Meta's internal AI incubator develops OpenRouter as a competitor to reduce code development computing power costs

Meta's internal laboratory responsible for incubating AI products and tools is developing a scheduling service that competes with OpenRouter. This service can divert different AI tasks to more cost-effective large models, thereby reducing overall computing power expenses.

This incubator, named AAI Labs, is part of Meta's Applied AI Engineering team. Meta officially established this department in March of this year, allowing employees to propose various AI products and service projects for internal use, with mature projects expected to be formally released externally later. Internal documents reviewed by the media indicate that after project proposals are approved, the company will form small specialized teams to complete product development and assess whether to launch externally.

A July internal memorandum shows that AAI Labs currently has about 200 project initiatives covering three main directions: consumer products for end-users, developer tools, and internal computing infrastructure for enterprises.

The model scheduling system being developed, codenamed Switchboard, is one of these initiatives. This system will first determine the difficulty level of requests initiated by users or intelligent agents, and then automatically match the appropriate AI model to handle the tasks: simple requests will be assigned to smaller models with lower invocation costs, operating logic that is fundamentally consistent with OpenRouter's automatic routing product.

Insiders reveal that, like most of AAI Labs' incubated projects, Switchboard is currently still in the early research and development stage and may not ultimately be formally commercialized. However, the project proposal document states that Meta can first deploy the system internally to reduce computing power costs, and once the technology matures, it will open external services to enterprises deploying AI code intelligent agents on a large scale.

The various projects of AAI Labs reflect Meta's strategic thinking: relying on massive AI investments to explore new product forms, business operations, and revenue channels beyond its advertising business. Meta estimates that the company's total spending on AI computing infrastructure and supporting hardware facilities this year may reach $145 billion, more than double the spending in 2025; at the same time, the company has been restructuring its engineering architecture team to enhance its AI research and development capabilities.

This incubator rapidly iterates AI product prototypes based on a company-wide creative solicitation model, with the outcomes either optimizing Meta's internal operational efficiency or being refined into independent products for market launch.

Meanwhile, Meta has also been finding ways to control the massive expenses brought about by AI tools available to programmers and all employees. The media previously reported that Meta just promoted the widespread use of AI tools across the company in June of this year, and shortly thereafter implemented AI token invocation limits, while also building an internal accounting platform to uniformly monitor AI cost consumption and enforce token budget control across departments The Industry Value of AI Model Routing Services

OpenRouter has gained a large number of developer users due to its one-stop access to multiple large models and its ability to help developers control costs and improve efficiency. Last week, it was reported that OpenRouter is in talks for acquisition offers from leading tech companies, and this deal is expected to boost its valuation by several billion dollars; the company's latest valuation in April was $1.3 billion.

Model routing technology gained widespread industry attention last year, triggered by OpenAI's release of GPT-5, which has a built-in scheduling mechanism that automatically switches to lower-cost models when users input simple commands. Companies like Databricks and Palantir have also developed proprietary routing tools to manage computing costs and enhance overall operational efficiency.

Internal documents from Meta specify the core pain point that Switchboard aims to address: "Even for the simplest coding requirements, we are currently paying according to the pricing standards of top large models." The project proposal mentions that the vast majority of tasks for code intelligence agents can be handled by lightweight models, with only a small number of high-difficulty tasks requiring the use of expensive cutting-edge flagship large models. The document states: "Currently, all tasks are handled by a single model, leading to financial waste on simple tasks and insufficient computing performance on complex tasks."

The document also points out that high inference costs are a core obstacle preventing Meta from widely deploying AI intelligence agents across the company: "The cost ceiling limits the scalable deployment of AI intelligence agents."

Meta declined to comment on this interview.

AAI Labs also has another product in development: an AI in-car guide application compatible with Apple CarPlay and Android automotive connectivity systems. Another internal planning document describes this product as relying on AI to provide real-time explanations of landmarks along the route, allowing drivers to directly voice-query information about attractions. The product is positioned as an extension of Instagram's map features, with future plans to integrate location-based short videos (Reels) and travel recommendation services, and it may later connect with Meta's Ray-Ban smart glasses. The document states that this application will first undergo internal testing with employees, and will be publicly released after evaluation.

The establishment of AAI Labs aligns with Zuckerberg's top-level strategic vision: AI technology can enable small teams to complete product iterations more quickly. In an analyst call in April, Zuckerberg stated that AI intelligence agents could allow small teams to efficiently advance research and development progress, leading to a multitude of innovative outcomes; he also suggested that Meta hopes to create up to 50 new independent applications based on AI