Author: Benedict Evans

Netflix, Shein and MrBeast

A couple of years ago, I wrote an essay arguing that Netflix is a ‘TV company’, not a ‘tech company’, simply on the basis that all the questions that matter for it are TV questions. What rights, what kinds of shows, what commissioning models, what stars on what basis... all the questions that matter are LA questions. Netflix goes to LA and hires LA people to buy LA things from other LA people. It used technology as a new channel to enter the market, and that technology has to be good, but if it was only showing Friends and Cheers, no-one would care if the compression ratio was 20% better.

Part of this argument is that the technology of streaming itself is mostly a commodity - a hard commodity, and one that you have to get right, but still a commodity - and the key point of leverage is somewhere else. Five years ago a lot of people in Silicon Valley would have laughed at the idea that ‘legacy media companies’ could ever write good software, but Disney has done OK. It might be no more than ‘OK’, but what matters is the TV.

I think you could make a similar point about Shein - both Netflix and Shein realised that you can make far more SKUs if you’re not constrained by physical inventory - the time slots on linear TV and the store rooms of physical retail. So Netflix makes far more TV shows than anyone else, and indeed (Read more...)

Retail, search and Amazon’s $40bn ‘advertising’ business

I first started paying attention to Amazon’s ad business a couple of years ago, when a line called ‘other’ in the back of the accounts started to grow really fast. In 2021 Amazon started splitting it out, and last year Amazon had $38bn of ad revenue, more than any traditional media company and bigger than the entire (now much reduced) global newspaper industry. 

The simplest way to look at this is as a very profitable new business. $38bn is a small percentage of Amazon’s overall $502bn 2022 revenue, but it probably had well over 50% operating margins, which would mean it brought in as much operating profit as AWS - $20-25bn, and without anything like as much capex. (Amazon's overall operating income, meanwhile, was only $12bn.)

More interesting, perhaps, is the fact that $38bn is also more than reported revenue for Prime, which also points to the problem with trying to calculate a stand-alone profitability for this. Prime has directly attributable revenue and costs, but it’s also an important marketing tool, and it drives increased purchasing: the real revenue from Prime is a lot more than subscription fee. In the same way, calculating an actual P&L for Amazon Ads would be pretty artificial (this is also why Apple doesn’t generate a P&L for the app store, even internally). 

So, what are the ads doing for Amazon? Some people think they’re degrading the experience - certainly, it’s hard to shop without wading though them. But you could also ask whether this is ‘advertising’ (Read more...)

ChatGPT and the Imagenet moment

A decade or so ago, systems based on something called ‘machine learning’ started producing really good results in Imagenet, a contest for computer vision researchers. Those researchers were excited, and then everyone else in AI was excited, and then, very quickly, so was everyone else in tech, as it became clear that this was a step change in what we could do with software that would generalise massively beyond cool demos that recognised cat pictures.

We might be going though a similar moment now around generative networks. 2m people have signed up to use ChatGPT, and a lot of people in tech are more than excited, and somehow even more excited than they were about using the same tech to make images a few weeks ago. How does this generalise? What kinds of things might turn into a generative ML problem? What does it mean for search (and why didn’t Google ship this)? Can it write code? Copy? Journalism? Analysis? And yet, conversely, it’s very easy to break it - to get it to say stuff that’s clearly wrong. The wave of enthusiasm around chat bots largely fizzled out as people realised their limitations, with Amazon slashing the Alexa team last month. What can we think about this, and what’s it doing? 

The conceptual shift of machine learning, it seems to me, was to take a group of problems that are ‘easy for people to do, but hard for people to describe’ and turn them from logic problems into statistics (Read more...)

Ways to think about a metaverse

Sometimes it seems like every big company CEO has read the same article about the same tech trend, and sent the same email to their team, asking “What’s our strategy for this?!“ A couple of years ago there were a lot of emails asking for a 5G strategy, and now there are a lot of emails asking about metaverse.

