are fundamentally different businesses because of their relationship to content; now, though, I think that Spotify’s position is preferable to Netflix, and has more long-term upside. Moreover, I don’t think this is a new development: I was wrong in 2018 to prefer Netflix to Spotify; worse, I should have already known why.

Categorizing Aggregators

In my original article about Aggregation Theory I wrote:

The value chain for any given consumer market is divided into three parts: suppliers, distributors, and consumers/users. The best way to make outsize profits in any of these markets is to either gain a horizontal monopoly in one of the three parts or to integrate two of the parts such that you have a competitive advantage in delivering a vertical solution. In the pre-Internet era the latter depended on controlling distribution…

The fundamental disruption of the Internet has been to turn this dynamic on its head. First, the Internet has made distribution (of digital goods) free, neutralizing the advantage that pre-Internet distributors leveraged to integrate with suppliers. Secondly, the Internet has made transaction costs zero, making it viable for a distributor to integrate forward with end users/consumers at scale.

Aggregation Theory

This has fundamentally changed the plane of competition: no longer do distributors compete based upon exclusive supplier relationships, with consumers/users an afterthought. Instead, suppliers can be commoditized leaving consumers/users as a first order priority. By extension, this means that the most important factor determining success is the user experience: the best distributors/aggregators/market-makers win by providing the best experience, which earns them the most consumers/users, which attracts the most suppliers, which enhances the user experience in a virtuous cycle.

Fast forward to 2017, where in Defining Aggregators I sought to provide a more specific definition of an Aggregator; specifically, Aggregators had:

  • A direct relationship with users
  • Zero marginal costs for serving users
  • Demand-driven multi-sided networks with decreasing acquisition costs

The rest of the post was devoted to building a taxonomy of Aggregators based on their relationships to suppliers:

  • Level 1 Aggregators paid for supply, but had superior buying power due to their user base; Netflix was here.
  • Level 2 Aggregators incurred marginal transaction costs for adding supply; “real-world” companies like Uber and Airbnb were here.
  • Level 3 Aggregators had zero supply costs; Google and Facebook were here.

Notice the tension between the two articles, specifically this sentence in Aggregation Theory: “no longer do distributors compete based upon exclusive supplier relationships”. Had I kept that sentence in mind then I ought to have concluded that Level 1 and Level 2 were not Aggregators at all: sure, companies in these categories could scale on the demand side, but they were going to hit a wall on the supply side. Indeed, this has been a problem for Uber in particular: the company has never been able to escape from the need to compete for supply, which is another way of saying the company has never been able to make any money.

Netflix has a similar problem: the company’s content investments have failed to provide the evergreen lift in customer acquisition I anticipated; despite having more original content than ever, Netflix has hit the wall in terms of subscriber growth, particularly in developed countries where it can charge the highest prices. While some of these challenges were forseeable — I predicted that Netflix would have a rough stretch where all of its once-content providers tried their hand at streaming1 — it does seem likely that Netflix is going to struggle to get significantly more leverage on its content costs than it has to date, as that is the only way to not just acquire but also keep its subscribers.

In short, Defining Aggregator was focused on the demand-side in its definition and the supply-side in its taxonomy; my original
Article — and, in retrospect, the better one — did the opposite. The only way you can truly control demand — the tell-tale sign of an Aggregator — is to have fully commoditized and infinitely scalable supply; streaming video fails on the former, and ride-sharing on the latter.

Spotify and Commoditized Supply

With that in mind, go back to the reason I was skeptical of Spotify, and those revenue-sharing deals with record labels. In 2017’s The Great Unbundling I noted that the music industry, much to the surprise of anyone who observed the Napster era, had ended up in pretty good shape:

While piracy drove the music labels into the arms of Apple, which unbundled the album into the song, streaming has rewarded the integration of back catalogs and new music with bundle economics: more and more users are willing to pay $10/month for access to everything, significantly increasing the average revenue per customer. The result is an industry that looks remarkably similar to the pre-Internet era:

A drawing of The New Music Media Industry Model

Notice how little power Spotify and Apple Music have; neither has a sufficient user base to attract suppliers (artists) based on pure economics, in part because they don’t have access to back catalogs. Unlike newspapers, music labels built an integration that transcends distribution.

