Anthropic Explores Custom Chips, Engages SK Hynix and Samsung
N.R. Finch
Anthropic is exploring a custom AI chip and has approached SK Hynix for memory supply and Samsung Electronics for foundry manufacturing. The custom-silicon race is spreading from hyperscale cloud giants to frontier model companies, opening a new front in the AI compute supply chain.
How did this come to light?
SK Group chairman Chey Tae-won disclosed the contact on stage alongside Anthropic CEO Dario Amodei at an AI event in San Francisco.
The reveal came during South Korean President Lee Jae-myung's Silicon Valley visit, giving the news a diplomatic backdrop.
The project itself remains at a very early stage — Anthropic has not yet defined the chip's function, compute level, or how it would integrate into server clusters.
What role would each Korean supplier play?
SK Hynix would most likely supply memory, not logic chips. It is a global leader in HBM — high-bandwidth memory designed specifically for AI accelerators — and server DRAM.
Samsung Electronics is seen as a potential foundry partner, offering advanced process nodes, advanced packaging, and memory products. Reports link the discussion to Samsung's 2 nm process, but no production agreement has been confirmed.
This means → Korean suppliers could cover multiple layers of the chip stack at once — Samsung handling fabrication and packaging, SK Hynix providing the memory alongside it.
Why would Anthropic design its own chip?
Anthropic is a frontier model developer that relies heavily on external compute: NVIDIA GPUs, Google TPUs — tensor processing units, Google's custom AI training chips — and Amazon's Trainium accelerators. Anthropic says these platforms remain central to its scaling plans.
A custom chip therefore would not replace existing platforms. It would more likely target specific bottlenecks — inference efficiency, data movement, interconnect performance, or memory utilization.
In plain terms = Anthropic is not building its own power plant. It wants to design one custom part for the exact spot where the system chokes.
What does this mean for the AI chip landscape?
Custom AI processors were previously the domain of hyperscalers — Google, Amazon, Microsoft — whose data-center scale justifies the high R&D cost and long design cycles.
Anthropic's move signals that independent frontier model companies are now engaging directly in chip design and locking in key components, seeking lower cost, more stable supply, and hardware better matched to their workloads.
This reflects a shift in the logic of the AI compute race: buying chips is no longer enough. Whoever shapes the chip's design gains an extra lever on cost and performance.
Content is for reference only, not financial advice.