Meta Reportedly Developing AI Router 'Switchboard' to Cut Inference Costs
0xBroomberg
Meta is incubating an internal AI routing tool called Switchboard that scores each request by difficulty and sends simple tasks to smaller, cheaper models — aiming to rein in inference costs approaching billions of dollars a year.
What problem is Switchboard trying to solve?
An internal Meta document states plainly: "We pay top-model prices for every coding request, including the easy ones."
Switchboard's logic is straightforward — score each incoming request for difficulty, route easy tasks to a small model, and reserve the top-tier model for hard problems only.
In plain terms = instead of sending every question to the most expensive specialist, Meta wants a triage desk — routine cases go to general practice, only the tough ones see the expert.
The project is run by AAI Labs, an internal employee-driven incubator. It remains early-stage; Meta declined to comment, and the product may never ship externally.
How big is Meta's AI bill?
An internal memo sent to roughly 6,000 employees revealed they consumed 73.7 trillion tokens in about 30 days. Meta even created a "Claudeonomics" leaderboard to track usage.
CTO Andrew Bosworth criticized the "token-maximizing" culture directly, writing: "Not all activity is progress. Token volume is not a measure of impact by any definition."
This means → the problem is not just technical waste — internal usage habits have spiraled, with employees defaulting to top-tier models regardless of task difficulty.
On the capex side, Meta has committed $125–145 billion in 2026 capital spending, mostly earmarked for AI infrastructure.
Where does the model-routing market stand today?
OpenAI's GPT-5 already has built-in routing, redirecting low-complexity requests to cheaper models.
Third-party results are concrete: Palantir's Evolve system cut inference costs by up to 97% in certain cases; McCarthy Building's token usage dropped 60% year-over-year; Cognition's routing system reduced costs by roughly 35%.
This reflects a shift — model routing has moved from experimental tooling to a must-have cost layer for enterprises.
How is capital pricing this market?
Routing-gateway startup OpenRouter recently closed a $113 million Series B led by Alphabet's CapitalG, at a valuation of roughly $1.3 billion — more than double its estimated valuation a year earlier.
The Information reports that OpenRouter is fielding acquisition interest in the multi-billion-dollar range.
This means → investors now treat "model routing" as a standalone infrastructure layer worth betting on independently, not just a feature bolted onto a frontier-model provider.
What does Meta building its own router mean for the landscape?
If Meta builds in-house rather than relying on OpenRouter or similar services, it can convert part of its massive AI capex into direct efficiency gains.
In plain terms = building your own triage desk not only saves money but puts Meta in direct competition with Databricks and Palantir on the routing layer.
Whether Switchboard evolves from an internal tool into an external product will be a key signal for Meta's AI monetization path.
Content is for reference only, not financial advice.