Anthropic in Talks with Samsung for Custom AI Chip Manufacturing
Taylor Wilson
Anthropic has begun early work on a custom AI chip and is negotiating with Samsung Electronics to use its 2nm process. Following OpenAI's playbook, yet another top AI lab is moving to control compute costs at the silicon level.
Why does Anthropic want its own chip?
Training and running large models depends heavily on Nvidia GPUs. Compute is the single biggest cost for AI companies.
This means → whoever can cut per-inference costs with custom silicon gains a durable edge in the commercial race.
OpenAI already partnered with Broadcom and last month shipped its inference chip, Jalapeño. Anthropic is tracing a nearly identical path.
How far along is the project?
Three people familiar with the matter say Anthropic is still in the early stages, working out the chip's function, specs, and deployment architecture.
Detailed design, testing, and manufacturing have not yet begun. The company is also talking to multiple chip-design firms.
In plain terms = this is the "blueprint" phase — tape-out and volume production are a long way off.
Why Samsung instead of TSMC?
Anthropic is considering Samsung's 2nm manufacturing process and advanced packaging — a technique that tightly integrates the main processor with high-speed memory to speed up data transfer.
TSMC's leading-edge capacity is chronically overbooked; Samsung has seized the opening to pitch its 2nm tech to more AI clients.
This reflects a broader shift: TSMC's capacity bottleneck is pushing orders toward Samsung. Google is also weighing moving some future TPU — tensor processing unit, Google's custom AI chip — production to Samsung.
What else ties Samsung to Anthropic?
In May, Samsung invested in Anthropic's funding round at a $65 billion valuation.
Two other major memory-chip makers — SK Hynix and Micron — joined the same round.
This means → the capital relationship came first; a foundry partnership is the natural next step.
What do the talent moves signal?
This month Anthropic hired Clive Chan, an early member of OpenAI's custom-chip team.
OpenAI began designing chips with Broadcom in 2024 and last month unveiled Jalapeño, an inference chip built to run large language models efficiently.
In plain terms = poaching from a rival's chip team signals Anthropic is not just exploring — it is actively building a roster.
What does Anthropic say officially?
Anthropic said Amazon Web Services' Trainium chips, Google TPUs, and Nvidia GPUs will remain core to its compute-scaling strategy.
The company declined to comment further on its chip roadmap. Samsung also declined to comment.
This means → the official line is still "multi-source," and custom silicon will not replace existing suppliers near-term — but the long-term direction is now visible.
Content is for reference only, not financial advice.