Benedict Evans: Predictions For Mobile, Machine Learning, eCommerce & Cars

Its hard to predict what will happen in the future of tech and marketing and even harder to get it right but one of the best presentations that ive seen on the topic does a pretty good job. Benedict Evans works for a16z, a venture capital firm in Silicon Valley and presented “Mobile is eating the world” at its inaugural summit event. In it he covers the growth of tech as a whole on a scale never seen before, mobile smartphone users surpassing PC’s and showing no signs of stopping as we head towards 5 billion smartphones on earth, machine learning and the shift by companies to an “AI-first” approach and how this may affect the eCommerce & Car industries in the future.

If you’ve got thirty minutes, the presentation is definitely worth watching or you can read the full transcription below.

Benedict Evans: Mobile is eating the world

I sometimes feel that talking about mobile now is a bit like talking about PCs in 2000. That this enormous thing has now happened and we’re thinking about what we can do with it.

I think the first way to talk about this is to talk about deployment. If we look at a snapshot of the world today, we’re about halfway to connecting everybody. There’s about five and a half billion people on Earth over 14. There’s close to five billion people with a mobile phone, about two and a half billion people have smartphones. That will go to five billion smartphones. What’s happening is that the mobile S-Curve is passing the PC S-Curve. We have about one and a half billion PCs on Earth and you can see that growth coming up in the last couple of decades and now flattening out and starting to decline.

Then you can see how the first mobile and now smartphones have accelerated past that, heading towards pretty much the entire global population. Then if you systematise that what tends to happen in technology is that each new technology follows an S-Curve and is then replaced by another one. In the early period of that S-Curve, it begins as a crazy idea and then it becomes something that’s grown very fast and there’s a frenzy and there’s bubbles. Then you move from the creation phase to the deployment phase where the questions are much more around scaling a maturity. Then over time then there will be another S-Curve that comes after that. Mobile now is making that transition from creation to deployment.

What that means is that the issues that matter are changing. When we’re in the creation phase, and we talked about platform wars and we talked about the latest hand set and argued about technology, and there were millions or tens or hundreds of millions of users. The key question is well is this going to work and who’s going to win? Now we’re in the deployment phase so winners are clear. We’re in a phase of steady incremental improvement, that each phone is a bit better than the one before. A lot of the technology becomes commodity. We’re at the phase we have billions of users. The question now is well what could we build with this thing? What can we build with this amazing, enormous thing that we built that everybody now has?

Then to think about the things I’m going to talk about in this presentation, I think there’s a couple of developments to this. Firstly, we have a new kind of scale for technology. We have new kinds of computing. Those two things together give us new ways that we can change other industries. I’m going to go down and talk about a couple of those industries, first e-commerce and then cars.

To talk about a new kind of scale, first the obvious observation. We’ve had a changing of the guard and the new guard are a lot bigger than the old guard. So, the new dominant companies in tech are Google, Apple, Facebook, and Amazon, GAFA for short. They’re about three times the scale of the old dominant companies in tech, which were Wintel, that is to say Microsoft and Intel. They’re about three times the size in revenue terms. That’s pretty unsurprising. I think what’s more interesting if you go onto the next slide is to think about how you compare the scale of the dominant companies today with the scale of Microsoft and Intel back in the 1990s when they were changing the world. We have this period of very rapid growth. Now, how does that compare to the period of very rapid growth in the past?

Back in the 90s Wintel grew their revenue 14 times and they looked amazing. If you compare that to the scale of Google, Apple, Facebook, Amazon today, they have 10 times the scale. They’re much, much bigger than Microsoft and Intel were back in the days when they were changing the world. If you think in terms of valuation, back in 1995 when Bill Gates was on every magazine cover on Earth, Microsoft was actually not the biggest company on the stock exchange. Today Google, Apple, Facebook, Amazon, and yes Microsoft are all the five biggest companies on the stock exchange. We’ve gone from being important companies to being completely dominant companies to being much bigger, much more powerful, much more significant companies.

Now that change in scale means much more investment, annual capex across these companies has gone from $1 billion in 2000 to $34 billion in 2015. It also means many more people, 10 times more people, now over half a million people employed. What it really means I think fundamentally is that the tech giants today have a completely different character of scale to IBM or Wintel before them. It’s not just that they’re bigger, it’s that they’re giants in the economy rather than just giants within the tech industry. Also, I think this is also very important, we have four of them and not just one. It’s not IBM, it’s not Microsoft, it is Google and Apple and Facebook and Amazon all competing with each other. So you have things that flow out of that.

