Key facts
- Anthropic is forming an in-house team to design custom silicon for its AI models.
- The company confirmed these plans after job listings for semiconductor engineers appeared.
- Anthropic will maintain a multi-chip strategy, using both its own designs and third-party hardware.
- This move mirrors efforts by competitors like OpenAI, Google, and Meta to develop custom AI chips.
- The goal is to reduce dependence on Nvidia and enhance model performance through co-design.
Anthropic has revealed its intention to establish an in-house silicon team dedicated to designing custom chips for its artificial intelligence models. This strategic move, confirmed by a company spokesperson, aims to reduce the company's reliance on hardware providers like Nvidia and enhance the performance of its AI models through integrated hardware and software development.
The decision follows a trend within the AI industry, where companies are increasingly pursuing vertical integration in their hardware strategies. OpenAI recently announced its custom chip, Jalapeño, developed in partnership with Broadcom. Google has long utilized its own Tensor Processing Units (TPUs) for its AI workloads, and Meta has also designed and deployed proprietary chips. Mistral AI is also reportedly considering similar custom silicon initiatives.
Anthropic's spokesperson clarified that the company will adopt a "multi-chip approach," continuing to utilize hardware from external vendors alongside its internally designed chips as it scales its operations. The company has previously co-designed hardware with partners but is now bringing more silicon expertise in-house to co-design new hardware and models side-by-side.
This pursuit of custom silicon is driven by several factors. The heavy dependence on Nvidia creates a potential strategic vulnerability, especially in a competitive market where AI compute infrastructure demand often outstrips supply. Furthermore, designing chips tailored to specific AI models can lead to significant performance improvements. Anthropic believes this co-design approach will provide a competitive edge, potentially enabling more efficient deployment of smaller or open-weight models on edge devices.