Product to Platform: When to Open Your Product to Third Parties

Why a platform is a price agreed with somebody not in the room

A platform, as this page uses the word, is a product other people build on and sell through. A product becomes one when a third party can earn more by building on it than by building alone. That is a fact about two businesses, and the meeting where the platform is approved usually holds only one of them.

So the room picks a rate from a phone platform and argues about whether developers will accept it. Neither number that would settle the argument is on the table: what the platform layer costs on its own, and what a builder earns without you.

Two rates and the corridor between them

The method turns the decision into two rates. The first comes from your costs. The second comes from the builder's business. The gap between them is the corridor, and the rate you charge has to sit inside it.

In the worked case used in the Platform Bet Strategy Pack, a fictional fleet-software company with 38,000 customers arrived with a 30 percent rate copied from a phone and a $21m layer cost that was the integration team's budget. Costed on its own, the layer needs 8.0 percent; the allocation had put it at 18.6. A builder stops building at 22.5 percent. So 30 percent earns nothing, because the volume it is charged on does not exist, and 15 percent earns $10.7m a year net of the layer and the $5.0m of contribution it displaces.

The weak input is the builder's alternative, which came from one conversation. The answer survives a 25 percent error in it and not a 50 percent one, which is the reason to ask more than one builder.

A waiting market it cannot reach alone

The first thing a builder needs is customers already standing on your product. Without them, a lower rate or better tools change nothing.

Salesforce introduced AppExchange at its Dreamforce conference in September 2005, years before Apple's consumer App Store. The pitch to developers was to stop building standalone software nobody could find and build where a waiting market of customers already stood. By the end of that first fiscal year AppExchange held 575 apps from 250 independent software vendors. As of December 2025 it hosted 6,233 apps from 3,668 developers, and 91% of Salesforce customers used at least one.

Amazon's first two attempts failed because they sent buyers somewhere else. Amazon Auctions launched in March 1999 and zShops in September 1999; of Auctions, Jeff Bezos later wrote, "I think seven people came, if you count my parents and siblings." Marketplace, launched in November 2000, put the third-party listing on Amazon's own product page next to Amazon's own offer. Within a year it accounted for roughly 5% of units sold, and as of the first quarter of 2026 third-party sellers accounted for 60% of paid units on Amazon.

Walmart opened Walmart.com to third-party sellers in 2009 and held most of them at a gate for thirteen years. In August 2022 Walmart replaced the Request to sell button with an open flow, and nearly 1,000 new sellers a day joined that week. Walmart Marketplace crossed 200,000 active sellers by mid-2025. A September 2025 CNBC investigation found Walmart had made vetting "more lax in a bid to compete with Amazon," with sellers impersonating legitimate businesses.

A layer that works without you

The second need is a layer a builder can use without your help. It is expensive, it is built ahead of the demand that justifies it, and its absence is hard to buy back later.

Nvidia introduced CUDA in November 2006 and kept building general-purpose computing into gaming chips whose buyers neither needed it nor paid for it. Nvidia's own executives have said it took roughly ten years before Wall Street believed the investment was worth anything. In 2012 AlexNet was trained on Nvidia GPUs through CUDA, and by Nvidia's fiscal 2025 filing over 5.9 million developers used CUDA and its other software tools.

AMD bought ATI's graphics business in October 2006 for about $5.4 billion in stated value and did not fund a comparable software layer. AMD's answer, ROCm, launched in 2016, a decade behind CUDA, and AMD's fiscal 2025 filing still names Nvidia the share leader in discrete graphics. AMD owned world-class silicon; Nvidia owned the layer developers had written for.

Hugging Face started as a consumer chatbot with roughly 100,000 daily users. In 2019 Hugging Face retired the chatbot and made its free library and Hub the company: it stopped building a product and started building the place other people build theirs. In August 2023 Hugging Face raised $235 million at a $4.5 billion valuation.

A rate that leaves the builder better off

The third need is the rate. Apple, Salesforce and Shopify converge on 15 percent for most developers, and the cases show both edges of the corridor: rates builders fight, and a low rate that brought no one.

Salesforce takes 15% of revenue on apps sold through AppExchange Checkout, with no minimum and no setup fees.

Shopify let developers keep all of their app revenue up to $1 million, with the exemption reset every year since August 2021. From June 16, 2025 Shopify made the $1 million a lifetime cap and charges 15% on everything above it. Of 40,556 registered Shopify partners in January 2025, 7,874 had an app listed, and 79.5% of developers list only one app, so the first million is most of what small developers will ever see.

Apple's 30% commission is the rate builders fight. On August 13, 2020 Epic routed Fortnite purchases around Apple's payment system on purpose, to trigger Fortnite's removal and sue; iOS was only about 7% of Fortnite's revenue. After a 2021 injunction required Apple to allow external payment links, Apple charged 27% on linked-out purchases. In April 2025 a federal judge found Apple in willful civil contempt, and in December 2025 the Ninth Circuit agreed the rate was "large enough that no rational developer would offer them."

A low rate does not create builders by itself. The Epic Games Store launched in December 2018 taking 12% from developers against Steam's 30%, yet between 2019 and 2024 its third-party revenue moved only from roughly $251 million to $255 million. A rate inside the corridor is necessary; it does not supply the waiting market.

A promise not to compete or close the door

The last need is governance rather than arithmetic: a builder has to believe you will not take its business or shut it out once it has built.

In 1985 Bill Gates and Jeff Raikes urged Apple to license Mac technology to three to five major manufacturers. Apple's Jean-Louis Gassée killed every deal to protect hardware margins above 55%. When Apple later allowed Mac clones, its CFO Fred Anderson told Wired in 1997 that "somewhere around 99 percent" of clone sales went to the existing customer base. Steve Jobs killed the clone program by shipping Mac OS 8 in July 1997, which clone makers were not licensed to ship, and Apple bought Power Computing's core assets for $100 million in Apple stock that September.

The Apple case holds the other end of the method. Opening moved Apple's own sales to someone else rather than adding buyers, which is the displaced contribution the rate the layer needs is meant to count. And the door, once closed, ended the builders' business.

The promise is not absolute. Amazon's Marketplace put third-party sellers against Amazon's own retail managers on the same page, and sellers came anyway, because the page was where the buyers were. Hugging Face went the other way and stopped competing with its users altogether. Which promise a builder needs depends on what it can earn without you.

What these cases show

Each case is drawn from the Stratrix case library and checked against the filings and reporting listed under Sources. The last column says what each one does not establish.

Running the decision

What this method cannot tell you

Platforms that worked are studied and the layers nobody built on are forgotten, so any set of cases is selected on its outcome. The cases above show what a builder needs. They cannot support a success rate for platform bets, and this page does not claim one.

The arithmetic has limits of its own. It prices one representative builder, and a platform has many. It holds the waiting market at maturity, which a real platform reaches over years. It does not price what a platform does to your own product's retention, which is a judgment rather than a number. And it says nothing about the promise not to compete or close the door, which is necessary and is governance rather than arithmetic.

Questions