Apple is the 10th largest retailer on Earth by revenue with over $53 billion in sales last year, which is not bad for a marketing operation. If you then look at content and remember the days when content was king, well for Amazon content is just a lever to get more Prime subscribers and sell more soap powder. They don’t even care about content as a primary method. Netflix of course is a content company, but Netflix itself has the fourth largest production budget in the USA. You have tech companies impinging a fundamental level on other industries. There’s a great quote from 2010 the Chief Executive of Time Warner said that, “Thinking of Netflix as a threat was like thinking that Albanian Army was going to take over the world.” I suspect he probably wouldn’t have that attitude today. Then to look at another sector again and enter a whole business of making chips has become an add-on for these companies. Google is making custom FPGA chips for machine learning. Amazon is making custom ASIC chips for networking. Apple makes CSAC, the system on a chip inside the iPhone. It’s the fastest SAC that there is on the market. It’s making all sorts of other custom chips which other devices, whether it’s touch ID, airports, or other things.

Again you have this scene that wars like the centre of the whole tech industry and the core to competing now becoming almost a side project for other technology companies as the scale has grown. What’s happening here is that we’re building on scale. We’re standing on the shoulders of giants. We have the scale of five billion mobile users with this vast supply chain of two billion devices being sold every year. We have the scale of the winners in this environment, Google, Apple, Facebook, Amazon have much greater scale than Wintel had before. Then we have new ways to compete, whether that’s custom hardware, retail, distribution, content, potentially in the future cause as well. We have these enormous giants giving us shoulders to stand on and build all sorts of other things.

Now in the meantime we have the new thing, or the new, new thing which is machine learning. I’m not going to give you the technical explanation. I have colleagues who produce material around that. Here really I’m thinking about what are the implications of machine learning for the broader tech industry and for industries outside technology. A good place to start is this image, which is a test from ImageNet. Is there a dog in this picture? After 50 years of work computer vision systems could get this right 72% of the time more or less. There’s a whole class of similar problems so we all know where there are things that are very easy for people and hard or impossible for computers. This was really one of the focuses for research into AI. The general consensus around this kind of stuff was that there were decades of more work before all of this stuff would really start working. Then in 2012 machine learning started working and then really everything changed.

For image recognition the best results were getting a 28% error rate, those went to 7%. For speech recognition the best results were getting a 26% error rate, those went to 4%. Similar results across a whole class of problems that people had been working on for decades, performance has completely transformed. This stuff started working in a way that it had never really worked before. To understand what happened there it’s really worth thinking about a fundamental change from using rules to using data. The old way that you would build something to recognise a dog was you built something that looked for ears, and noses, and legs, and fur. You’d hired linguists and you’d write grammar rules and you’d try and codify how human intelligence works, which is different because we have no idea how human intelligence works really.

Then the new way that we do this, which I said all started working in 2012, was machine learning where you use data. You give it 10,000 pictures labelled “dog” and 10,000 pictures labelled “no dog” and you let a neural network that will work it out. You let the system by itself work out what the key criteria are, what the ways of deciding and choosing and analysing are. Hence we call this learning because it’s the machine that’s actually learning by itself. This becomes possible because we have millions and millions of times more computing power and millions and millions of times more data than we had when anyone tried to do this in the past.

What that means is best summed up by this quote from Frederick Jelinek, “Every time I fire a linguist, the performance of the speech recognizer goes up.” What’s happening here is that we’re analysing data, we’re not looking for dogs. This works for many other data sets. You can ask which of our customers are going to turn. Is that car going to let me merge? Is there anything weird happening on our network? What are the patterns inside of this information that I haven’t thought to look for? You have a generalizable solution. To give an extreme example, Google took the system that it used to win AlphaGo, the Chinese board game, in the spring and applied it to managing or optimising the cooling in one of its data centres. It achieved a 15% energy saving after Google having spent decades trying to optimise that by hand.