It’s easy to see why this would be a bad thing for Spotify; on the flipside, you can see why many of the most optimistic assumptions about Spotify’s upside rested on the company somehow eliminating the labels completely, even though the fact that new music immediately becomes back catalog music means that the label’s relative position in the music value chain is quite stable. This is why I’ve always been skeptical about the possibility of Spotify displacing the labels entirely, and the short-lived exclusive wars confirmed that point of view: it’s better for artists and labels for all music to be available everywhere.

My contention — and the key insight undergirding this Article — is that this is better for Spotify, as well. Go back to the point about commoditization: an input does not need to be free to be commoditized. Sure, the marginal cost of streaming music — which is nothing but bits — is zero; it is a credit to the music labels’ new-music-to-back-catalog flywheel that they are able to charge for access to these bits. At the same time, the bits are available to anyone for roughly the same price: that is why not just Spotify and Apple, but also YouTube, Amazon, Tidal, Deezer, etc. all have roughly the same catalog for roughly the same price. Streaming music isn’t free, but it is an infinitely available non-exclusive commodity.

Look again at the contrast to the other companies I highlighted above: there are a limited number of potential ride-share drivers, which means that Uber has to compete with Lyft with a never-ending set of driver incentives that make profitability difficult if not impossible to achieve; Netflix has some of the content — not all of it — and it has to bid against competitors to get more of it.

This does, to be clear, make it easier to achieve a direct profit: as I noted above, because Netflix pays a fixed cost for content it can earn a surplus without having to pay anything extra to the content producer, whereas Spotify can only eke out profitability from its subscribers by reducing its operational costs. The real opportunity for an Aggregator, though, is building a business model that is independent of supply and instead predicated on owning demand. Here Spotify is better placed than Netflix — although the latter is finally making moves in that direction.

Spotify’s Investor Day

Earlier this month Spotify had an investor day (which I covered in an Update last week); Charlie Hellman, vice president and head of music product, presented a textbook case as to why Spotify is an Aggregator for music, with the business model to match. Hellman started by emphasizing Spotify’s role in new music discovery:

It’s important to remember that first and foremost Spotify is a music company. All of our music team’s strategies ladder up to two primary goals: making a unique and superior music experience for fans, and creating a more open and valuable ecosystem for artists. These two goals really complement one another, which is clear to see when you look at the playlisting ecosystem we’ve spent the last decadee defining and perfecting. Whatever your mood, your style, whatever the occassion, Spotify has something for you, and as Gustav mentioned, Spotify drives around 22 billion discoveries a month. On top of that, 1/3 of all new artist discoveries happen on personalized algorithmic playlists. Listeners love this exposure to new music, as well as the personalized touch. Discovery is our bread and butter, and it’s driving a level of engagement that no streaming service can claim.

This is the number one characteristic of an Aggregator: in a world of scarcity distribution was the most valuable; in a world of abundance it is discovery that matters most. To that end, Hellman emphasized that the world of digital music was one of ever-increasing abundance, thanks in part to Spotify:

The music industry is changing fast. There have never been fewer barriers to entry, and that’s enabling more and more talented artists to be discovered…but with this reduction in barriers comes an increase in the number of artists seeking success. We’re in the midst of an explosion in creativity where tens of thousands of songs are uploaded each day, and that rate of daily uploads has doubled in the last two years. In this rapidly growing landscape, artists need an evolving toolkit that works for the millions who will make up tomorrow’s music industry. One that mirrors their creativity and ambition by offering speed and scale.

Hellman highlighted free tools Spotify offers, including analytics, custom art and videos, and the ability to pitch your song for a playlist. What is far more important from a business perspective, though, is the fact you can pay for promotion.

