San Francisco startup Sapiom says it has materially lowered a client’s token-based spending by routing requests to the most cost-efficient models and running open-weight models on its own infrastructure.
Who, what, when
- Polsia, a San Francisco company that operates swarms of AI agents for other businesses, saw projected annual revenue jump this year from $100,000 to $10 million, and its token spending on Anthropic reportedly climbed to $1.2 million per month.
- Polsia’s founder hired Sapiom, which specializes in token-cost optimization, to run evaluations. According to Ilan Zerbib, Sapiom’s founder and CEO, those evaluations reduced Polsia’s token consumption roughly tenfold to about $100,000 per month.
- Zerbib said the pricing charged by frontier-model providers such as OpenAI and Anthropic is “unsustainable” for many startups, preventing them from deploying products even when demand exists.
Sapiom’s business model and products
- Sapiom is raising a $35 million Series A led by Dragonfly’s Haseeb Qureshi. Earlier this year the company raised $15 million in a seed round led by Accel.
- The company is launching a single-API model router that automatically sends requests to the most cost-effective model for the task, positioning Sapiom against routing providers such as OpenRouter.
- Unlike many routing firms, Sapiom runs open-weight models on its own server racks located in a San Jose data center. Most competitors don’t operate their own infrastructure and act more as intermediaries between providers.
- Sapiom says it serves a few thousand customers and does not charge a markup on models it provides; because it operates its own compute, it invoices customers directly for the compute used.
- Sapiom also helps customers connect autonomous agents to external systems and the economy, offering services such as payment processing and SMS.
Market context and why it matters
- The company’s fundraising comes as corporate AI budgets face closer scrutiny. Forrester predicted enterprises will postpone a quarter of planned AI spending to 2027. A KPMG survey of more than 2,100 senior executives in June found only 7% could point to established returns from AI initiatives.
- Uber’s chief operating officer said in May that AI costs were proving “harder to justify” than expected.
- Zerbib contends that in about 95% of cases it does not make sense to use an expensive frontier model. He also argues that lower costs can ultimately benefit frontier-model providers by enabling companies to build more agents, some of which will require the most powerful models.
Competition and risks
- Routing and cost-optimization are becoming crowded fields: Amazon and Microsoft bundle intelligent prompt routing into Bedrock and Azure, open-source routers are freely available, and OpenRouter reportedly routes 25 trillion tokens per week while offering Auto Router and Fusion, which combines cheaper models to approximate frontier performance.
- One market tracker counts about 80 active routing competitors.
Conclusion
Sapiom’s proposition is that reducing inference costs can extend startups’ runways and make agent-based automation economically viable. Its combination of a single-API router and on-premise-style infrastructure could be an advantage, but the company faces rapid commoditization of routing services and strong competition from cloud providers, open-source projects, and established intermediaries.
