When Does a Middleman Become Infrastructure?

In the middle of 2025, JPMorgan Chase told the companies that connect financial apps to bank accounts that access would no longer be free. The bank had received 1.89 billion data requests in June alone, and said only about 13 percent of them were tied to something a customer had actually just done.¹

The largest of those companies is Plaid. If you have ever linked a bank account to a budgeting app, a brokerage, or a payment service, you have probably used it without knowing. By 2025 Plaid sat inside thousands of financial products. It was, by any reasonable definition, infrastructure.

JPMorgan did not cut it off. In September 2025 the two companies announced a renewed agreement with a new pricing structure, under which Plaid pays for access it had previously received for nothing.¹ CNBC reported the cost could reach $300 million a year, which Plaid has said it will absorb rather than pass to its customers.¹ A former regulator's summary of the moment was blunt: the biggest player had already caved.

That outcome is worth sitting with, because it is better than it sounds and worse than it looks. Plaid was too embedded to remove. That protected the company. It did not give the company any say over what it pays.

So what actually turns a middleman into durable infrastructure, and what does durability buy you once you get there? We looked at companies that started life inside somebody else's transaction, across payments, advertising, travel and commerce, and traced what happened when the party above them changed the rules.

Being hard to replace is not enough

The intuitive answer is that you become safe by becoming deeply wired into your customers. Make yourself expensive to rip out. Two companies show why that reasoning breaks.

Marqeta

Marqeta builds the plumbing behind payment cards. When a fintech company issues a card, someone has to authorize transactions in real time, and Marqeta was that someone for Block, the parent of Cash App and Square. At the point its contract came up for renewal in 2023, public filings and analyst coverage put Block at roughly three quarters of Marqeta's revenue.²

Replacing Marqeta would have been genuinely painful for Block. Morningstar assessed at the time that a failure to renew could have rendered Marqeta's business model unworkable, which tells you how deep the integration went in both directions.²

Block renewed. It also cut Marqeta's take rate by roughly 40 percent on a gross margin basis. Marqeta's net revenue fell from $676 million in 2023 to $507 million in 2024.² The switching cost stopped the customer from walking. It did nothing to stop the customer from renegotiating.

Segment

Segment did something different and, on paper, even stickier. It collected the data a company generates about its own customers and routed it to the hundreds of other tools that needed it. Analytics, email, advertising, support. More than 700 prebuilt connections, one of the largest integration footprints anyone has ever assembled.³

Twilio bought it for $3.2 billion in 2020, then wrote down $285.7 million of the acquisition in late 2023, and split the business into two units in a 2025 restructuring.³

Here is the difference that matters. Every one of those 700 connections joined a customer to that customer's own tools. When cloud data warehouses got cheap and good, a company could decide on its own to do the routing itself. Nobody else's permission was required. No competitor had to agree. No bank, no platform, no regulator.

Infrastructure is a coordination problem you solved

The companies that held their position had something structurally different going on. Removing them would have required several parties who do not trust each other to agree on a replacement.

The clearest case is AppsFlyer.

Companies that advertise mobile apps need to know which ads actually produced installs. The trouble is that the advertising platforms selling those ads also count the results. Google reports its numbers, Meta reports its numbers, and no advertiser is willing to take either at face value. So the industry settled on independent measurement companies, and AppsFlyer became the largest of them.

For years that business rested on an identifier Apple assigned to every iPhone. In April 2021 Apple started requiring users to opt in to being tracked, and roughly 96 percent of U.S. iPhone users initially declined.⁴ The technical foundation of the industry disappeared more or less overnight.

AppsFlyer's position survived intact, for two reasons that had nothing to do with the identifier. Its software was already built into 15,000 companies' own apps, so removing it meant each of those companies rewriting their own product. And the job it did, acting as the referee that competing platforms could all live with, became more valuable once reliable measurement got scarce rather than less.

In June 2026, the position became unusually visible. Google, Meta, Unity and Moloco together put more than $1 billion into AppsFlyer at a $2.7 billion valuation. All four took minority stakes with no control rights, no exclusivity, and no preferential access to the measurement data.⁴ Four companies that compete ferociously with one another paid, in effect, to keep the scorekeeper neutral. Chief executive Oren Kaniel put the logic plainly: attribution and measurement have to be independent, unbiased and trusted.⁴

AppsFlyer's protection was never the identifier Apple took away. It was that its customers' suppliers, and its suppliers' competitors, all needed a referee they could agree on.

Two sides, counted separately

Which brings us back to Plaid, and to the thing most companies get wrong when they assess their own exposure.

