Sapiom Secures $35M to Slash AI Agent Spend, With Anthropic as Backer

A Startup Takes Aim at AI Agent Bills

San Francisco startup Sapiom has raised a $35 million Series A to tackle a rapidly emerging pain point for businesses: runaway spending on AI agents. The round, led by Dragonfly, comes just 11 months after the company launched and six months after a $15 million seed round from Accel, pushing total funding to $50 million.

Sapiom sits between an AI agent and the models it calls, acting as a smart router. Before a request is processed, it evaluates which model, tool or service can handle it at the lowest cost, and enforces a preset budget. The company says its platform has already processed more than 270 million transactions.

The most striking example involves Polsia, an AI startup that runs swarms of agents to help operate other businesses. Polsia’s revenue exploded from $100,000 to $10 million, but its token bill on Anthropic reached $1.2 million a month. After deploying Sapiom, that bill fell roughly tenfold to about $100,000.

In a strange twist, Anthropic is one of Sapiom’s backers, returning for the Series A alongside Okta Ventures, Menlo Ventures and Array Ventures. Founder Ilan Zerbib frames it as complementary: cheaper inference allows companies to deploy more agents, and some of that work will still demand frontier models. The funding lands as corporate AI budgets face their first hard audit. Gartner forecasts that over 40% of agentic AI projects will be cancelled by end-2027, with escalating costs a leading reason.

Behind Sapiom's $35M Bet on Cheaper AI Inference

The Anthropic Irony: Invested in Its Own Bill Reduction

Sapiom’s core value proposition is cutting the enterprise bill for frontier models—exactly the kind of bill Anthropic’s own customers face. Anthropic’s participation signals a strategic acknowledgement that not every agent call needs its most expensive model. Founder Ilan Zerbib argues that in 95% of cases, a cheaper model suffices, and that enabling more agents overall expands the total market for frontier models. This investment can be seen as a hedge: Anthropic gets a seat at the table in a layer of the AI stack that could otherwise bypass it entirely.

Routing Wars: OpenRouter, Hyperscalers, and 80 Competitors

Sapiom enters a crowded field. OpenRouter already moves about 25 trillion tokens a week. Amazon Bedrock and Microsoft Azure are bundling model routing directly into their platforms. One market tracker counts 80 active routing competitors. The core routing function is rapidly becoming a commodity—a feature rather than a standalone business. Sapiom differentiates by running open-weight models on its own hardware in a San Jose data centre, billing for compute directly rather than adding a margin on model access. Whether that infrastructure edge is defensible remains an open question, especially as cloud providers integrate routing deeper into their ecosystems.

Is Owning the Racks Enough?

The startup’s bet is that direct control over inference hardware delivers cheaper, more predictable performance than acting as a middleware layer. In the short term, this could attract customers who want to avoid per-token markups. But the landscape is evolving fast. If large cloud platforms can match that cost structure by running inference at scale, or if open-source routing alternatives mature, Sapiom’s hardware advantage could shrink. The $50 million total funding gives it a runway to prove that ownership of compute plus smart routing is a durable moat—not just a temporary solution to a spending spike.

What the Funding Round Means for AI-Intensive Businesses

• CFOs auditing AI spending can expect 5–10x cost reduction on non-frontier tasks. Polsia’s experience—slicing $1.2M monthly Anthropic bills to $100K—provides a real-world benchmark for similar agent-heavy operations.
• Competitive routing is becoming standard; plan for hybrid approaches. OpenRouter, Microsoft Azure and Amazon Bedrock are embedding routing, meaning the days of standalone routing tools as a premium service may be numbered. IT procurement should evaluate whether a dedicated tool or an integrated cloud feature better fits their stack.
• Sapiom’s owned data centre gives it a direct compute cost advantage—for now. Companies with latency or cost-sensitive workloads can use that as negotiating leverage, but should track how quickly hyperscalers can replicate the model.
• The Gartner prediction of 40%+ project cancellations by 2027 is a warning. Enterprises without a clear AI cost strategy risk shelving projects before they prove value. Start with a hard budget for agent inference, leveraging routers where frontier models aren’t essential.