Anthropic is exploring the development of its own AI chip and is reportedly in talks with Samsung Electronics about it. The move is part of a broader trend in which AI companies are seeking greater control over their own hardware and want to become less dependent on Nvidia.
According to The Information, the discussions with Samsung are still in the early stages. Many technical decisions have not yet been made. For example, it remains unclear whether the chip will be intended exclusively for AI inference or whether it will also be suitable for training models. It has also not yet been determined how the processor will be deployed within the server architecture or what performance levels the chip will ultimately need to deliver.
Response to chip shortages
The fact that Anthropic is exploring the development of its own processor comes as no complete surprise. Reuters reported as early as April that the company was exploring this possibility in response to the ongoing shortage of AI chips. With its own accelerator, Anthropic can better tailor the hardware to its own models while simultaneously reducing its dependence on external suppliers.
The processor’s ultimate role remains unclear. Some AI processors, such as Nvidia’s Rubin GPU, are designed for both training AI models and inference. Others, however, focus on a much more specific part of the AI workload. For example, the startup Etched developed an accelerator intended exclusively for inference, while Nvidia introduced the LPU 30 earlier this year for a specific category of computations within inference workloads.
This development is part of a broader trend within the AI sector. Last week, OpenAI unveiled its first in-house inference processor, Jalapeño, which was developed in collaboration with Broadcom. Broadcom had previously assisted Google in developing its TPU processors. Should Anthropic indeed develop its own chip, a partnership with a specialized chip designer would also be a logical step.
Samsung as a manufacturing partner
The Information does not specify which manufacturing technology Anthropic intends to use. Samsung currently has several advanced manufacturing processes at its disposal, including a 4-nanometer process and the newer 2-nanometer SF2P process. The latter was specifically developed for high-performance data center chips and is set to enter production later this year.
SiliconANGLE adds that the SF2P process uses so-called gate-all-around transistors. In this design, the gate completely encloses the conduction channel, thereby limiting current leakage and improving energy efficiency. The connections between the transistors have also been further optimized to boost performance.
In addition, Samsung produces HBM (High Bandwidth Memory), the high-speed stacked memory used in many modern AI processors to keep large amounts of model data available with very high bandwidth. It is not yet known whether Anthropic will use this technology, but it would be a natural fit if the company is developing a powerful AI accelerator.
Having its own chip does not mean that Anthropic is parting ways with existing suppliers. In a response to TechCrunch, the company declined to confirm or deny reports of discussions with Samsung. However, Anthropic emphasized that chips from Amazon Web Services, Google, and Nvidia remain an important part of its infrastructure. A proprietary processor thus appears to be primarily intended as a complement to the existing hardware strategy.
This aligns with its close collaboration with AWS. In April, Anthropic announced that it had committed to more than $100 billion in AWS infrastructure over the next ten years.
Part of a broader data center strategy
The potential chip development fits within Anthropic’s expansion plans. Last year, the company announced a $50 billion initiative to build AI data centers in the United States, in partnership with infrastructure specialist Fluidstack.
The expectation is that a proprietary processor will ultimately be deployed within that infrastructure. This will give Anthropic greater control over performance, energy consumption, and the availability of computing power. With demand for AI computing growing faster than supply, more and more AI companies are investing not only in models but also in the silicon on which those models run.