AI Pulse · Business Models & Winners

Sapiom just closed a $35 million Series A six months after launch, and the pitch wasn’t a vision deck. It was a spreadsheet. The company routes AI agent requests to the cheapest model that can handle the job, and it’s already processing over 270 million transactions with 100,000 agent runs a day. One customer reported cutting inference costs by 75% after switching, according to the funding announcement. Anthropic came back as a returning investor.

The infrastructure play here is simple. Most companies fire up the biggest, most expensive model for every agent task, even when a smaller one would work fine. Sapiom sits in the middle and routes each request to the right model based on cost and capability. The math adds up fast when you’re running agents at scale. That 75% cut isn’t theoretical, it’s a line item someone can show their CFO.

What matters for business owners is the pattern, not the product. Investors are writing checks for companies that show up with real usage numbers and customer outcomes, not promises. If you’re building or buying AI tools, the question shifts from “what can this do” to “what has this already done, and for how much less.” The companies that win funding, and customers, are the ones that can answer with a percentage and a reference.

This is the kind of thing we build into an AI command centre, where you can track what each model costs per task and route work accordingly. You don’t need to be processing 270 million transactions to care about the unit economics. You just need to know what you’re spending and whether a cheaper model gets you the same result. That’s the difference between running AI as an experiment and running it as infrastructure.