Sapiom raises $35M Series A to route AI agents to cheaper models

The interesting number in Semafor's scoop on Sapiom is not the $35 million Series A led by Dragonfly's Haseeb Qureshi, or the $15 million Accel-led seed that came earlier this year. It is $1.2 million a month. That, according to founder and CEO Ilan Zerbib, is what one customer, a San Francisco AI agent startup called Polsia, was spending on Anthropic tokens as its projected annual revenue lurched from $100,000 to $10 million. Sapiom claims it got that bill down to roughly $100,000 a month.

The pitch is straightforward. Sapiom runs open-weight models on its own server racks in a San Jose data center and offers a single API key that routes each request to whichever model it thinks can handle the job most cheaply. Zerbib's line, quoted by Semafor, is that "in 95% of cases, it doesn't make sense to go to a very expensive frontier model." Unlike most of the roughly 80 routers now in this market, Sapiom is not a middleman on top of other providers' APIs; it owns the compute and charges customers for it directly rather than stacking a premium on top.

Why this matters if you are not a router shopper: it is the clearest recent signal that the economics of running real AI agents at scale, not demos, are forcing a second layer of infrastructure to exist between the app and the frontier lab. Polsia is a useful poster child precisely because it has no employees and its cost line is almost entirely tokens. When a company like that says its bill is "unsustainable," the ceiling on the current pricing model is visible.

The honest caveat is that this is a single customer anecdote sourced through the vendor, and Semafor's piece does not tell you which open-weight models Sapiom is routing to, how quality held up after the tenfold cut, or who else joined the round beyond Dragonfly. It also does not mention a board seat. Take the Polsia numbers as reported by the CEO who saved them, not as an audited benchmark. And the competitive picture is crowded: OpenRouter alone is moving 25 trillion tokens a week, before you count Amazon Bedrock or Microsoft Azure.

The part worth watching is whether owning the racks turns out to be a durable moat or a temporary arbitrage. If Zerbib is right that trillions of agents are coming, the routers who own inference capacity, rather than resell it, get to keep the margin the frontier labs are currently collecting.