The Amazon machine

When you look at large manufacturing companies, it becomes very clear that the machine that makes the machine is just as important as the machine itself. There’s a lot of work in the iPhone, but there’s also a lot of work in the machine that can manufacture over 200m iPhones in a year.

More than any of the other big tech platform companies, Amazon is a machine that makes the machine. People tend to talk about the famous virtuous circle diagram - more volume, lower costs, lower prices, more customers and so more volume. However, I think the operating structure of Amazon - the machine that makes the machine - is just as important, and perhaps less often talked about.

Amazon at its core is two platforms - the physical logistics platform and the ecommerce platform. Sitting on top of those, there is radical decentralization: Amazon is hundreds of small, decentralized, atomized teams sitting on top of standardised common internal systems. If Amazon decides that it’s going to do (say) shoes in Germany, it hires half a dozen people from very different backgrounds, maybe with none of them having anything to do with shoes or ecommerce, and it gives them those platforms, with internal transparency of the metrics of every other team, and of course, other people (and Jeff) have internal transparency to their metrics.

These are the famous ‘two pizza teams’. The obvious advantage of a small team is that you can do things quickly within the team, but the structural advantage of them, in Amazon at least (and in theory, at least) is that you can multiply them. You can add new product lines without adding new internal structure or direct reports, and you can add them without meetings and projects and process in the logistics and ecommerce platforms. You don’t (in theory) need to fly to Seattle and schedule a bunch of meetings to get people to implement support for launching make-up in Italy, or persuade anyone to add things to their roadmap.

This means not so much that products on Amazon are commodities (this is obvious) but that product categories on Amazon are commodities.

This model has two obvious consequences for Amazon. The first is that it can scale almost indefinitely - if you can launch x in y without a meeting or a new org structure, the speed of expansion into new categories is limited mostly by your ability to hire and to procure (and also by consumers’ willingness to buy a new category online, of course). The second is that the buying experience for any given product category ultimately needs to fit a lowest-common-denominator model. The platform teams cannot easily create custom experiences for each new category.  You can see this sometimes as a weakness if you poke around many categories: Amazon can go almost indefinitely broad, but not necessarily deep. Hence, there are questions as to what categories might need a deeper experience - most obviously, now, higher-end clothing.

There’s a third consequence, though: those atomised teams don’t actually need to work for Amazon. If the constraint to the model’s growth is how fast you can hire product teams and sign supplier agreements, letting other people do it for you and charging them a margin (and of course the internal teams also have margin targets too) lets you scale faster and with less risk.

This is the insight behind Marketplace, which gives external teams wholesale access to the logistics platform. AWS (which does something similar for the cloud platform) is now 10% of Amazon’s revenue, but the fees charged to third-party vendors using Marketplace are close to 20%, and this represents over half of the total volume of goods sold through Amazon.

Amazon does not include the total funds paid by consumers for Marketplace purchases in its reported revenue and does not report the number either - rather, it reports as revenue only the service fees it charges to the vendors. Estimates of the total value of goods being sold both though Amazon itself and through Marketplace vendors (together this is termed gross marketplace value, or GMV) are generally about double Amazon’s reported revenue. In other words, Marketplace means that Amazon handles (but does not, incidentally, itself set prices for) double the share of ecommerce that it reports as revenue.

Meanwhile, when AWS launched the general consensus was that this must be burning cash and that it was another example of the idea that the company operates by buying market share at breakeven or a loss. At a certain point, though, AWS became big enough that financial regulations required Amazon to split it out in its reporting, and we discovered that it was profitable - it now has a 25% operating margin. Then the narrative inverted - the cash from AWS was supposedly subsidizing the loss-making purchase of market share elsewhere in the business.

In my view, both narratives were based on a false premise. As I discussed in detail in this post a few years ago (written before AWS was broken out), those atomised teams are all at varying stages of development - some are large and some small, some old and very profitable, some new and making startup losses. The net income and free cash flow numbers you see are the aggregate of all of those hundreds of teams, and so don’t really tell you anything much - Amazon invests cash from profitable units into the creation of new, unprofitable units, and you have no real idea what the distribution looks like. This, I think, is how we should see both AWS and the marketplace business: Amazon is uniquely obliged to disclose the profitability of AWS, but it’s not the only profitable part of the company.

Prime fits into this, incidentally, as a third pillar, next to logistics and ecommerce. Every piece of perceived value that Amazon can add to Prime makes you more likely to sign up and less likely to cancel, and once you have it, as a sunk cost, you are much more likely to route other online (and increasingly many offline) purchases though it. The best parts of Prime, from Amazon’s point of view, are things with no marginal cost, such as TV. Amazon buys TV shows to get you to buy toilet paper, and your shifting your toilet paper buying from Walmart to Amazon has an LTV that dictates the TV acquisition budget.

Amazon, then, is a machine to make a machine - it is a machine to make more Amazon. The opposite extreme might be Apple, which rather than radical decentralization looks more like an ASIC, with everything carefully structured and everyone in their box, which allows Apple to create certain kinds of new product with huge efficiency but makes it pretty hard to add new product lines indefinitely. Steve Jobs was fond of talking about saying ‘no’ to new projects - that’s not a very relevant virtue to Amazon.

For both Amazon and Apple (and indeed Google or Facebook), this means that there are certain kinds of project that they can deliver very well and very repeatably and predictably, but also, crucially, that there are certain kinds of project that they are much less well suited to deliver. Google doesn’t tend to be better at cloud platforms than Apple and worse at UIs because there are better or worse people in each team, but because each company is set up to deliver certain kinds of things, and the closer a project is to that machine’s direction the more reliable the result. If the machine is designed to do X, it will struggle at Y no matter how clever the people. A lot of the story of Amazon for the last 20 years is of how many Ys turned out to be Xs - how many categories that people thought could not be sold online and could not be sold as commodities turned out to be both.

AmazonBenedict Evans