In addition to these free tools, we’ve also invested in building the most performant and effective commercial tools for promotion in the streaming era. Because there’s so much being added to Spotify every day, artists need tools that will help them stand out, now more than ever, and we’re uniqely positioned to deliver effective promotion for artists for a few reasons. First, unlike like, say, social media marketing, our promotion tools reach people that have already actively made a decision to open Spotify and listen to music. It’s contextual. Plus, our ability to target listeners based on their listening activity, their taste, is second to none. And further, we have the unique ability to actually report back how many people listened to or saved the music as a result. Our ability to deliver the best promotional opportunities for artists presents a tremendous opportunity.

This is an opportunity that is paying off; CFO Paul Vogel explained how Spotify’s heavy investments in podcasting (more on this in a moment) were obscuring major improvements in the financial performance of the company’s music business:

Our music business has been a real source of strength, driving strong revenue growth, and strong margin expansion. This may not be immediately evident in our consolidated results, but make no mistake, we have delivered against the expectations and framework we rolled out at the time of our direct listing. Isolating just the performance of our music operations, you’d see that our music revenues, which consist of premium subscriptions, ad-supported music, our marketplace suite of artist tools, and strategic licensing, great at a 24% compound annual growth rate, in-line with our expectations on an FX-neutral basis, and importantly, our music gross margins have increased over the same time frame, reaching 28.3% in 2021. This is approximately 150-basis points higher than our total 2021 consolidated margin of 26.8%…Looking at our progression in another way, since 2018, the last year before our major podcast investment, our music margins have expanded on average by approximately 75 basis points per year. At our last investor day we told you to expect gross margins in the 30-35% range over the long-term. This was, and still remains, the goal for our music operations. As you can see from these numbers, we are clearly on our way.

Let me unpack how we have expanded our music gross margins. At the beginning of 2018, we announced the development of our marketplace business — all of the tools and services that Charlie described earlier. Our thesis back then was that by providing increased value to artists, creators, and labels, they would see material benefit, and so would Spotify, and that is exactly what we are seeing today. We’ve long maintained that our success is not solely tied to renegotiating new headline rates. It’s about our ability to innovate, right along with our partners, to grow a business that benefits both artists and Spotify, and that’s what we’ve done with Marketplace. In 2018 our Marketplace contribution to gross profit was only $20 million. In 2021 it grew to $160 million, 8x the size in just four years. We expect that number to increase another 30% or more in 2022. We see tremendous upside in Marketplace, and anticipate that it’s financial contribution will continue to grow at a healthy double-digit rate in the years ahead. Marketplace is the quintessential example of our approach to capital allocation. There was a significant up-front cost to build-and-launch these offerings, but we saw compelling data which gave us the confidence to double-down and invest aggressively against our goals. It may have taken time to build up momentum, but our patience and conviction has paid off, and we are seeing material benefit from our investment.

Notice how this business — in its mechanics, if not its financial numbers — looks more like a Google or a Facebook than a Netflix: Spotify isn’t earning money by making margin on its content spend; rather, it is seeking to enable more content than ever, confident that it controls the best means to surface the content users want. Those means can then be sold to the highest bidder, with all of the margin going to Spotify. Spotify calls this promotion — it certainly looks a lot like the old radio model of pay-to-play — but that’s really just another word for advertising. Moreover, this isn’t Spotify’s only advertising business: the company has long been building an ad-supported music business, and is now heavily investing in doing the same for podcasting (which has always clearly been an aggregation strategy).

Acting Like an Aggregator

Being an Aggregator is not the only way to make money; indeed, it is amongst the most difficult. It is far more straightforward to make a differentiated product and charge for it; that is also the only possibility for most products and companies. What makes Aggregators unique is that they are best served by doing the opposite of what is optimal for traditional businessess:

  • Instead of having differentiated content, Aggregators want commoditized content, and more of it.
  • Instead of increasing margin on their users, Aggregators want to reduce it, ideally to zero, or at least to the same level as their competitors (as in the case of Spotify).
  • Instead of introducing friction in the market, the better to lock-in users, Aggregators want to decrease friction, confident the gravitational pull of their user experience will, all things being equal, draw in more users than their competitors, increasing their attractivness to not just suppliers but also advertisers (who, in the case of Spotify’s music business, may be the same entities).