What we have here is a way of building systems that can find patterns in data in similar ways to the ways that people would have found that data, but at massive scale. Hence, a statement from Sundar Pichai, Chief Executive of Google that, “We will move from mobile-first.” Google will move from mobile-first to AI-first. Everything in the tech industry is now being refocused from mobile to mobile plus AI. There’s a whole bunch of use cases or indeed a whole bunch of ways of thinking about what the use cases might be, because it’s still very, very early. I think the fundamental thing is that any data set can have new analysis applied to it. Secondly, that this opens up all sorts of opportunities for interaction because now computers could look at images or video or voice in completely new ways. That enables new devices and new interaction models. A particular thing that I want to pull out is to think about images, to think about computers being able to read images, and then of course there will be our bottom two surprises. There will be a whole bunch of things that don’t actually look like machine-only use cases, but they turn out to be just that.

This also happens at the top and the bottom of the stack. You train your machine-only models at vast scale in the cloud with huge amounts of data. Once you’ve created that model, you can very often put them on very small, very cheap devices. You can run them natively on a smart fan, but you can also run them on something even smaller. You could make a $10 or $20 widget with a camera and a machine-only running on a DSP that could look for people or could look for cars or could look for some specific use case. That becomes particularly interesting when we think about imaging. One of the small things the supply chain has done is make image sensors available for almost anything. You can see the size of global camera market. You really have cameras in absolutely everything, except of course in cameras. Your orange line here is the camera market itself which is now shrinking down to a professional lump. The image centre itself becomes this universal thing that can sit in anything or can apply to any kind of image problem. I don’t think we quite know what it means when we say that computers can read images in the ways that they’ve always been able to read text. We don’t quite know all the implications of that might be.

As we build all of this stuff, we have this surge in complexity. We also have a rush to extraction. What I mean by that is that on the one hand we’re going from the fundamental scientific inside to building an enormous amount of engineering as we work out, well what is the right kind of machine I need to solve this particular problem, as we build all of the implementation and the product and the execution on top of the underlying concept, even as society itself has much further to go. Both the engineering and the science are both moving at a very, very rapid speed. In parallel, we have Google, Amazon, Microsoft, IBM, other companies putting this into the cloud and making it available as a platforms to other people. You can take the speech recognition that Amazon uses for Alexa and you can plug it into your end product using an Amazon cloud product, the same for Google, the same for Microsoft, the same for IBM.

On the one hand you have this huge increase in complexity of what is possible and what’s being built. On the other hand, you also have again the GAFA competition, trying very hard to make that available as a simple layer to anybody else at the same time. What I think is really going on here, again to return to the S-Curve, is that machine learning is almost at the beginning of the creation phase and so we’re having this explosion and this frenzy of creation, and invention, and discovery, and argument about who would win and what would winning mean, with an awful lot of low-hanging fruit and an awful lot left to build.

One of the things that that enables then to think about the next section is new kinds of computers or new kinds of computing. There’s an obvious boiling point to think about mobile, which is out of one smartphone app to now 60% of all time spent online in the USA. We all kind of know this. We’ve talked about this an enormous amount. That also means increased concentration, so 15 to 20% of mobile time spent is actually happening inside Facebook, which makes it the world’s largest mobile web browser. There’s increased concentration, it’s harder to break out, it’s harder to become a breakthrough app, it’s harder to get lots of users. We know all of this. I think what’s more interesting is to think about a new generation of computing that’s happening as a result of this.

So on the left you have to original Apple Mac. It didn’t really change that for 30 years until we went to smartphones. You had a keyboard and a mouse and a screen. You kind of moved your hand somewhere over on the right and a cursor moves somewhere over on the left two or three feet away. Then we go to smartphones and we really change a huge amount. We have this device, it’s intensely personal, and that’s always with you, and that has imaging, and a screen as a primary input, that has touch, and tilt, and senses and an app store. Where very often you can ignore bandwidth and you can ignore battery and I’ll talk about that in a moment. Well of course you’re a thousand times faster, you have a cloud machine learning, and of course much easier to use.

On a fundamental level, you have this new computing model that is both much more sophisticated and much easier to use at the same time. That enables, on the one hand, a wave of new senses and inputs and interfaces, machine learning, all kinds of new products that can create. On the other hand, it enables all kinds of new creation live, social, broadcast, touch, imaging, all sorts of things that a consumer can do with a device that they could never have done with a PC. A good illustration of this I think at the moment are all of the current crop of live video apps. I think it’s interesting to think about how many assumptions live video breaks. You assume a high quality camera, two high quality cameras, consumer GPU that can do live image recognition that can do live effects, assume attached screen, assume it’s always on, always connected, assume it’s always in your pocket, assume it knows who your friends are, assume unlimited bandwidth, assume unlimited battery. Bare in mind roughly half of all smartphone use happens at home so you can plug it in in your wifi, so you do have unlimited bandwidth and battery.

