Meta’s internal AI incubator, AAI Labs, is developing a model routing tool called “Switchboard” that reduces overall inference costs by assigning task difficulty scores and routing simple requests to cheaper, smaller models. The project is currently in its early stages and falls under AAI Labs, an internal incubator established by Meta in March this year. Internal documents reveal that Meta currently pays top-model prices for every programming request, including simple tasks that could be handled by smaller models—making inference costs the “primary barrier” to large-scale agent deployment. Meta’s AI infrastructure spending this year is projected to reach as high as $145 billion, and the company is exploring multiple ways to control AI expenses. Switchboard also targets the model routing market, which already includes competitors such as OpenRouter, and potential acquisitions could value the project in the billions of dollars.Author and source: Wall Street Journal
Under pressure from continuously rising spending on AI infrastructure, Meta is seeking cost-reduction strategies internally.
On July 21, according to tech media The Information, Meta’s internal AI incubator AAI Labs is developing a model routing tool called “Switchboard,” whose core logic closely resembles OpenRouter’s Auto Router product—routing simpler requests to more cost-effective small models based on task difficulty scores to reduce overall inference costs.
Switchboard is still in its early stages, and it remains uncertain whether it will ultimately be realized. However, internal documents obtained by The Information show that Meta’s team has clearly outlined two potential paths: one is to deploy it internally to reduce costs, and the other is to publicly release it to external organizations running AI programming agents at scale.
Analysis suggests that this tool is not only a cost-reduction initiative but may also represent Meta's attempt to create a new revenue stream in the AI tools market.
According to reports, the proposal directly addresses a current practical pain point for the company: paying for top-tier models for every programming request, including simple tasks that could easily be handled by smaller models.
Addressing the core issue: inference cost is the biggest barrier to scalable deployment.
The report states that the rationale for launching the Switchboard project was expressed quite plainly in internal documents: “We pay the top model price for every programming request, including simple ones.”
The document further notes that the majority of programming agent tasks can be fully handled by smaller models, with only a few tasks truly requiring the capabilities of state-of-the-art large models. However, the current situation is that "all requests are sent to the same model, leading to overspending on simple tasks or underperformance on complex ones."
The document explicitly identifies inference costs as the "primary barrier" to broader internal deployment of agents, stating bluntly: "Cost is the key factor limiting our ability to run agents at scale."
This statement corroborates Meta’s recent series of actions to control AI spending. According to prior reports from The Information, in June, Meta notified employees that, just weeks after encouraging broader company-wide adoption of AI tools, it would begin imposing limits on AI token usage and build an internal platform to track AI expenditures and enforce token budgets.
Notably, the Switchboard project is part of AAI Labs, an internal incubator under Meta’s Applied AI Engineering team, which was officially established in March this year. According to internal documents reviewed by The Information, this mechanism allows employees to submit proposals for AI products and services; once approved, small teams are assigned to build them and have the opportunity to publicly release them.
As of July this year, AAI Labs has approved approximately 200 projects across three key areas: consumer products, developer tools, and internal infrastructure. Switchboard is one of them.
This mechanism reflects Mark Zuckerberg’s broader strategic vision—using AI to enable small teams to rapidly build products. In April this year, Zuckerberg told analysts that AI agents mean “small teams can make very rapid progress,” and predicted the technology would drive “massive innovation.” He also stated that Meta could develop as many as 50 new applications.
Model Routing赛道: Beyond OpenRouter, major players are entering the arena
The model routing space targeted by Switchboard is attracting increasing attention.
OpenRouter has gained significant popularity among developers for helping them access a variety of AI models at lower costs. According to The Information last week, OpenRouter has engaged in discussions with a larger technology company regarding a potential acquisition, which could elevate its valuation by billions of dollars—the company was valued at $1.3 billion in April this year.
The concept of model routing has garnered broader attention following OpenAI's integration of a routing feature in GPT-5, which automatically switches to a more cost-effective model when user prompts are relatively simple. Since then, companies such as Databricks and Palantir have developed their own routing tools to manage costs and improve efficiency.
Meta's development of Switchboard is both a proactive response to its own cost pressures and a strategic move to build proprietary capabilities in this rapidly growing field.
Greater Ambition: AI Investments Seek Diversified Revenue Streams
Behind the Switchboard project is Meta’s broader ambition to transform its substantial AI investments into new tools, businesses, and revenue streams.
Previously, Meta estimated that its spending on AI infrastructure, along with other equipment and facilities, could reach as high as $145 billion this year, more than doubling the 2025 level. Meanwhile, Meta is reorganizing its engineering teams to strengthen its AI development capabilities.
The projects incubated by AAI Labs extend beyond Switchboard. According to another internal document obtained by The Information, AAI Labs is also developing an AI-powered driving guide app that runs on Apple CarPlay and Android Auto, using AI to narrate nearby landmarks and allow drivers to ask questions.
The document positions the product as an extension of Instagram’s map experience, with potential future integration of location-based Reels content, travel recommendations, and even Meta Ray-Ban smart glasses.
Reports indicate that these projects collectively outline Meta’s strategy: starting with employee ideas, rapidly prototyping AI products, and then selectively launching them to external markets—exploring incremental revenue opportunities beyond advertising while controlling costs.