Anthropic is in early discussions with Samsung Electronics to manufacture a custom artificial intelligence chip, marking a potential first step by the Claude developer toward building its own silicon stack as compute demand becomes one of the defining cost and capacity constraints in the AI industry.
The talks, first reported by The Information and subsequently cited by TechCrunch and Korean media, remain preliminary. Anthropic has not finalized the chip’s design, target workload, performance profile or server architecture, according to the reports. The company also has not committed to using Samsung as a manufacturing partner. Anthropic told TechCrunch that a diversified hardware stack, including chips from Google, Amazon and Nvidia, will remain central to its compute strategy, while declining to add further comment on any potential Samsung deal.
A Move Toward Compute Control
The discussions underscore how leading AI model companies are moving beyond pure software and cloud contracts into deeper infrastructure planning. Training and serving frontier models require large clusters of GPUs, custom accelerators, high-bandwidth memory, networking equipment and long-term power commitments. For Anthropic, which competes directly with OpenAI, Google DeepMind and Meta, control over compute costs and availability is increasingly strategic.
Reuters reported in April that Anthropic was considering designing its own AI chips, though the initiative was then at an early stage and the company had not finalized a design or assembled a dedicated chip team. Since then, the economics have become more pressing. In April, Broadcom announced a long-term agreement with Google to develop future generations of custom AI chips through 2031, while also providing Anthropic access to roughly 3.5 gigawatts of AI computing capacity using Google’s AI processors beginning in 2027. Anthropic said at the time that Claude demand had accelerated sharply, with run-rate revenue surpassing $30 billion in 2026 from about $9 billion at the end of 2025.
A proprietary chip could eventually help Anthropic optimize inference costs, reduce dependence on scarce Nvidia GPUs and tailor hardware to Claude-specific workloads. But custom silicon is expensive, technically risky and slow to commercialize. Google’s TPUs took years to mature, and OpenAI, Meta and Amazon have all faced the challenge of matching Nvidia’s software ecosystem, developer support and rapid product cadence.
Samsung’s Foundry Opportunity
For Samsung, an Anthropic mandate would be strategically important. The South Korean group has been trying to close the gap with Taiwan Semiconductor Manufacturing Co., which dominates advanced-node foundry production and has captured the bulk of high-end AI accelerator demand. Counterpoint Research estimated that TSMC held 73% of the pure-play foundry market in the first quarter of 2026, illustrating the scale of the competitive gap Samsung is trying to narrow.
Samsung’s potential appeal lies in its advanced 2-nanometer process, its memory leadership and its push to offer more integrated AI chip manufacturing services. Reuters reported in June that Google was also in talks with Samsung to manufacture part of a next-generation AI processor, using Samsung’s 2nm technology for a component linked to memory connectivity. Samsung previously secured a $16.5 billion Tesla AI chip supply deal in 2025, a contract expected to support its delayed Taylor, Texas fab and strengthen its loss-making foundry business.
An Anthropic-Samsung agreement would not immediately change the AI chip market. Any chip would likely take years to design, tape out, test and scale. Still, the talks show that hyperscalers and model developers are increasingly treating custom silicon as a strategic hedge against GPU shortages, margin pressure and geopolitical concentration in semiconductor supply chains. For investors, the signal is clear: the AI race is expanding from models and cloud capacity into the foundry layer that will determine who can scale intelligence most efficiently.