Then also assume, and this is an inverted point, assume there are a billion high-end smartphones. For 15 years people in mobile have always said, “I will remember not everyone has a high-end phone and remember not everyone has an iPhone.” Well, there’s a billion high-end smartphones on Earth now, so a lot of companies actually can just assume everyone has a high-end smartphone. What you’re doing here is you’re turning the whole phone into a camera or turning the camera into a computer. You’re really breaking all sorts of interface assumptions around what you could build on a PC or what you could have built on a mobile phone in the past.

The other extreme, you have frictionless computing I think in devices like Alexa, the Amazon Echo, or Snapchat spectacles. You don’t have any buttons or any apps or any intermediate steps. There’s no cable, you don’t have to think about charging it really or it’s charged for a week. There’s no computer stuff, no clerical work you have to do, no admin you have to do before you use it. I think what’s interesting about these devices is that they’re not just a camera or a microphone or speaker. They’re actually unbundled pieces of apps. The Amazon Echo is the unbundled Amazon app. The spectacles are the unbundled Snapchat app. You’re moving that app into a new context. Of course, they were built with commodity components from the smartphone supply chain. There’s no unique hardware to it all here in any of these things.

Then you light them up with machine learning. Say you have the Senses, which is a smartphone, all the smartphone components, and then you layer machine learning on that, speech recognition, computer recognition, whatever else it is, and then whatever it is you actually want to do, the action on the end of that. One should hesitate just for a moment and remember that just because recognise what somebody said doesn’t mean that you can do anything about it. You have to narrow the domain to the things that you can actually do and work out how you can tell your users, “Actually I can do this but I can’t do that.” I think that’s one place where Amazon has been very successful with Echo and perhaps more successful than a bunch of previous voice assistants.

When you’ve done that, when you’ve removed that friction you also … You’re removing friction and you’re changing choices. I think it’s interesting to think about the direction of travel of computing is being about removing questions. Do you have the printer driver? What’s your password? Where did you save that file? Do you want to restart the computer? Do you want to turn it on? Do you want to turn it off? Do you want to plug this is? All these all kind of questions that we don’t think about anymore that computers used to ask us before plug and play and then smartphones got rid of. If you ask Alexa to order you soap, does it ask you what kind of soap? Does it ask you what app you want to use? If you take a picture with spectacles does it ask you what app you want to share it with or who you want to send it to?

Again, you’re removing friction, you’re removing questions, you’re removing admin. Of course, as you remove those questions you’re also removing choices. That’s creates a huge strategic incentive for the big platforms, particularly for Google, for Apple, for Facebook, and for of course for Amazon above all to be the ones who can answer that question. These kind of devices and these applications and interfaces create all kinds of new questions as to who it is that actually earns a customer and who it is that’s going to be the one who answers that question for you.

I think really at a fundamental level what’s going on here is a direction of travel. It used to be that you would have direct physical interaction with data. You would actually hold the data in your hand. You could make new zeros with a pen. You were holding the ones and zeros in your hand. We’ve gone from direct, physical interaction with data on the one hand, to direct physical interaction with data quite at the other extreme. Now you’re touching it, you’re looking at it, you’re talking to it, you’re holding it. The next generation is removed to augmented reality. You’ll just be looking at it and waving your hand at it. You have direct physical interaction from one extreme then to a complete opposite extreme.

All of this gives us shoulders of giants to stand on, new computing, new scale, new interaction models, new ways of interacting with computers. I’m now going to talk about two ways that that changes things that are not actually technology, that change other industries. The first one those is e-commerce. I think it’s useful to think about retailers as newspapers. That is to say newspapers have a fixed cost base with falling revenue. They shifted from a paid model to a freemium model. Their entire product got unbundled. Their fundamental distribution advantage disappeared and as we went to new medium, you had different consumption. It wasn’t just that you read the same things in a different place, you read different things. Just with e-commerce, it’s not just that you buy the same things in a different place, you buy different things.