Plaid has thousands of customers. No single fintech customer appears to define the business. On that side, the dispersion is real and it is protective.

On the other side sits a small number of very large banks that control access to the accounts. That is the side where Plaid got taxed, and it got taxed precisely because that side is concentrated.

Cardlytics makes the same point in a harsher register. It ran the offers you see inside your banking app, and its top three bank partners accounted for over 80 percent of the money it paid out to partners in 2025. Bank of America gave notice in April 2025 and the relationship ended in February 2026. Second quarter revenue in 2026 came in at $36.9 million against $58 million a year earlier, and the number of monthly qualified users fell by 39 million.⁵ One departure.

Marqeta had the same shape on its customer side. So the rule generalizes, and it is worth stating in the plainest terms available. Thousands of customers do not protect you from three suppliers. Diversification only counts on the side where the revocation right actually sits.

So cut out the network?

If a concentrated upstream party controls your access, the obvious move is to go around it and contract with the underlying suppliers yourself. Sometimes that works. It almost never works the way people expect.

The Trade Desk is the best-documented attempt. It buys digital advertising for its clients, and historically it reached publishers through a layer of intermediaries called supply-side platforms, which typically take somewhere between 15 and 20 percent.⁶ In 2022 the company launched OpenPath, letting publishers connect straight to it and pay a flat fee instead.⁶

Four years on, roughly 300 publishers have taken the direct route, including Reuters, The Washington Post, Gannett and Condé Nast.⁶ Every one of them kept their existing intermediary relationships. The Trade Desk's own materials note that the top 20 publishers each connect to an average of 24.5 supply-side platforms.⁶ The company said publicly at the time that it was not trying to remove the intermediaries, and one of its executives conceded early that direct connections make sense for large, technically capable publishers while smaller ones are better served by the existing layer.⁶

This is not a half-finished project. It is the pattern that repeats in the successful cases.

Why a supplier says yes

Suppliers do not go direct because you asked nicely, and they do not go direct for a marginally better rate, because the incumbent can always cut its rate. They go direct when you offer something the incumbent structurally cannot.

Zillow is the cleanest illustration. It used to receive property listings through a syndication service called ListHub. In December 2014 ListHub's parent was acquired by News Corp, which also owns Realtor.com, Zillow's principal competitor. Zillow was now getting its most important input from a rival.⁷

It declined to renew and went to the regional listing organizations directly. What it offered them was not money. It was control: a dashboard giving brokers and agents authority over where their own listings appeared and how they were reported. A syndicator, by definition, is the thing that takes that control away. It could not match the offer without dismantling itself.

The other side responded in kind, terminating the feed to Zillow's newly acquired Trulia on five business days' notice, which Zillow blocked with a restraining order.⁷ By March 2015 Zillow had signed 25 additional listing organizations directly and said two of the three largest had agreed to send feeds.⁷

The Trade Desk's publishers went direct for a related reason. They were not offered a discount on a commodity. They were offered access to advertising demand they could not reach any other way, on economics they could actually see.

Two ways it goes wrong

American Airlines used force

Corporate travel gets booked through systems that travel agencies and company travel departments live inside all day. American Airlines decided to push bookings toward its own newer channel, and from April 2023 it withheld more than 40 percent of its fares from those systems.⁸

Corporate agencies reported bookings down around 40 percent. Meanwhile Delta and United both reported corporate bookings up 14 percent in the first quarter of 2024.⁸ The buyers had not followed American to its new channel. They had booked competitors whose flights were still available in the system they already used.

American posted a first-quarter loss despite record revenue, cut its guidance, and on 28 May 2024 announced the departure of the executive who had designed the strategy. Chief executive Robert Isom conceded the company had moved faster than it should have, and the following day framed the correction with unusual candor: they had used a lot of sticks, and needed to put some carrots in place.⁸

The lesson holds well beyond airlines. If the network owns your customer's daily workflow, withholding your supply from it does not pull the customer toward you. It hands them to whoever stayed.

Spotify counted the cost

In September 2018 Spotify began testing something that looked overdue: letting independent artists upload music directly to the service, keep ownership, and take a larger cut. It shut the program down by 30 July 2019 and sent participating artists back to approved distributors.⁹

The company's stated reason was that its distribution partners were already doing the work well.⁹ That work turns out to be considerable. Distributors verify who owns what, clean up the metadata that makes a track findable, handle infringement claims, and calculate and pay royalties. Removing them meant building all of it, with real legal exposure attached, at a moment when Spotify was heading into licensing renewals with the major labels. Contemporary coverage found the timing suggestive, though Spotify never confirmed a connection.⁹

This is the most commonly underestimated number in any disintermediation plan. The fee you can see is not the cost you would take on. Whatever the network quietly does for contracting, settlement, fraud, compliance and dispute resolution becomes your job the moment you route around it, and it becomes your job for every supplier you take direct.