Spotify has, for the most part, acted like an Aggregator: the company has fought exclusives in the music business, kept its subscription prices as low as possible, and in the case of podcasts ensured its Anchor platform supports all podcast players.2 Netflix has not: the company has invested heavily in its own content, steadily increases its prices, and is now embarking on a campaign to make sure its best customers pay more for sharing access.

What is noteworthy are the exceptions: on Spotify’s side the most obvious one are its podcast exclusives. The potential payoff in terms of taking podcast share is obvious; being an Aggregator means being the biggest player in a particular space, and by all accounts Spotify’s strategy has delivered exactly that. There are, though, risks in the approach: exclusive content creators are liable to become ever more expensive over time as they seek to seize their share of the value they create. Moreover, they risk triggering a response by competitors making their own exclusive deals. To put it in the terms discussed above, exclusive content is de-commoditized content, and that is bad for Aggregators (that noted, Spotify’s biggest long-term competitor is not Apple but YouTube, a formidable Aggregator in its own right; maybe exclusives aren’t the worst idea).

Netflix, meanwhile, is finally building — or contracting to build — an ad business. While this still seems like a reaction to slowing growth instead of a considered strategy, what is important is the shift from selling exclusive supply to selling exclusive demand: the latter is far more scalable and defensible, although the transition will be very dificult (and may not fully pay off if Netflix isn’t willing to invest on its own).

The takeaway I am most interested in, though, is a selfish one: I had an essential part of Aggregation Theory — the commoditzation of supply — right originally, only to forget that insight and make several bad calls along the way. In the case of Netflix and Spotify I was right in observing that they were different businesses; my mistake was mislabeling which had more potential as an Aggregator.

I wrote a follow-up to this Article in this Daily Update.

  1. What I got wrong about my prediction was the timing, thanks to COVID

  2. Spotify has also ensured that independent podcasters like Stratechery can deliver their content on Spotify

Data and Definitions

Last week the German Bundeskartellamt (“Federal Cartel Office”) announced in a press release:

The Bundeskartellamt has initiated a proceeding against the technology company Apple to review under competition law its tracking rules and the App Tracking Transparency Framework. In particular, Apple’s rules have raised the initial suspicion of self-preferencing and/or impediment of other companies, which will be examined in the proceeding.

The press release quoted Andreas Mundt, the President of the Bundeskartellamt, who stated:

We welcome business models which use data carefully and give users choice as to how their data are used. A corporation like Apple which is in a position to unilaterally set rules for its ecosystem, in particular for its app store, should make pro-competitive rules. We have reason to doubt that this is the case when we see that Apple’s rules apply to third parties, but not to Apple itself. This would allow Apple to give preference to its own offers or impede other companies.

The press release continues:

Already based on the applicable legislation, and irrespective of Apple’s App Tracking Transparency Framework, all apps have to ask for their users’ consent to track their data. Apple’s rules now also make tracking conditional on the users’ consent to the use and combination of their data in a dialogue popping up when an app not made by Apple is started for the first time, in addition to the already existing dialogue requesting such consent from users. The Identifier for Advertisers, classified as tracking, which is important to the advertising industry and made available by Apple to identify devices, is also subject to this new rule. These rules apparently do not affect Apple when using and combining user data from its own ecosystem. While users can also restrict Apple from using their data for personalised advertising, the Bundeskartellamt’s preliminary findings indicate that Apple is not subject to the new and additional rules of the App Tracking Transparency Framework.

John Gruber disagrees at Daring Fireball:

I think this is a profound misunderstanding of what Apple is doing, and how Apple is benefiting indirectly from ATT. Apple’s privacy and tracking rules do apply to itself. Apple’s own apps don’t show the track-you-across-other-apps permission alert not because Apple has exempted itself but because Apple’s own apps don’t track you across other apps. Apple’s own apps show privacy report cards in the App Store, too…

If you want to argue that Apple engaged in this entire ATT endeavor to benefit its own Search Ads platform, that’s plausible too. But if Apple actually cared more about maximizing Search Ads revenue than it does user privacy, wouldn’t they have just engaged in actual user tracking? The Bundeskartellamt perspective here completely disregards the idea that surveillance advertising is inherently unethical and Apple has studiously avoided it for that reason, despite the fact that it has proven to be wildly profitable for large platforms.

