Anthropic in Early Talks With Samsung to Manufacture Custom AI Chip on 2nm Process
- Anthropic is in early-stage talks with Samsung Electronics to manufacture a custom AI chip using Samsung's 2nm process and advanced packaging facilities [1]
- Anthropic recently hired Clive Chan, an early member of OpenAI's custom chip team, signaling serious intent to build in-house silicon capabilities [2]
- The project is preliminary — Anthropic has not finalized what the chip will do, how powerful it will be, or how it fits into a server [1]
- The move follows OpenAI's June 24 unveiling of Jalapeño, a custom inference chip built with Broadcom in a nine-month development cycle [3]
- Anthropic says AWS Trainium, Google TPUs, and Nvidia GPUs remain central to its compute strategy, and is also in discussions with Microsoft and UK startup Fractile [2]
Anthropic has begun early-stage work on a custom AI chip and is in discussions with Samsung Electronics to serve as the manufacturing partner, The Information reported on Wednesday [1]. The talks specifically involve Samsung's 2-nanometer fabrication process and advanced packaging capabilities, though no formal agreement has been reached and no detailed design work has commenced [2].
The discussions mark a significant step for the $965 billion AI company, which recently hired Clive Chan — an early member of OpenAI's custom chip team — as part of a deliberate engineering buildout [2]. Samsung, along with SK Hynix and Micron, participated in Anthropic's $65 billion Series H fundraising round in May, giving the chipmaker an existing relationship with the AI lab [4].
The talks come just eight days after OpenAI unveiled Jalapeño, a custom inference processor co-developed with Broadcom in what the companies described as a nine-month development cycle — believed to be the fastest ASIC development in high-performance semiconductors [3]. The rapid proliferation of custom silicon projects among frontier AI labs reflects growing pressure to reduce dependence on Nvidia, which holds roughly 74% of the AI chip market [2].
What We Know
The project remains at an early stage. Anthropic has not determined what the processor should do, how powerful it should be, or how it would fit into a server, according to The Information's report [1]. The company may ultimately decide not to proceed with the partnership.
An Anthropic spokesperson declined to elaborate on the chip roadmap but confirmed the company's existing compute relationships remain intact. "AWS Trainium, Google TPUs and Nvidia GPUs will remain central to our compute strategy," the company said [2].
Beyond Samsung, Anthropic is also in discussions with Microsoft regarding its Maia AI chips and with UK-based startup Fractile about inference solutions — indicating a deliberate multi-vendor diversification strategy rather than a single-partner bet [2].
Why Samsung
Samsung occupies a unique position in the semiconductor supply chain as both a memory giant and a contract chip manufacturer. Its foundry division competes directly with TSMC, the world's dominant chipmaker, for leading-edge manufacturing contracts.
However, Samsung's foundry business has struggled with yields at advanced process nodes compared to TSMC, raising questions among analysts about whether it can close the manufacturing gap [2]. Google has separately considered Samsung for manufacturing future tensor processing units, which would give Samsung additional high-profile AI chip clients [2].
Samsung and SK Group announced a combined $518 billion decade-long investment plan for four South Korean memory-chip plants, underscoring the scale of capital flowing into semiconductor manufacturing [4]. Samsung's participation in Anthropic's fundraising round provided a financial relationship that likely facilitated the chip manufacturing discussions.
The Custom Silicon Race
Anthropic's exploration follows a pattern established by the largest technology companies. Google developed its Tensor Processing Units for internal AI workloads, Amazon built Trainium and Inferentia chips for AWS customers, and Microsoft developed Maia for Azure AI infrastructure.
OpenAI's Jalapeño, unveiled June 24, represents the first time a pure-play AI lab has brought a custom chip to production [3]. The inference-focused ASIC was co-designed with Broadcom and is targeted for initial deployment by the end of 2026. OpenAI said Jalapeño delivers substantially better performance per watt than current alternatives [3].
The custom chip push reflects the economics of running frontier AI models at scale. Training and inference costs remain a dominant expense for AI labs, and purpose-built silicon can deliver meaningful efficiency gains over general-purpose GPUs for specific workloads. Nvidia shares closed at $194.83 on Wednesday, down 1.4%, while Broadcom fell 2.4% to $360.45 [5][6].
What's Next
For Anthropic, the chip discussions represent optionality rather than an imminent product. Custom chip development typically takes years from initial design to production deployment — OpenAI's nine-month timeline with Broadcom was considered exceptionally fast and relied on Broadcom's deep ASIC design expertise [3].
Samsung's 2nm process is still ramping, adding another variable to the timeline. If Anthropic proceeds, the chip would join an increasingly crowded landscape of custom AI accelerators competing for workloads currently dominated by Nvidia's data center GPUs.
The broader signal is clear: frontier AI labs with sufficient capital are moving to vertically integrate their compute stacks, treating chip design as a strategic capability rather than outsourcing it entirely to Nvidia. Anthropic, valued at $965 billion after its May fundraise, has the resources to pursue this path [4].
Further sources
The stories that matter, in one email. Free — unsubscribe anytime.