What the cases have in common

Underneath all of it are two questions that companies tend to collapse into one.

Plaid answered the first well and the second badly. So did Marqeta. Segment could not answer the first at all, which is why the second never came up. AppsFlyer is the only company here that answered both, and it managed the second one because the parties with an interest in repricing it could not agree among themselves on who would take its place.

1. Removing you recreates a coordination problem

   Not a technical migration. An actual negotiation among parties who would have to agree on what replaces you, and who have reasons not to trust each other's answer.

2. No upstream party controls too much of your economics

   Count each side of your business separately. Thousands of customers do not offset three suppliers who can turn you off.

3. You directly hold the supply that carries the value

   Not all of it. Enough that losing the network would be survivable, and enough that the network knows it.

4. The network keeps the long tail, permanently

   Nobody in this study eliminated their upstream layer. They demoted it to the part that was never worth owning.

A company standing at this fork now

Wildfire Systems is a San Diego company that builds the shopping rewards programs banks and fintechs put their own brand on. When a customer of one of those banks shops online and earns cashback, Wildfire is usually the layer making that work. It reaches merchants largely through affiliate networks, chiefly CJ, Impact.com and Rakuten Advertising, and its pitch to customers is that they get access to more than 50,000 merchant programs without onboarding a single one.¹⁰

Two positions sit inside that company, and they are not equally durable.

The browser extension business is the rented one. It operates under rules set by Google's extension store and by the affiliate networks, and both sets of rules tightened during 2025 and early 2026 after a competitor's conduct drew scrutiny. Google updated its extension policies in March 2025. Rakuten Advertising terminated PayPal's Honey from its network in January 2026, with Awin and Impact.com taking their own action.¹¹ Honey had roughly 20 million users and had been bought for about $4 billion. None of that mattered, because the right to participate was never Honey's to hold.

The second position is RevenueEngine, which lets AI shopping applications monetize the purchases they generate. Structurally this is a much better place to stand. It sits inside the customer's own product rather than in a browser someone else governs. It solves a real coordination problem, connecting many AI applications to many merchant programs, which nobody wants to solve one relationship at a time. And if agentic commerce needs a commission attribution layer that the platforms can all accept, that is the AppsFlyer position.

The unresolved part is upstream. If merchant access still ultimately runs through a small number of affiliate networks, then Wildfire could become genuinely difficult for its customers to remove while remaining dependent on suppliers who set the terms. That is the Plaid problem, and Plaid is a good outcome rather than a bad one, but it is an expensive one.

The history says Wildfire does not need direct relationships with 50,000 merchants. Nobody in this study built anything close to that. The question is narrower and more answerable: how concentrated is the commission volume?

If a few hundred merchants drive most of it, the move the record supports takes two sentences. Keep the networks for breadth. Own the relationships that carry the economics.

Those merchants would still need a reason to sign, and it cannot be a better rate, because a network can always match a rate. The plausible reason is attribution inside AI shopping flows. A merchant that wants to be found and correctly credited when an agent does the buying has a problem the traditional affiliate-network model was not designed around, and solving it is the kind of offer an incumbent cannot match without rebuilding itself. That is what Zillow did with control, and what The Trade Desk did with demand.

If instead the volume is genuinely spread across tens of thousands of merchants, no company examined here has escaped a position of that shape, and the sensible strategy is not escape. It is becoming valuable enough that the networks would rather charge you than lose you.

The public record does not disclose which of those is true. It is also worth noting that Wildfire currently reads to those networks as a source of new demand rather than a threat. CJ's chief executive publicly praised a Wildfire product launch in March 2025, and Rakuten Advertising has published a favorable interview with a Wildfire executive.¹⁰ That is the safest available position for a network-dependent company, and it is also the thing that changes fastest once you start moving existing volume around them.

What survives

Across every case here, the platform stayed in the value chain. Apple did not stop mediating iPhone advertising. The banks did not stop controlling account data. The affiliate networks, the listing syndicators and the travel distribution systems are all still there. What changed was how much of them any given company actually needed.

That is the useful version of the question. Not how to remove the layer above you, which mostly does not work, but which specific part of it you cannot afford to have someone else control.