This strikes me as a situation where Gruber — my co-host for Dithering — is right on the details, even as the Bundeskartellamt is right on the overall thrust of the argument. The distinction comes down to definitions.

It’s striking in retrospect how little time Apple spent publicly discussing its App Tracking Transparency (ATT) initiative — a mere 20 seconds at WWDC 2020, wedged in between updates about camera-in-use indicators and privacy labels in the App Store:

Next, let’s talk about tracking. Safari’s Intelligent Tracking Prevention has been really successful on the web, and this year, we wanted to help you with tracking in apps. We believe tracking should always be transparent, and under your control, so moving forward, App Store policy will require apps to ask before tracking you across apps and websites owned by other companies.

These 20 seconds led, 19 months later, to Meta announcing a $10 billion revenue shortfall, the largest but by no means only significant retrenchment in the online advertising space. Not everyone was hurt, though: Google and Amazon, in particular, have seen their share of digital advertising increase, and, as Gruber admitted, Apple has benefited as well; the Financial Times reported last fall:

Apple’s advertising business has more than tripled its market share in the six months after it introduced privacy changes to iPhones that obstructed rivals, including Facebook, from targeting ads at consumers. The in-house business, called Search Ads, offers sponsored slots in the App Store that appear above search results. Users who search for “Snapchat”, for example, might see TikTok as the first result on their screen. Branch, which measures the effectiveness of mobile marketing, said Apple’s in-house business is now responsible for 58 per cent of all iPhone app downloads that result from clicking on an advert. A year ago, its share was 17 per cent.

These numbers, derived as they are from app analytics companies, are certainly fuzzy, but they are the best we have given that Apple doesn’t break out revenue numbers for its advertising business; they are also from last fall, before ATT really began to bite. They also exclude the revenue Apple earns from Google for being the default search engine for Safari, and while Google’s earnings indicate YouTube has suffered from ATT, search has more than made up for it.

I explained in depth why these big companies have benefitted from ATT in February’s Digital Advertising in 2022; I wrote in the context of Amazon specifically:

Amazon also has data on its users, and it is free to collect as much of it as it likes, and leverage it however it wishes when it comes to selling ads. This is because all of Amazon’s data collection, ad targeting, and conversion happen on the same platform — Amazon.com, or the Amazon app. ATT only restricts third party data sharing, which means it doesn’t affect Amazon at all…

That is not to say that ATT didn’t have an effect on Amazon: I noted above that Snap’s business did better than expected in part because its business wasn’t dominated by direct response advertising to the extent that Facebook’s was, and that more advertising money flowed into other types of advertising. This almost certainly made a difference for Amazon as well: one of the most affected areas of Facebook advertising was e-commerce; if you are an e-commerce seller whose Shopify store powered-by Facebook ads was suddenly under-performing thanks to ATT, then the natural response is to shift products and advertising spend to Amazon.

This is where definitions matter. The opening paragraph of Apple’s Advertising & Policy page, housed under the “apple.com/legal” directory, states:

Ads that are delivered by Apple’s advertising platform may appear on the App Store, Apple News, and Stocks. Apple’s advertising platform does not track you, meaning that it does not link user or device data collected from our apps with user or device data collected from third parties for targeted advertising or advertising measurement purposes, and does not share user or device data with data brokers.

I note the URL path for a reason: the second sentence of this paragraph has multiple carefully selected words — and those word choices not only impact the first sentence, but may, soon enough, lead to its expansion. Specifically:


Apple’s advertising platform does not track you, meaning that it does not link user or device data collected from our apps with user or device data collected from third parties for targeted advertising or advertising measurement purposes, and does not share user or device data with data brokers.