Conceptually, I think what I would say is that everything that the internet did to media will happen to retail. You break up all bundles and aggregate whether that bundle is an album, or magazine, or a newspaper, or a store, or a shopping district, or a mall. They’re all aggregators and bundles. Then new aggregators emerge and those new aggregators shape consumption in different ways. They result in you buying things or consuming things in different ways.

So far what e-commerce has done is give you the stuff you already knew you wanted. You know that you want to buy that thing so you can go on Amazon and get it. Today e-commerce is about 10 to 12% of US retail revenue, a little bit more in some other countries, a little bit less in some others. Amazon is at least 2% of that. If you include marketplace revenue it might be 4% or even 6% depending upon your estimates. So far e-commerce has been much better at logistics than at demand generation. It’s been much better at retailer’s logistics then at retail as experience or it’s suggesting what you might want to buy. That said, it’s got to almost every product category. 15, 20 years ago people were saying, “Well no one will every buy clothes online.” Today Amazon is the fourth largest clothes retailer in the USA, not the fourth largest online clothes retailer, the fourth largest clothes retailer. People will buy anything online if they know that it exists, particularly with free returns which is just another piece of logistics.

The question is how do you know about it? How do you know what you want to buy? Amazon is Google for products. That’s where you go to find something where do don’t have Facebook for products. Not Facebook in the sense of friends, but Facebook in the sense of all the suggestions it gives you for random stuff you might want to look at. We don’t have buzzfeed for photos, we don’t have working viable, systematic, scalable ways of suggesting things and discovering things. We have a lot of great interesting companies trying. That’s not a solved problem where there’s one great company that does everything. What that means is that the internet lets you buy, but it doesn’t really let you shop. It lets you buy anything you could get in New York or London or Tokyo. It doesn’t let you shop the way you can shop if you live in New York or Tokyo or London.

The first response to this, the first way of digital demand generation is advertising. The advertising industry together with marketing globally is about a trillion dollars every year. $500 billion spent on ads, of which about a third is on digital and half of that third is on Google, and then another a half trillion dollars go on marketing, so end-caps, coupons, promotions, discounts, everything else, ways of getting people to buy things. When I look at that chart and I’m reminded there’s that line from Jeff Bezos, “Your margin is my opportunity.” I rather suspect that advertising is an opportunity, just as much as margin is an opportunity. I think what happens here to channel [Marcia McLewan 00:20:14], is that one should think that the channel is the product. As you change where you buy things, you change how you buy things. You buy different things in a supermarket, from a grocery store, you buy different things online. Again, the channel shapes what you want to buy.

There’s two examples to think about here. One of them is what you might call Soap as a Service. That is to say you ask Alexa to get you more soap or you buy an Amazon Dash Button and you put it in your washing machine. You never make a brand decision again. You don’t decide what soap to buy anymore. You’ve made that decision. You’ve just intermediated Walmart, but you’ve also just intermediated an ad agency. In a sense, you’ve also just intermediated Proctor & Gamble who, among other things, bundling chemicals up with a brand and brand equity and distribution. You’re completing changing the whole journey of how that soap powder gets from a chemical factory to your home and why that brand rather than another brand gets to you.

At the opposite extreme, our portfolio company Walker & Company is building a men’s grooming product is able to build that and take that to market in ways for an enormous amount less money than what it would cost you in the offline days because of the amount you have to spend on promotion, the amount you have to spend on inventory, the amount you have to spend on distribution has been completely changed. On the one hand you can totally change something with a Dash Button, on the other hand you can take a conventional product and just sell it in a different way online.

I think if we then go a step further and think about machine learning and how that applies to these kinds of questions, how imaging applies to these kind of questions, we get some other kind of interesting opportunities. Now what would happen if you were to buy the last 10 years of ELLE Decoration on Ebay and drop that onto your neural network and then take a photograph of your living room? What might it be able to suggest? Would it be able to tell you what sofa you would like, what cushions you would like, what lamp you might like? Those are things you could never do with computing in the past. We know how you might do that with computing today, whether that’s for clothes, for makeup, for travel, for cooking. Whole categories of product that are around personal taste and around what you see and what you think you would like, rather than just the things that you already know. What would I like that I have never seen? Well, in the today you have to go to a store or look in a magazine. You need some kind of manual editing or curation. We may be able to do that in a completely different ways in the future with imaging, with machine learning in particular.

