Anthropic Evaluates a Proprietary AI Chip: Samsung Among Potential Manufacturing Partners

Anthropic would be accelerating plans to develop a proprietary chip dedicated to artificial intelligence. The company, according to sources, has initiated contact with Samsung Electronics for a possible collaboration regarding the production of the chip. According to rumors published by The Information, the project is still in its early stages and there are no definitive decisions on the chip's characteristics, its use, or even the continuation of the initiative.

The idea of a proprietary chip had already emerged in April, when Reuters reported that Anthropic was considering the possibility of designing a dedicated accelerator to address the growing scarcity of computational resources necessary for training and inferring AI models. At the time, the project appeared to be little more than an idea, lacking a dedicated team and without a defined technical direction. Today, however, the situation seems to have evolved. There was also speculation about the acquisition of the British startup Fractile to secure a very promising technology.

One of the most evident signs of Anthropic's intentions is the entry of Clive Chan into the company, a figure who previously contributed to the development of OpenAI's custom chip program. The hiring suggests a shift from a purely exploratory phase to a more concrete development activity, although there are no guarantees that the project will actually come to market.

When asked about the matter, Anthropic reiterated that its infrastructure strategy will continue to rely on a diversified hardware platform, which includes accelerators developed by Google, Amazon, and NVIDIA. The company did not provide further comments regarding the rumors about talks with Samsung.

The interest in custom chips reflects an established trend in the industry. Last week, OpenAI announced its first processor dedicated to inference, developed together with Broadcom, aiming to improve the performance-to-energy consumption ratio and reduce dependence on NVIDIA hardware. Google and Amazon have long been using proprietary accelerators, namely TPU and Trainium/Inferentia, within their cloud infrastructures.

For Anthropic, the issue is becoming increasingly significant in light of the rapid expansion of the business. According to available information, the annualized revenue growth rate has surpassed $30 billion during the year, more than tripling from the approximately $9 billion reported at the end of 2025. Such growth makes it economically more attractive to invest in the development of dedicated hardware capable of optimizing operational costs and energy consumption.

At the same time, the company has already signed a long-term agreement with Google and Broadcom to secure about 3.5 gigawatts of TPU-based computational capacity starting in 2027. Even if the proprietary chip project materializes, it is likely that Anthropic will continue to rely on a heterogeneous infrastructure where internally developed silicon would represent an additional option to increase control over the hardware intended for running the Claude models.

It remains to be seen whether Samsung will indeed be the chosen partner for production or if Anthropic will decide to explore other avenues, which would inevitably involve the congested TSMC or the outsider Intel.