Anthropic Launches Early Work on Self-Developed AI Chips: Negotiating 2nm Process with Samsung
According to The Information, Anthropic has initiated early work on self-developed AI chips and is in collaboration negotiations with Samsung Electronics, considering the use of Samsung's 2nm process and advanced packaging services to shorten the distance between processors and high-bandwidth memory to enhance transmission efficiency.
To advance the project, Anthropic recently recruited Clive Chan, an early member of OpenAI's self-developed chip team. Faced with high computing costs, leading large model manufacturers are accelerating hardware self-construction, and the project is still in the early planning stage.
Anthropic stated that Amazon's Trainium, Google's TPU, and NVIDIA's GPU remain the core of computing power; for Samsung, potential collaboration would be a key breakthrough in competing with TSMC in the 2nm foundry market.
Market mechanisms indicate that the demand for AI training and inference is driving capital from pure GPU procurement to vertically integrated self-developed chips, with event-driven collaborations accelerating funds towards advanced processes and packaging. Companies like Anthropic benefit from cost optimization and supply chain control, while foundries like Samsung face pressure to compete for large orders to increase market share.
Source: Public Information
ABAB AI Insight
Anthropic previously relied mainly on Amazon and Google cloud computing infrastructure and had rapidly expanded its training clusters through large-scale procurement of NVIDIA GPUs between 2024-2025. This self-development move continues the hardware vertical integration path initiated by competitors like OpenAI.
In terms of capital strategy, Anthropic is mobilizing engineering resources through talent poaching and negotiations with Samsung, motivated by the need to reduce long-term computing dependencies and marginal costs. Specific actions include Clive Chan's joining to accelerate chip planning, forming a capital and technology transfer from cloud service procurement to self-developed accelerators.
Similar to OpenAI's collaboration with Broadcom on the first inference chip and Meta's self-developed projects, Anthropic is in the early stages of expanding from a pure model manufacturer to a soft-hard integration, coinciding with Samsung's competition cycle to catch up with TSMC in 2nm technology.
Essentially, this represents a restructuring of the industry chain: the cost pressure of large model training is shifting pricing power from GPU suppliers to end manufacturers. Through advanced process collaborations, technological substitution and supply chain diversification are achieved, with the mechanism being talent mobility and process breakthroughs lowering entry barriers, accelerating the structural shift of AI hardware from oligopoly to multi-party competition.
ABAB News · Cognitive Law
The higher the computing costs, the more self-developed chips become a necessity for survival.
The speed of talent poaching determines the length of the hardware catch-up window.
The deeper the cloud dependency, the more vertical integration can reshape industry pricing power.