Meta is reportedly working on its own AI model router, similar to the OpenRouter service, which has gained traction as businesses look to offload some tasks to cheaper AI models.
OpenRouter aggregates a large number of AI models and allows businesses to decide which tasks should be sent to each one. It handles billing, usage tracking, and request routing, making it much easier for developers to reduce costs and use the best model for a particular task.
According to The Information, the project is codenamed Switchboard and is intended to help Meta reduce its internal AI costs. The project is currently being handled by Meta’s incubator, AAI Labs, which pitches and builds prototypes of AI-powered products for internal use.
While AAI Labs is primarily focused on internal projects, it leaves open the possibility for other teams to develop them for public use. It reportedly has around 200 approved projects spanning the company’s technology stack, with some eventually pushed to public release.
The switch from tokenmaxxing to cost cutting
Meta was one of the first tech companies to encourage employees to spend as much as possible on AI, an approach that became known as tokenmaxxing.
However, it has since shifted, like the rest of the industry, toward reducing costs. Researchers had been burning through millions of dollars in AI costs, to the point that Google reportedly limited Meta’s access to its own AI models because of overuse.
The social media giant has increased its planned capital expenditures for 2026 by almost 100 percent if it spends the $145 billion at the top end of its forecast. While it cuts back on internal AI spending, it is expanding the amount of AI infrastructure it operates and accelerating the timeline for launching new versions of its custom AI chip.
That expenditure may be partly funded by Meta selling excess data center capacity, either to partners or rivals through a neocloud platform. Following the success of SpaceX, which has signed deals with Anthropic, Google, and Reflection AI totaling $27.6 billion per year for access to its Colossus data centers, Meta may be able to generate even more income by selling its own capacity.
However, this would reduce Meta’s ability to use that capacity to run its own AI models and services, which has been the primary aim behind CEO Mark Zuckerberg’s infrastructure buildout.
AI pushing Meta to explore new markets
The expansion of AI services has also pushed Meta to explore more new markets than it had previously.
The main one is smart glasses, which have been moderately successful, with seven million units sold last year. Meta is now trying to embed its AI model into the glasses, allowing users to interact with both the assistant and the world around them.
The AI model may also be embedded into other hardware in the near future, including a smart pendant.
It is also reportedly looking at prediction markets, similar to Polymarket and Kalshi, as a possible way to increase engagement across its platforms.