en.Wedoany.com Reported - On July 2, US artificial intelligence company Anthropic has initiated early work on self-developed AI chips and held discussions with South Korea's Samsung Electronics regarding potential manufacturing collaboration. The relevant processors are still in the planning stage, and Anthropic has not yet determined the chip's functional positioning, computing power level, or future deployment method in servers or server clusters.
The key point of this plan lies in its "early" stage. Anthropic has not completed chip design, nor is it immediately handing over mass production to Samsung; rather, it is evaluating whether a self-developed chip roadmap can support the long-term computing power demands of AI services like Claude. The computing pressure on large model companies primarily comes from three stages: training, fine-tuning, and inference. Among these, inference demand will continue to grow with user scale, enterprise client calls, and the expansion of multimodal applications. If the chip is designed for inference, it needs to prioritize energy efficiency, cost, batch concurrency, and memory access efficiency; if for training, it must address multi-card interconnection, matrix computing performance, high-bandwidth memory, cluster communication, and software stack adaptation.
Anthropic has already contacted multiple chip design companies but has not yet entered the detailed design phase. This means the chip specifications remain highly uncertain, including whether to adopt an ASIC architecture, whether to specifically serve Claude model inference, whether to also handle training tasks, whether to be placed in its own server clusters, or whether to be deployed through cloud partners. AI chips are not just about a single processor; they also require simultaneous determination of packaging form, HBM memory configuration, host interface, board design, server cooling, power supply, cluster networking, and compiler toolchain. South Korea's Samsung Electronics' potential role in such discussions may include wafer foundry, advanced packaging, memory collaboration, and chip manufacturing services. AI processors increasingly rely on high-bandwidth memory and advanced packaging, and the computing chip itself, HBM, interconnect structure, and cooling design must be considered as a whole. For Samsung, participating in Anthropic's self-developed chip project would not only secure an AI client but also enter a critical link in the self-built computing supply chain of US large model companies.
US large model companies are attempting to reduce their complete reliance on general-purpose GPU supply. Nvidia GPUs remain the mainstream choice for current AI training and inference, but their price, supply cycle, energy consumption, and cluster deployment costs put pressure on AI companies. Self-developed chips may not fully replace GPUs; they are more likely to be used first for specific models, specific inference workloads, or internal cloud clusters, reducing per-inference costs in controlled scenarios. If Anthropic continues to advance this project, it must bind model structure, operator characteristics, inference latency, throughput requirements, and hardware design together; otherwise, even if the chip is manufactured, it may struggle to achieve sufficient efficiency advantages in actual services. This news also places Samsung back in the competition for AI chip foundry. TSMC has long dominated the high-end AI chip manufacturing position, while Samsung hopes to attract more AI clients through advanced processes, HBM, and packaging capabilities. Previously, other US tech companies have explored AI chip-related manufacturing collaborations with Samsung, and Anthropic's addition to the discussion list indicates that large model companies and cloud service providers are seeking more diversified chip supply paths. For the entire AI infrastructure chain, the computing power competition is extending from "buying GPUs" to an integrated competition involving chip design, foundry, packaging, memory, servers, and data center deployment.
Currently, Anthropic's self-developed AI chip has not yet entered a clear product stage. Known information only indicates that the company has initiated early work, contacted chip design companies, and discussed potential manufacturing collaboration with Samsung. Whether the chip will be formally approved, how specifications will be determined, who will participate in the design, which generation of process technology will be used, and whether it will enter server cluster deployment all remain without final results. The value of such early-stage chip projects lies in revealing that large model companies are reassessing long-term computing costs and supply security.
