The switchboard is becoming the AI product

The telling detail in Runway’s new product is that the user is no longer meant to care which model did the work.

TechCrunch reported that Runway has launched an AI model router as generative media gets crowded. That sounds like plumbing. It is actually a product strategy.

The obvious read is that generative media is filling up with model choice, so Runway needs a way to stay useful when customers want options. The sharper read is that model choice itself has become a burden. If every product team has to compare quality, latency, price, failure rates and policy behaviour across a rotating cast of models, the “best model” stops being a fixed answer. It becomes a decision made at runtime.

That decision layer is becoming the product.

Stripe reportedly circling OpenRouter makes the point louder. Techmeme linked a Wall Street Journal report saying Stripe is in talks to acquire OpenRouter, a startup that helps developers use AI models, in a deal that could fetch around $10B. Strip away the acquisition drama and the strategic logic is plain: whoever sits between developers and models can measure usage, route requests, abstract suppliers and turn inference into a billable flow.

That is Stripe’s home turf. Payments companies do not need to manufacture the thing being bought. They win by sitting at the moment of transaction, making it reliable, measurable and easy to charge for. AI inference is starting to look less like software licensing and more like payments: small events, priced dynamically, routed through infrastructure, reconciled later.

The model is becoming a supplier

This is uncomfortable for model companies because it changes the source of power. If the user interface, billing relationship and routing logic live somewhere else, the model risks becoming one supplier in a marketplace. A very expensive supplier, perhaps. A technically impressive supplier. But still a supplier.

Meta’s new coding move fits this tension. Reuters via Yahoo Finance reported that Meta launched a new AI coding tool powered by Muse Spark 1.2. On the surface, that is another model company entering one of AI’s clearest commercial categories.

But coding is also where switching costs may move away from the model fastest. Developers do not wake up wanting Muse Spark 1.2 in the abstract. They want a bug fixed, a test written or a pull request reviewed. If those actions happen inside a workflow that can swap models beneath the surface, the brand on the model matters less than the reliability of the system around it.

GitHub seems to understand this. Its Copilot SDK, a multi-platform SDK for integrating GitHub Copilot Agent into apps and services, points to Copilot becoming an ingredient embedded in developer tools rather than a feature trapped inside one interface. That matters because the battleground shifts from “which assistant has the best chat window?” to “which agent can be inserted into the places work already happens?”

The historical parallel is the travel industry. Airlines still matter, but the booking layer gained enormous power by comparing routes, prices and availability across carriers. Customers usually buy the trip, not the aircraft. In AI, the router is the booking engine, the meter and sometimes the travel agent.

For builders, the lesson is blunt: do not hard-code your product strategy to model loyalty. Build for routing, observability, fallback paths and unit economics from day one. The winning AI product may not be the one with the flashiest model. It may be the one that knows when to use which model, how much it costs and who pays.

---