AI Pulse · Frontier Labs Watch

Anthropic just confirmed it’s building custom AI chips, targeting roughly 50% lower inference costs for Claude. The company is co-designing silicon alongside the model architecture itself, reportedly working with Samsung for manufacturing. Job postings for semiconductor engineers are live, offering $320,000 to $485,000, which tells you how serious they are about this.

This isn’t a moonshot. OpenAI shipped a Broadcom-built inference chip in June. Every major lab except xAI now has a custom silicon effort underway. The pattern is clear: the next competitive edge isn’t raw model capability, it’s how cheaply you can run inference at scale. If you can cut the cost of each API call in half, you can either pocket the margin or undercut competitors on price. Both matter when you’re serving millions of requests a day.

For anyone running a business that leans on AI, this matters because inference cost directly affects what you can afford to automate. Right now, a complex workflow that calls Claude ten times per task might be too expensive to run on every customer interaction. Cut that cost in half and suddenly it’s viable. The same logic applies to real-time analysis, customer support routing, document processing, anything that scales with volume. Lower costs mean more use cases pencil out.

This is the kind of shift we track inside the Omni Command Centre, where cost per task and model performance sit side by side so you can see when a pricing change makes a new workflow worth building.

The takeaway: inference cost is becoming a competitive lever. The labs know it. If you’re planning AI infrastructure for the next two years, assume costs will drop and plan your build-versus-buy decisions accordingly. According to the original report, Anthropic is betting heavily on this shift, and the salary bands suggest they expect results soon.