A canonical quote here I think from Eric Raymond, “A computer should never ask the computer for anything that it should be able to work out by itself.” Well, what will the computer be able to work out? What do you want to buy? What questions will it be able to answer for you again? What suggestions will it be able to make? I think one of the ways of thinking about what’s happening here then is that data is working its way through retailing. In the 90s we had computing coming into the supply chain with ERP and completely changed that part of the industry. In the 2000s we have advertising, certainly at least digital advertising, giving a little bit more data as to how you think about how people buy stuff. What’s really going to happen now as we think much more about demand, about how you choose what you might buy because with digital as you go online and as you add imaging, as you add e-commerce, as you add machine learning, you still aren’t having completely new ways of what you might want to buy.

If we then pull back to 1980, then here are some more S-Curves. This is the S-Curve for Walmart and Walmart of course used new technology to change retail. It used trucking, and refrigeration, and computing, and various other things in order to change what retail would look like. Amazon is just starting to that now. We will see more S-Curves as we use new technology again to change how retail would work. If one then thinks about the scale of the opportunity here, this chart shows you some of the sectors that have been upended by software today. I can see my media, mobile phones, perhaps pay TV, perhaps TV production. If you then zoom out, retail is something over $20 billion opportunity. All of that is potentially going to be affected by software, by machine learning, by e-commerce, by mobile.

There was these shops, to now to think about from the small cardboard boxes flying around the world, to think about very large metal boxes speeding around the world with cars. One extreme to the other. If retailers are newspapers, cars are definitely famous. Where history doesn’t repeat itself, but it rhymes. Ask for phones, ask for cars, today’s technology may disappear. Key components today, key points of differentiation today become commodities. All of the value moves to software. Network effects move away from the hardware into software and into the cloud. The word phone and the world car mean something completely different. I’m going to talk about all of these in turn.

First, as we disrupt cars as a tech industry disrupt cars, we need to almost completely separate paths. The first is electric and the second is autonomy. The move to electric is happening today. What electric does is that removing the engine and the transmission totally destabilises the car industry and its suppliers. Moving from gasoline to electric is not about taking the gas tank out and putting a battery that it completely changes how the car gets made. It doesn’t however change how the car gets used very much. Autonomy changes that as well. Autonomy is in its very early stages today. It is perhaps five to 10 years away. There were an awful lot of challenges to resolve, one after another, but when autonomy works, it changes the value. Of course it accelerates on demand, but it changes what cars are and it changes cities probably as much as cars themselves change cities.

To talk about electric first. The shift to electric is about the battery cost curve and about a function of battery economics starting to work. Right now battery per dollar per kilowatt hour is something over $200. As it gets down closer to $100, it becomes completely cost substitutional with gasoline. That just is a question of time. As you move to electric you unbundle the cart. Your complex proprietary gasoline engines and transmission systems disappear and get replaced by much simpler commodity batteries and motors. This whole central spine of the car that was where all the money and the investment and the differentiation was, that all just goes away. You have ten times fewer moving parts. The basis of competition in making the car goes away. It’s not about the gasoline, it’s about the engine and not having an engine`anymore in the same way.

Now, that doesn’t mean that everything changes. Scaled design brand distribution still matter for people who can navigate the change, but a huge amount of value moves up the stack into the software, which is an entirely new skillset. As you go to autonomy, you get new layers of value being built on top of this. A great precedent for thinking about this … Again, back to the phone analogy, is what happened to phones? Back in 2000 or 2001 you had to have co-invented GSM to make phones. All of the technology and the invention was in the cellular, it was in how you talk to the network and in understanding that. Nokia said that manufacturing was a core competency. That was a fundamental competitive advantage. 75% of all Nokia phones were made in eight factories. They had 40% of the market. They dominated the world. Today Apple’s iPhone has 189 suppliers, 789 locations, none of those are owned by Apple. A small number of those components are unique, and distinctive, and created by Apple. But, an enormous amount of that has nothing to do with Apple at all. The whole component supply chain of the phone industry got completely unbundled, which is why you now have not just hundreds, but in a purely literal sense, thousands of companies in China making mobile phones.

Then to think about the scale here. This is kind of an unfair, but relevant comparison. Leading technology companies are now spending as much on capex as car OEMs. These are big companies. Apple, meanwhile, has $237 billion in gross cash on its balance sheet. Google has $73 billion. Questions of scale and engineering and bashing metal are entirely accessible to new kinds of companies. There is electric. Now as we think what happens as we add autonomy to this.

