In a significant move within the AI industry, Anthropic has quietly entered the race to develop its own custom AI chips. This strategic shift is driven by the need to optimize the performance and cost-efficiency of its AI models, particularly the widely used Claude.
The company has begun hiring engineers with experience in semiconductor design, signaling a serious commitment to building hardware tailored to its specific needs. This initiative is part of a broader trend among leading AI labs to move away from relying solely on third-party hardware providers like Nvidia.
Anthropic’s Co-Design Approach
Anthropic’s stated goal is to employ a co-design approach, where the chip and the AI model are developed in tandem. This method ensures that each component is optimized for the other, leading to faster and more efficient performance. Apple and Google have successfully used this approach with their M-series chips and TPUs, respectively.
The company emphasized that this move is not about abandoning existing hardware providers but about enhancing efficiency. Anthropic will continue to use a multi-chip approach across AWSGoogleNvidia and AMD. However, the development of custom chips will allow Anthropic to fine-tune its hardware for specific workloads, potentially reducing costs and improving performance.
Current and Future Chip Partnerships
Anthropic has not waited for its in-house team to start utilizing custom chips. In April 2026, the company expanded its partnership with Google and Broadcom to secure roughly 3.5 gigawatts of next-generation TPU capacity, set to come online in 2027. This follows a gigawatt of capacity arriving in 2026 under a previous agreement.
Broadcom plays a crucial role in this partnership, developing and supplying custom TPUs and committing to providing networking and other components for Google’s next-generation AI racks through 2031. This substantial investment underscores Anthropic’s rapid growth and the increasing demand for its AI services.
The Arithmetic Behind Custom AI Chips
The decision to develop custom chips is largely driven by the economics of serving billions of tokens daily. The cost of each token is significantly influenced by the hardware used, and optimizing this cost can lead to substantial savings over time. General-purpose accelerators are designed to handle a wide range of tasks, which means they include silicon area for functionalities that a specific AI model may never use.
A custom-designed chip can eliminate this overhead, leading to more efficient and cost-effective operations. However, developing advanced chip programs requires significant investment and time. Anthropic’s substantial revenue and stable model architecture make this endeavor feasible and potentially highly beneficial.
Despite the move towards custom chips, Anthropic’s chief financial officer, Krishna Rao noted that the company will continue to rely on existing hardware providers. The software ecosystem and interconnect standards, such as CUDA remain critical for integrating and managing large-scale AI systems.
By optimizing its hardware for specific workloads, the company aims to enhance the efficiency and cost-effectiveness of its AI models, ultimately benefiting its growing customer base.
