Anthropic has confirmed plans to build an in-house silicon team, a move aimed at giving the Claude maker more control over the hardware that powers its models.
Ars Technica reports that Anthropic and OpenAI are racing to scale while reducing dependence on Nvidia. That does not mean Anthropic is about to replace GPUs across its infrastructure. Designing competitive AI silicon is expensive, slow, and difficult, especially when software ecosystems and manufacturing capacity matter as much as chip design.
The strategic reason is clear. Frontier AI labs spend enormous sums on compute, and access to chips can shape product timelines, model training, and margins. Even partial custom hardware work can help a lab understand bottlenecks, negotiate better, or optimize systems for its own workloads.
The plan also reflects how AI labs are becoming infrastructure companies. The biggest model developers now need expertise in data centers, networking, power, inference economics, and chips, not only neural-network research. Hardware control is becoming part of model strategy.