Answering the 5G email was actually pretty easy, partly because almost no-one needs a 5G strategy at all (I wrote about this here), but also because we knew what 5G meant. We probably don’t know what ‘metaverse’ means. More precisely, we don’t know what someone else means. This word has become so vague and broad that you cannot really know for sure what the speaker has in mind when they say it, since they might be thinking of a lot of different things. Neal Stephenson coined the word but he no longer owns it, and there’s no Académie Française that can act as the tech buzzword police and give an official definition. Instead ‘metaverse’ has taken on a life of its own, absorbing so many different concepts that I think the word is now pretty much meaningless - it conveys no meaning, and you have to ask, ‘well, what specifically are you asking about?”

If you do ask that, I’d suggest that there are two broad sets of things that people might mean when they say ‘metaverse’.

First, the narrow definition is simply that some combination of VR and AR will (Read more...)

Rocket ships and tractors

"Annual income twenty pounds, annual expenditure nineteen pounds nineteen and six, result happiness. Annual income twenty pounds, annual expenditure twenty pounds nought and six, result misery" - Mr Micawber

“If you're offered a seat on a rocket ship, don't ask what seat” - Eric Schmidt

In the early years of Facebook, a common criticism we heard over and over was that this company and many like it had little or even no revenue and were burning huge amounts of cash, even years after launching, and so they would clearly never make any real money. By extension, they couldn’t possibly be worth the valuations they were attracting. It’s worth remembering now that people said this not because they turned out to be wrong (people are wrong about tech all the time, on both the positive and negative sides) but because their premise was wrong. This was a misunderstanding of the nature of a consumer internet company with network effects and very little marginal cost. 

For such businesses, knowing how, exactly, you’re going to make money might not be the most important thing to focus on. If you are acquiring tens or hundreds of millions of users with a new kind of service, and they are attributing value and attention to you, and the users, attention and value have network effects and hence probably winner-takes-all effects, and if they come with little or no marginal cost, then the revenue can and probably should come later. It is probably more important to focus on building (Read more...)

Within and tech M&A

Lina Khan, the head of the FTC, first became known for a law school paper arguing that US antitrust policy in recent decades has become far too narrowly focused on low consumer prices as the only test for harm, ignoring choice, competition and innovation, and that Amazon was an exemplar of this, since it uses its market power in ecommerce to push prices down, not up.

The paper was somewhat flawed by its unquestioning acceptance of the myth that Amazon sells a loss and is subsidised by investors, neither of which have been true for some 20 years, but the general analysis of US policy is perfectly valid, and while it wasn't really a new observation, the paper probably catalysed a shift in conversation about antitrust policy in the USA.

Of course, one could argue that this shift has also been driven by how much bigger leading tech companies have become in the last decade (simply because tech has become so much bigger), and how much more central to our lives (and politics) tech has become. There are plenty of US industries that look a lot less competitive than tech (meat-packing, for example), but tech is new and different and much more visible, and politically friendless. So it’s in the crosshairs, and Biden appointed Khan to the FTC with that mandate.

A second strand, meanwhile, is that this general shift in attitude towards companies with large market shares, particularly in tech, is accompanied by a new degree of attention paid (Read more...)

Back to the trend line?

Back in 2020, as we were all locked down and forced to do everything online, we got very excited about ecommerce penetration. All sorts of charts went viral showing that we’d jumped forward anything from three to five years in a couple of months. This was a big part of the ‘Covid Rotation’, and now we’re on the other side of that rotation - people went back to the office, and back to stores, and back on planes. And for retail and ecommerce, it looks like a lot of that growth was temporary, and we’re reverting to the trend line.

There’s a pretty obvious question in this chart, though - percentage of what? This is what the absolute numbers for US e-commerce look like. That reversion to the trend line suddenly doesn’t look quite so obvious.

The issue, of course, is that total retail sales have been far from stable, and so as the denominator has swung around, so has the e-commerce penetration. US ecommerce has, in fact remained at the new higher level - for now - but there has been a surge in physical (and, of course, inflation has suddenly shot up to 10% or so).

