There's a classic distinction between one- and two-sided markets. A one-sided market is something like Twitter, where everyone can be a follower or get followed, and there aren't distinct categories of users. A two-sided service might be something like Booking.com, where the hotels (and other travel-service providers, e.g. airlines) and travelers are distinct groups, or like Tinder, where men will join the site with the most women and vice-versa. There are some markets that are, in a sense, many-sided—medical providers getting paid by health insurers whose relationship with the policyholder is mediated by their employer, for example—but in these cases, one of the economic jobs of the intermediaries is to simplify things to the point that "how can I get this weird rash looked at?" is not a combinatoric nightmare.

But there's also a strange case I like to think of as 1.5-sided markets, ones where it's helpful to have both sides of a transaction on a platform, but not strictly necessary. One case of this is bond trading: there's a minor economic puzzle here, in that most kinds of trading moved from voice to electronic quickly once it became obvious that having two people talk to one another on the phone was not the highest-bandwidth, highest-fidelity, lowest-latency way to transmit bits of information like "I want to buy 100 shares of Apple at no more than $205.14/share." Bonds have been comparatively slow to move online, because both sides of the network need to care enough to make it happen.

In bond trading, it's tempting to look at things as a one-sided network, where everyone is there to exchange bonds for cash or vice-versa. But in practice, there are two kinds of participants: liquidity takers trade when it occurs to them that they ought to make a transaction, and liquidity makers trade when someone asks them for a quote. If the market functions based on phone calls, it doesn't go electronic until a critical mass of liquidity makers are on a given platform, or a critical mass of liquidity takers insist that they prefer transacting digitally. Liquidity makers are 1) much bigger participants than the average taker, and 2) beneficiaries of a system where nobody is entirely sure what the fair value of a given bond is, and it's inconvenient to canvass lots of trading desks and find out for sure. So the nodes in the network that are most densely-linked are the most reluctant to change protocols.

But it's at least hypothetically possible to have a setup like this:

1. You use some text-to-voice service to call up every bond dealer in parallel and ask them for a quote

2. You hop on the call with whoever gave the best price, and execute the trade.

Suddenly it's a 1.5-sided market! One side can move over, the other side doesn't have to (but will probably find that it's a very good idea to do so). And this kind of market shows up in other places: Affirm wants to be offered by merchants at the point of sale so they get the merchant as a user, but there are cases where they're happy to get a transaction even if they don't fully monetize it, so they offer a virtual card that can be used like a typical credit or debit card. So, they're giving you part of the network but not the whole thing. When you type the name of a movie or show into Netflix search, at first Netflix matches to movies with similar titles—but once it's clear that you're looking for a specific movie they don't actually have, they switch to showing you results for similar ones! Type "wall" and the top results are Wallace & Grommit, but add "street" and suddenly the #1 result is a documentary about Bernie Madoff (with "Wall Street" in its name, to be fair), but the #2 result is The Founder, which doesn't focus on the stock market at all but which is a movie about business dealings and betrayal. Netflix can't get every single movie you might be looking for, but once it knows exactly what you want, it can easily offer you approximately what you want instead.

Social networks solve the onboarding problem in a similar way: long-term, LinkedIn wants you connecting with peers and colleagues. But when you sign up, you might not have very many of either! So they want to connect you to business-adjacent influencers instead, who will at least give you something to look at and interact with when you log in.

Many of these 1.5-sided networks get easier to execute once you're not limited to the actual space of content, connections, partners, etc., but can explore the latent space of hypothetical ones instead. That's basically what generative AI is, in the end: a statistical guess about what could exist, based on what already does exist.

Some networks are moving in the opposite direction: there are many domains where models still underperform humans, but a growing share where they win, and one of the things they'll sometimes outperform people at is authentic-feeling emotional connections. They can also fill in the smaller gap between the medium in which you want to consume something and the medium in which it actually exists, as in the bond trading example. This, too, can move in the opposite direction: for many early Uber users, part of the appeal was getting a ride without having a conversation with someone, but some users actually want that, so if you want to, Uber will let you call 833-USE-UBER to basically tell someone to use the app for you.¹

Networks win when they fill in gaps, and the default gap is a missing node in the network. But there's no rule stating that this node needs to have the same form or substance as the rest of the network, and as it gets easier to emulate human behavior automatically, networks built for humans will increasingly be able to achieve critical mass when they're still populated mostly by bots.

The Diff has used the 1.5-sided network model many times:

• Building an incomplete graph ($) was an early look, before AI tools got big (and includes thoughts on DoorDash burning cash by offering delivery from restaurants that hadn't signed up for the platform yet.

• This piece on MarketAxess ($) looks at how hard it's been to make bond trading electronic.

• We looked at Google using AI to call businesses on your behalf ($).

• DoorDash does this, too ($).

• And there's a startup that tries to reduce the sidedness of Airbnb by promoting house swaps ($).

Join the discussion!

1  "Someone" was probably human when the feature launched, but this is exactly the kind of customer service task that can be replaced by AI in most cases, with humans sticking around for escalation.