OpenAI opened a marketplace on 29 September where enterprise customers can pay for other companies’ AI products out of money they have already committed to OpenAI. Thirty-two partners were live at launch, announced at the company’s DevDay event. For those partners, the shop is also a competitor’s front door.
The most detailed reading so far comes from Software Synthesis, a newsletter that published its analysis on 5 October 2026. Its argument is a forecast, not a report of what OpenAI has decided. The newsletter’s version: OpenAI wants to be the place where the rest of the AI industry gets built and sold, much as the big cloud providers became the place where companies bought business software. The rest of this piece follows that argument and flags where it is speculation.
The launch facts are thin. Eligible customers can steer part of an existing commitment toward the listed products. The partner the newsletter singles out is Baseten, which serves open-weight models (models anyone can download and run). Through Baseten, OpenAI customers can now spend their OpenAI budget on those open models inside Codex or through the Responses API. The newsletter reads that as OpenAI accepting that it will not capture every token a customer buys, in return for keeping the buyer close.
That reading has numbers behind it. Software Synthesis cites Ramp’s AI Index, which puts open-weight models at 5 percent of business spend, and says OpenAI and Anthropic have been cutting prices to hold on to customers. It also cites Suger, a startup that helps vendors sell through cloud shops, estimating that enterprise buyers hold roughly $348 billion in committed cloud budget. The newsletter puts OpenAI near a $70 billion revenue run rate, with more than half from enterprises.
Today the plumbing is plain. Each vendor still invoices the customer separately, and OpenAI reconciles the accounts. As far as the newsletter knows, there is no listing fee and no joint selling. Its forecast is conditional: if the marketplace gains traction, the cloud pattern could repeat, with labs and apps selling together, apps paying a fee to be listed, and buyers receiving a single invoice. That is a hypothesis about a market that has existed for a week.
The cost to the listed apps is old and familiar. Selling through the company whose models you depend on, or compete with, is among the oldest bargains in software. The vendor gets a buyer who can spend without a new procurement fight. The shop owner gets a cut, sees which products sell, and holds the customer relationship. The newsletter’s cloud analogy implies exactly that fee structure, though nothing in its reporting says OpenAI has set one.
The author sketches two outcomes. In the bullish one, OpenAI gains enough market power to require more use of its own models as a condition of listing, the way cloud providers profit however customers divide their spend. In the bearish one, buying through OpenAI turns out no easier than buying direct, and most reallocated money flows to Baseten’s open models, where OpenAI earns nothing from the tokens yet still owns the customer. Both cases hinge on OpenAI having enough enterprise reach to justify charging for access.
History argues for caution. Plugins, the GPT store, and the apps SDK were earlier attempts to turn OpenAI’s reach into a platform of other people’s products, and all of them lived inside ChatGPT. The newsletter notes that this attempt differs, because partners attach to OpenAI’s contracts and APIs rather than its chat interface. It also argues that Anthropic is less likely to copy the move, since Anthropic competes more openly with its own customers.
The newsletter expects a verdict by the first half of 2027, when OpenAI’s S-1 filing, the document a company files with US regulators before going public, is expected to land. Any AI application vendor weighing a listing before then should ask three things: whether a fee is coming, whose name goes on the invoice, and what happens to the customer data.
Analysis from Software Synthesis, published 5 October 2026.