“Tracking” is not a neutral term! My strong suspicion — confirmed by anecdata — is that a lot of the most ardent defenders of Apple’s ATT policy are against targeted advertising as a category, which is to say they are against companies collecting data and using that data to target ads. For these folks I would imagine tracking means exactly that: the collection and use of data to target ads. That certainly seems to align with the definition of “track” from macOS’s built-in dictionary: “Follow the course or trail of (someone or something), typically in order to find them or note their location at various points”.

However, this is not Apple’s definition: tracking is only when data Apple collects is linked with data from third parties for targeted advertising or measurement, or when data is shared/sold to data brokers. In other words, data that Apple collects and uses for advertising is, according to Apple, not tracking; the privacy policy helpfully lays out exactly what that data is (thanks lawyers!):

We create segments, which are groups of people who share similar characteristics, and use these groups for delivering targeted ads. Information about you may be used to determine which segments you’re assigned to, and thus, which ads you receive. To protect your privacy, targeted ads are delivered only if more than 5,000 people meet the targeting criteria.

We may use information such as the following to assign you to segments:

  • Account Information: Your name, address, age, gender, and devices registered to your Apple ID account. Information such as your first name in your Apple ID registration page or salutation in your Apple ID account may be used to derive your gender. You can update your account information on the Apple ID website.

  • Downloads, Purchases & Subscriptions: The music, movies, books, TV shows, and apps you download, as well as any in-app purchases and subscriptions. We don’t allow targeting based on downloads of a specific app or purchases within a specific app (including subscriptions) from the App Store, unless the targeting is done by that app’s developer.

  • Apple News and Stocks: The topics and categories of the stories you read and the publications you follow, subscribe to, or turn on notifications from.

  • Advertising: Your interactions with ads delivered by Apple’s advertising platform.

When selecting which ad to display from multiple ads for which you are eligible, we may use some of the above-mentioned information, as well as your App Store searches and browsing activity, to determine which ad is likely to be most relevant to you. App Store browsing activity includes the content and apps you tap and view while browsing the App Store. This information is aggregated across users so that it does not identify you. We may also use local, on-device processing to select which ad to display, using information stored on your device, such as the apps you frequently open.

Just to put a fine point on this: according to Apple’s definition, collecting demographic information, downloads/purchases/subscriptions, and browsing behavior in Apple’s apps, and using that data to deliver targeted ads, is not tracking, because all of the data is Apple’s (and by extension, neither is Google’s collection and use of data from Safari search results, or Amazon’s collection and use of data from its app; however, a developer associating an in-app purchase with a Facebook ad is).


Apple’s advertising platform does not track you, meaning that it does not link user or device data collected from our apps with user or device data collected from third parties for targeted advertising or advertising measurement purposes, and does not share user or device data with data brokers.

One thing should be made clear: there has been a lot of bad behavior in the digital ad industry. A particularly vivid example was reported by the Wall Street Journal last month:

The precise movements of millions of users of the gay-dating app Grindr were collected from a digital advertising network and made available for sale, according to people familiar with the matter. The information was available for sale since at least 2017, and historical data may still be obtainable, the people said. Grindr two years ago cut off the flow of location data to any ad networks, ending the possibility of such data collection today, the company said.

The commercial availability of the personal information, which hasn’t been previously reported, illustrates the thriving market for at-times intimate details about users that can be harvested from mobile devices. A U.S. Catholic official last year was outed as a Grindr user in a high-profile incident that involved analysis of similar data. National-security officials have also indicated concern about the issue: The Grindr data were used as part of a demonstration for various U.S. government agencies about the intelligence risks from commercially available information, according to a person who was involved in the presentation.

Clients of a mobile-advertising company have for years been able to purchase bulk phone-movement data that included many Grindr users, said people familiar with the matter. The data didn’t contain personal information such as names or phone numbers. But the Grindr data were in some cases detailed enough to infer things like romantic encounters between specific users based on their device’s proximity to one another, as well as identify clues to people’s identities such as their workplaces and home addresses based on their patterns, habits and routines, people familiar with the data said.

