Anthropic, the company behind the Claude model, is exploring custom silicon. Reports indicate it has entered discussions with Samsung to develop a chip tailored for inference and training of its large language models (LLMs). The news – surfacing about a week after OpenAI announced its own custom chip partnership with Broadcom – marks another step in the race for hardware independence in the AI industry.

NVIDIA’s GPU dominance has been nearly absolute for years. But soaring costs, strained supply chains, and the need to optimize architectures for specific workloads are pushing major players to seek alternatives. Google with its TPUs, Amazon with Trainium, Microsoft with Maia: each is investing in internally designed silicon, betting that efficiency depends on hardware control. Anthropic, now backed by billions in investment, is no exception.

A custom chip would allow the company to calibrate every aspect of the architecture – core count, memory bandwidth, compute precision – to the needs of Claude models. This means lower inference latency, higher throughput, and potentially reduced energy consumption. But perhaps the most relevant aspect for on-premise deployments is control: a proprietary accelerator can be integrated into servers that reside physically in an organization’s data center, free from cloud provider licensing or dependencies. In scenarios where data sovereignty and regulatory compliance (such as GDPR) are paramount, relying on independent hardware becomes a competitive factor.

It remains unclear whether the collaboration with Samsung will target a chip for training, inference, or both. No details have emerged about the manufacturing node, architecture, or timelines. However, the choice of a partner like Samsung – which has proven experience in AI chip production and an advanced foundry – indicates a different path from OpenAI, which with Broadcom focuses on a specialized neural network design without directly entering manufacturing.

For enterprises evaluating local stacks, this fragmentation of the hardware market presents both opportunity and complexity. On one hand, chips optimized for specific models could make on-premise deployments economically viable that today require prohibitive investment. On the other, the proliferation of proprietary platforms threatens to complicate standardization and software portability. Those designing LLM infrastructure will need to navigate an increasingly diverse ecosystem where compatibility among frameworks, libraries, and accelerators is not a given.

AI-RADAR tracks these developments closely, aware that the AI game also hinges on the ability to choose the right hardware for each operational context.