Anthropic is building an in-house silicon team as Claude’s compute strategy deepens

Lead

Anthropic is extending the Claude race into the hardware stack. According to Ars Technica, the company has confirmed that it is hiring for a “custom silicon” team, including roles such as silicon engineer and technical program manager for silicon projects. That puts Anthropic on the same path as several other major AI labs: not just building models and products, but also trying to shape the infrastructure those models run on.

This does not mean Anthropic is about to abandon external chip suppliers. A company spokesperson said Anthropic will continue to pursue a “multi-chip approach,” using hardware from other companies alongside its own designs. In other words, the effort looks less like an immediate replacement for today’s GPU-heavy deployments and more like a long-term strategy to diversify and optimize compute.

Key points

• Anthropic has confirmed a custom silicon push: Hiring for semiconductor design and silicon program roles suggests the effort has moved from rumor to internal capability building.
• The plan is not chip self-sufficiency overnight: Anthropic says it still expects to use third-party hardware, with its own designs becoming one part of a broader compute portfolio.
• Hardware and models may be co-designed: The company wants teams to work on future models and hardware together, potentially improving efficiency, performance, or cost.
• The industry is moving in the same direction: OpenAI has announced a custom inference chip called Jalapeño with Broadcom, Google has long used its own AI hardware, Meta has deployed in-house chips, and Mistral is reportedly exploring similar options.

Why it matters

The immediate backdrop is the AI industry’s dependence on high-end accelerators, especially Nvidia GPUs. Demand for AI compute continues to exceed available capacity, and access to infrastructure has become a strategic constraint. For frontier model providers, relying too heavily on one dominant hardware ecosystem can create cost pressure, supply risk, and limited bargaining power.

Custom silicon offers another possible advantage: vertical integration. A general-purpose accelerator must serve many customers and workloads. A model provider, by contrast, knows its own architectures, inference patterns, memory needs, and deployment priorities. If hardware can be tuned around those requirements—and if models can be designed with the hardware in mind—the result may be better throughput, lower latency, or improved economics at data-center scale.

Still, Anthropic appears to be early in this journey. The company is hiring key people, and chip design requires long timelines, manufacturing partners, validation, and deployment work. Any practical benefit for Claude users is unlikely to appear immediately. The significance is more strategic: competition among AI labs is moving beyond model quality and product features into the deeper layers of data-center hardware, supply chains, and compute economics.

Source: Ars Technica AI