You can see this more dramatically in the UK, which had a much bigger lockdown and hence much bigger swings in the denominator and in the penetration. The penetration percentage has spiked all over the place.

Equally interesting, I think, is the question of which ‘retail’ number we should use as the (Read more...)

There’s no such thing as data

Technology is full of narratives, but one of the loudest is around something called ‘data’. AI is the future, and it’s all about data, and data is the future, and we should own it and maybe be paid for it, and countries need data strategies and data sovereignty. Data is the new oil!

This is mostly nonsense. There is no such thing as ‘data’, it isn’t worth anything, and it doesn’t really belong to you anyway.

Most obviously, ‘data’ is not one thing, but innumerable different collections of information, each of them specific to a particular application, that aren’t interchangeable. Siemens has wind turbine telemetry and Transport for London has ticket swipes, and you can’t use the turbine telemetry to plan a new bus route. If you gave both sets of data to Google or Tencent, that wouldn’t help them build a better image recognition system.

This might seem trivial put so bluntly, but it points to the uselessness of very common assertions, especially from people outside tech, on the lines of ’China has more data’ or ‘America will have more data’ - more of what data? Meituan delivers 50m restaurant orders a day, and that lets it build a more efficient routing algorithm, but you can’t use that for a missile guidance system. You might not even be able to use it to build restaurant delivery in London. ‘Data’ does not exist as one, single, unified thing, where you can add every row and table of every different kind (Read more...)

TV, merchant media and the unbundling of advertising

About five years ago, a revenue line buried in the back of Amazon's accounts started to get quite big. ‘Other revenue’ was over $4bn by the end of 2017, and if you looked at the notes to the notes, you discovered that this was ‘predominately advertising’. By 2019 this had grown to $14bn, and I wrote about it here, pointing out that ‘Amazon’ was no longer just e-commerce and AWS, and had become a bundle of lots of different businesses, many of which were probably just as profitable as AWS. However, we still didn’t know how what ‘predominantly’ meant. At the end of 2021 this changed: Amazon started splitting out the ad revenue directly, telling us that this is now a $31 billion business.

$31bn is roughly the same size as Google Display, YouTube, or the entire global newspaper industry’s ad business.

This is only about 6.5% of Amazon’s net revenue, but it has much higher margins. Google’s ad business has close to 60% operating margins excluding TAC; Amazon’s ads should be higher margin, given that it’s mostly leveraging the core businesses’s existing cost base (in other words - this is ‘found money’), but even assuming the same 60%, that would be $18bn of operating income in 2021, almost exactly the same as the $18.5bn that Amazon reports for AWS. Given AWS’s capex requirements, this makes it extremely likely that the ad business produces more cashflow.

The margin differential between e-commerce and advertising has become a much bigger story (Read more...)

Tech questions for 2022

Sometimes the centre of gravity in tech is very clear - everything is about PCs, or the web, or smartphones. But at other times, there are lots of things going on and none of them are The Thing, and all of them are full of questions. Of course, for some crypto people crypto is the only question and the only answer, but as we enter 2022 there are lots of areas where trillion dollar questions are wide open. These are the questions I wonder about at the moment - there are others.


Crypto is so big and potentially important, and yet so vague and so early, that we can’t even agree what to call it, and at times the noise of both irrational, religious hype and straw-man attacks can seem overwhelming. There is a set of ideas that could in principle be as central to tech as machine learning or open source, but after that, everything is a question. 

The tech itself is in a period of massively increasing sophistication and complexity, as everyone builds on an open canvas and builds capability on a simple idea - early PCs or indeed the early consumer internet looked like this. But the more layers, abstractions, building blocks and primitives are created, the harder it is to know which will resolve into things normal people can use, and, paradoxically, the more likely gatekeepers become. We’re imagining the metaverse while arguing about how TCP-IP should work and whether this new ‘WWW’ thing is (Read more...)