Anthropic Builds In-House Silicon Team to Design Custom AI Chips

Anthropic has confirmed it is establishing an internal silicon team to design custom AI chips specifically for its Claude models. The move signals a strategic shift toward vertical integration as the company seeks to optimize the hardware layer for its proprietary AI architectures.

According to reports from Ars Technica and TechBuzz AI, the company is in discussions with Samsung to develop and manufacture these custom chips. The partnership is specifically exploring the use of Samsung's 2nm technology to produce hardware tailored for both the training and inference of Claude. This initiative aims to enhance performance and lower the operational costs associated with running large-scale frontier models.

The Push for Hardware Independence

The AI industry is currently defined by a heavy reliance on Nvidia's H100 and H200 GPUs, which has created significant financial burdens and supply chain bottlenecks for model developers. To mitigate this, Anthropic has already diversified its compute strategy. The company has utilized Amazon Trainium and secured a massive deal with Google Cloud for up to 1 million Tensor Processing Units (TPUs) in a buildout exceeding 1GW of power.

Thomas Kurian, CEO of Google Cloud, noted that Anthropic’s expanded use of TPUs reflects the "strong price-performance and efficiency" the company has experienced with that hardware over several years. However, moving from third-party accelerators to proprietary silicon represents a deeper level of control over the compute stack.

The Strategic Stakes

Controlling the silicon layer is increasingly viewed as an existential requirement for frontier AI labs. By designing its own hardware, Anthropic can optimize for the specific matrix multiplications used in its transformer architectures, effectively bypassing the "Nvidia tax" and reducing dependence on a single vendor. Furthermore, leveraging Samsung's foundry capabilities allows Anthropic to diversify its manufacturing pipeline away from the industry's heavy concentration at TSMC.

This transition mirrors a broader trend among AI leaders. Meta has developed its MTIA chips, Microsoft has introduced Maia, and OpenAI is reportedly partnering with Broadcom to achieve similar goals. The objective for all these players is to align the physical hardware precisely with the software requirements of their models.

The Infrastructure Race

As the AI arms race evolves, the competition is shifting from purely algorithmic improvements to a battle over physical infrastructure and hardware efficiency. The ability to iterate on chip design in tandem with model architecture could provide a significant competitive advantage in terms of latency and throughput.

While the partnership with Samsung marks a concrete step forward, the industry remains watchful for the actual deployment timelines of these chips. For now, Anthropic continues to balance its immediate compute needs through its existing partnerships with Google and Amazon while building the internal expertise necessary to sustain its own silicon ecosystem.