There is a set of steady steps or steady progress towards autonomy. This is the standard industry terminology from level one through to level five. Level one is the cruise control that your grandparents car had. Today we are at the stage of building cars that can more or less stay in their lane and more or less slow down, but you’ve really got to be sitting in the driver’s seat ready to take control and watching the wheel all of the time. That’s a car that will probably have fewer accidents, but it’s not a car that can drive itself. What we move towards is level four and then level five. The car first where the driver can sit in the backseat, but the car might have a problem. Then level five where you don’t have a steering wheel and where a commercial vehicle might have no cabin and might not have a human involved at all. That level five is probably five to 10 years away with the emphasis on 10 years depending on who you talk to.

At the point that you do that, you start thinking, “Well what’s a competitive mate in cars here?” We should presume that batteries and sensors will become commodities. Where are the strategic levers? There is a clear strategic level in physical scale. Bashing metal is still about physical scale. Design, manufacturing, there is a lot of expertise and difficulty in that. Then as you move to the software, well there are network effects. There are network effects in the driving data to feed your machine learning agent. There are scaling and network effects in the possession of your high density 3D maps, so the cars need to move around. It’s not yet clear how many winners there will be in this or even who the winners might necessarily be. There were very strong network effects. With on-demand, that you can layer on top of that. Again, there are strong network effects, a virtual circle of a density of colours, a density of writers, which tends to support Uber and Lyft as opposed to anyone trying to create their on-demand fate again from scratch. Really for all of these, these are unanswered questions. It’s so early that it’s not yet clear quite when the answers to these will look like.

When you have that though, when you have autonomy, talking about a self-driving car, a term I deliberately haven’t used until now, a self-driving car is a bit like a horseless carriage. That is to say you take the horse off, but that’s not really the only thing you change. When you’ve taken the horse off then you change everything else about the vehicle. The same thing with both electric and autonomy, when you take the gasoline engine out and you take the steering wheel out, you will change everything else about the car. We will have completely new kinds of vehicle types. We will also have completely new ways of using those vehicles. What that means is we can focus on the obvious impacts of this stuff, which oil and safety. It’s easy to talk about those. Half of global oil production goes for gasoline for cars. One and a quarter million people are killed every year in road accidents around the world. Those are the obvious things. The second order of facts, if anything wouldn’t be much bigger.

What happens to the industry and servicing cause if you have ten times fewer moving parts? What happens to the machine tool industry? What happens to power generation and the storage with all those cars with electric? What gets sold in 150,000 gas stations if nobody’s buying gas anymore? What happens to taxes? Then when you look at autonomy, if every car on the road is fully autonomous there’s no more parking and there’s no congestion, unless my chest congestion. Every car can drive down a freeway two feet apart from the car in front of it. What happens to housing in that environment? How far are you willing to live? What happens to logistics in retail? What happens to commercial real estate? What happens to trucking if you don’t need a driver? What happens to the ownership of cars? What happens to insurance?

Really once you think about autonomy changing cities as much as cars themselves change cities and in just as many unpredictable ways. If we’d been having this conversation a hundred years ago we probably could have said that everybody on Earth would have had a car, the other half of the people in the room would have said you were crazy to suggest that. But, we wouldn’t have predicted Walmart. We wouldn’t have predicted freeways. We wouldn’t have predicted all the other changes that happen to cities as a result of cars for better and for worse.

Where I think that takes us is that the biggest changes here are probably unknowable. This is an illustration from the year 1900 of what going to the opera in 2000 would look like. We have flying cars, nothing else has changed. We have Paris in the year 1900 with flying cars. That tends to be the fallacy that people make in looking at the future, that our current society will go into the future totally unchanged, except now we’ll have robots or flying cars or computers. A housewife will have a robot to do the dishes for her. That tends not to be how it works. That wouldn’t be how it works for autonomous cars as well. Thank you.

About The Speaker – Benedict Evans

Benedict works at Andreessen Horowitz (‘a16z’), a venture capital firm in Silicon Valley that invests in technology companies. He tries to work out what’s going on and what will happen next and you can find his own blog here.

Leave a Reply

Your email address will not be published. Required fields are marked *