Meta develops Switchboard AI router to cut costs by routing simple tasks to cheaper models

Meta is reportedly building an internal AI model router called Switchboard, developed by its in-house incubator AAI Labs. The project aims to optimize AI costs by automatically routing simpler tasks to cheaper models, much like the external service OpenRouter. Switchboard handles billing, usage tracking, and request routing, allowing Meta to reduce the skyrocketing token costs that came from its earlier 'tokenmaxxing' culture—where employees were encouraged to maximize AI usage. The shift became urgent after Google reportedly limited Meta's access to its own AI models due to excessive consumption. AAI Labs has around 200 approved internal projects, with some eventually released publicly, though Switchboard's current focus is internal cost reduction.

This cost-cutting initiative comes as Meta simultaneously ramps up AI infrastructure spending, with planned capital expenditures reaching up to $145 billion in 2026—nearly double previous forecasts. Meta is accelerating its custom AI chip timeline and may sell excess data center capacity through a neocloud platform, following the model of SpaceX's Colossus data center deals with Anthropic and Google. Beyond internal routing, Meta is exploring new AI markets including smart glasses (seven million units sold last year), a smart pendant, and even prediction markets similar to Polymarket. The Switchboard project reflects the broader industry trend of balancing massive AI infrastructure investment with operational efficiency.