Two goods are economically complementary when having more of one of them leads to more demand for the other. There are plenty of didactic examples: bread and peanut butter make jelly more useful, for example. But the PB&J system actually illustrates some of the varieties of complementarity: you need a spoon and a knife, so those are complements, too. But more spoons and more knives won't lead to more sandwiches; past some threshold (in this case, one of each) they don't provide much more complementarity. But the raw materials do keep scaling: whether you're making dinner or managing a manufacturing facility that produces Uncrustables, you want roughly constant proportions of each ingredient.

For some applications, the complementarity is even fuzzier. Back to the kitchen, some ingredients like butter and salt will show up almost everywhere. Others are closer to one-off, and some things show up in a large number of different dishes. So each food item, and each piece of kitchen equipment, has some amount of complementarity with everything else.

Complementarity matters a lot in tech. One of the drivers of growth at the sector level is that many products complement one another at the product level: if there are smarter phones, there will be more apps; if there are more apps, there will be even smarter phones. This was also a driver of the more recent runup in AI: R&D budgets for designing chips and capex budgets for building them only make sense in light of demand from companies building AI models, and that demand is reasonable because there are so many more consumer use cases.¹

Some things are more complementary than others. In labor markets, a good general rule is that an increase in the supply of low-skill labor pushes wages down (if there are more people out of work, someone will mow your lawn or watch your kids for less than otherwise). But in high-skill markets, supply creates its own demand. The flow of talent into software created more demand for a category that previously would have been considered incomprehensible and is now a massive market: software tools that make software companies more efficient at building software. And of course, these tools create their own new source of demand because the total possible number of integrations between products is equal to products-in-use multiplied by that number minus 1, all divided by two.²

Some products are almost universal complements: "information," as a category, is a way to make many other inputs better (though it's tricky to treat it as a discrete category, since the theoretical definition of "information" makes it by definition the most heterogeneous good around). Intermediaries also tend to be a universal complement, and a tool for increasing complementarity: any time shopping data indicates that purchasing product X correlates with purchasing product Y, their somewhat complementary nature becomes an emergent property of the recommendation algorithms that will cause them to be bought together, the logistics algorithms that discover that they ought to be shipped together, etc.

One place where complementarity shows up a lot, in fact, is in recommendation engines: every e-commerce provider, and every SaaS company with a long menu of add-ons and "call us for" pricing, is implicitly making decisions around the complementarity between goods. For e-commerce, that's a second-order issue because the main question is whether buying one product predicts a higher propensity to buy another when offered, whether or not they're actually additive—gasoline and lottery tickets don't make one another more useful, but they do get bought together. But in B2B, identifying valuable complements has a different payoff function, those complements both give the company something to cross-sell and ensure that their customer has to churn off of two products instead of one. If chosen right, the whole becomes greater than the sum of its parts (the platform becomes much more valuable to a customer’s workflow etc.), which makes churning less palatable or feasible.

In general, being in the business of identifying new complements is a good one; it's the kind of thing that leads to compounding over surprisingly long periods. But that's also the kind of characteristic companies have a strong incentive to fake: in a stylized sense, the most promotional companies always promise investors that they're directly tied to a hot theme, and that they get disproportionate benefit from it relative to everyone else. So, complement-spotting can lead to plenty of adverse selection. But it also rewards people who think deeply about a given business: when there's a complementary relationship between companies in very different businesses (chip companies and software companies, for example, or, in an earlier era, automakers and the oil business), it's entirely possible that neither side has the whole picture and even the people actively making it happen underestimate the size of what they'd stumbled on.

The Diff considers the economics of complements and substitutes in many different contexts. We’ve looked at:

• How hybrid work changes the elasticity of oil ($).

• Some kinds of economic growth accelerate other industries, and some kinds come at their expense.

• A great place to explore this is the economics of factory electrification.

• Complementarity is why part of the impact of fast food mobile apps shows up in real estate prices ($).

Information as a product is a universal complement and universal substitute.

Join the discussion!

1  From Dylan Patel on the Dwarkesh Podcast: “There's a strong argument that funding the next node wouldn't be economically viable anymore if it weren’t for AI taking off and generating humongous demand for the most leading-edge chip.” The dollars coming from smartphone growth were beginning to peter out before LLMs saved the day.

2  This also creates a nice form of on-the-job training for junior developers, since it's fairly straightforward to scope out what an integration looks like—the newly-signed contract goes into a Dropbox here and leads to a slack message there—but it also offers good practice at some combination of reading documentation, dealing with busy third parties, or using trial and error.