It’s difficult to defend any aspect of this, and this isn’t even a worst case scenario: there are plenty of unscrupulous apps and ad networks that include explicit Personal Identifiable Information (PII) in these data sales/transfers as well; as Eric Suefert noted in 2020, the industry has had this reckoning coming for a very long time.

That, though, is why the “and” from Apple is so meaningful; here is the sentence again:

Apple’s advertising platform does not track you, meaning that it does not link user or device data collected from our apps with user or device data collected from third parties for targeted advertising or advertising measurement purposes, and does not share user or device data with data brokers.

This definition conflates two very different things: linking and sharing. The distinction between the two undergirded a regular feature of Meta CEO Mark Zuckerberg’s appearances in Congressional hearings; here is a representative exchange between Senator Edward Markey and Zuckerberg in 2018:

Should Facebook get clear permission from users before selling or sharing sensitive information about your health, your finances, your relationships? Should you have to get their permission?…

Senator…I want to be clear: we don’t sell information. So regardless of whether we get permission to do that, that’s just not a thing we’re going to do.

Meta doesn’t sell data; it collects it, and the third parties that leverage the company’s platforms for advertising very much prefer it that way. PII is like radioactive material: it’s very valuable, and can certainly be leveraged, but it’s also difficult to handle and can be dangerous to not just the users identified but to the companies holding it. The way Meta works is that its collective advertising base has effectively deputized the company to collect data on their behalf; that data is not exposed directly, but is instead used to deliver targeted advertisements that are by-and-large bought not by targeting specific criteria, but rather by specifying desired results: app installs, e-commerce conversions, etc. Everything user-related is, to the companies buying the ads, a complete black box.

This is where linking comes in: apps or websites that leverage Facebook advertising (or any other relevant advertising platform, like Snap) include a Facebook SDK or Pixel that tracks installs, sales, etc., and sends that data to Meta where it can be linked to an ad that was shown to that user. Again, this is completely invisible to the developer or merchant; technically they are sending data to Meta, since the conversion data was collected in their app or on their website, but in reality it is Meta collecting that data and sending it to themselves.

The reason why developers and merchants are happy with this arrangement is that advertising is a scale business: you need a lot of data and a lot of customers to make targeted advertising work, and no single developer or website has as much scale as, say, a Google or an Amazon; Meta et al enable all of these smaller developers and merchants to effectively work together without having to know each other, or share data.

Google, Amazon, and Facebook's ad businesses operate similarly, but only Facebook is affected by ATT

You can, to be clear, object to this arrangement, but it’s worth pointing out that this is very different than selling or sharing data with data brokers; all of the data is in one place and one place only, which is broadly similar to the situation with Google or Amazon (or Apple, as I’ll get to in a moment). The big difference is that Meta doesn’t own all of the customer touch points: whereas a Meta advertiser may own their own Shopify website, an Amazon advertiser has to list their goods on Amazon’s site, with all of the loss of control that entails. Apple’s definition, though, lumps Meta’s approach (which again, is representative of other platforms like Snap) in with the worst actors in the space.


Apple’s advertising platform does not track you, meaning that it does not link user or device data collected from our apps with user or device data collected from third parties for targeted advertising or advertising measurement purposes, and does not share user or device data with data brokers.

To the extent you think that the Bundeskartellamt is right, then it is this word that is the most problematic definition of all. One would assume that “our” means Apple-created apps, like News or Stocks: just as Amazon collects data from the Amazon app, of course Apple collects data from its own apps. The actual definition, though, is much more expansive; go back to the Epic trial and the exchange I recounted in App Store Arguments:

The argument that Judge Gonzales Rogers seemed the most interested in pursuing was one that Epic de-emphasized: Apple’s anti-steering provisions which prevent an app from telling a customer that they can go elsewhere to make a purchase. Apple’s argument, in this case presented by Cook, goes like this:

A tweet from Adi Robertson

This analogy doesn’t work for all kinds of reasons; Apple’s ban is like Best Buy not allowing products in the store to have a website listed in the instruction manual that happens to sell the same products. In fact, as Nilay Patel noted, Apple does exactly this!

A tweet from Nilay Patel

The point of this Article, though, is not necessarily to refute arguments, but rather to highlight them, and for me this was the most illuminating part of this case. The only way that this analogy makes sense is if Apple believes that it owns every app on the iPhone, or, to be more precise, that the iPhone is the store, and apps in the store can never leave.

Let me be precise in a different way that is relevant to this Article; Apple doesn’t particularly care about or claim ownership of the content of an app on the iPhone, but:

  • Apple insists that every app on the iPhone use its payment system for digital content
  • Apple treats all transactions made through its payment system as Apple data
  • Ergo, all transactions for digital content on the iPhone are Apple data

The end result looks something like this — i.e. strikingly similar to Facebook, but with App Store payments attached:

Apple's ad model looks similar to Facebook's

Here’s the key point: when it comes to digital advertising, particularly for the games that make up the vast majority of the app advertising industry, transaction data is all that matters. All of the data that any platform collects, whether that be Meta, Snap, Google, etc. is insignificant compared to whether or not a specific ad led to a specific purchase, not just in direct response to said ad, but also over the lifetime of the consumer’s usage of said app. That is the data that Apple cut off with ATT (by barring developers from linking it to their ad spend), and it is the same data that Apple has declared is their own first party data, and thus not subject to its ban on “tracking.”

This, needless to say, is where legitimate questions about self-preferencing come to the forefront. Developers who want to link conversion data with Facebook are banned from doing so, while they have no choice but to share that data with Apple because Apple controls app installation via the App Store; this strikes me as a clear example of the President of the Bundeskartellamt’s claim that “Apple’s rules apply to third parties, but not to Apple itself”.

I have been very clear that I disagree with those who want to ban all targeted advertising; I believe that targeted advertising is an essential ingredient in a new Internet economy that provides opportunities to small businesses serving niches that are only viable when the world is your market. After all, people who might love your product need some way to know that your product exists, and what is compelling about platforms like Facebook is that it completely leveled the advertising playing field: suddenly small businesses had the same tools and opportunities to advertise as the biggest companies in the world. At the same time, I understand and acknowledge those who disagree with the concept on principle.

What is frustrating about the debate about ATT, though, is that Apple presents itself as a representative of the latter, with its constant declarations that privacy is a human right, and advertisements that lean heavily into the (truly problematic) world of data brokering, even as it builds its own targeting advertising business. Gruber asked me on this morning’s episode of Dithering whether or not I would feel better about ATT if Apple weren’t itself doing targeted advertising, and the answer is yes: I would still be disappointed about the impact on the Internet economy, but at least it wouldn’t be so blatantly anti-competitive.

Apple, to its credit, has made moves that very much align with its privacy rhetoric by cleaning up some of the worst abuses by apps, including significantly fine-tuning location permissions, providing a new weather framework that makes it significantly cheaper to build a weather app (reducing the incentive to monetize by selling location data), and increasing transparency around data collection. Moreover, at this year’s WWDC the company introduced significant enhancements to SKAdNetwork that should make it easier for developers and platforms like Facebook to re-build their advertising capabilities.

At the same time, an increasing number of signals suggest that Apple is set to significantly expand their own advertising business; an obvious product to build would be an ad network that runs in apps (given that these apps run on the iPhone, Apple would in this scenario claim that collecting data about who saw what ad would be first party data, just like transactions are). Yes, Apple tried and failed to build an ad network previously, but a big reason that effort failed is because Apple didn’t collect the sort of data necessary to make it succeed.

What has changed is not just Apple, but also the data that matters: when iAd launched in 2010, digital advertising ran like people still think it does, leveraging relatively broad demographic categories and contextual information to show a hopefully relevant ad;1 what matters today is linking an ad to a transaction, and Apple has positioned itself to have perfect knowledge of both, even as it denies others the same opportunity.

  1. This is the era when Facebook earned its reputation for being far too cavalier with user data; Facebook was also the company that built the modern advertising approach that depends on linking data instead of sharing it.