Anthropic Eyes Custom AI Chips, Asks SK Hynix for Memory Supplies

Anthropic has taken a concrete step toward building its own AI semiconductors, asking South Korean memory giant SK Hynix for component supplies — a move that signals the Claude maker is serious about reducing its dependence on Nvidia's GPUs.

The disclosure came from SK Group Chairman Chey Tae Won during an AI summit in San Francisco, where he appeared on stage alongside Anthropic CEO Dario Amodei. Chey described the request as "remarkable" for an AI developer to have its own chip ambitions. His conglomerate controls SK Hynix, the world's leading producer of high-bandwidth memory (HBM) — the specialized chips essential for training and running large AI models.

Anthropic has not confirmed a chip design, manufacturing partner, or production timeline. A company spokesperson said it was "always exploring options" while drawing on a range of chips for a diversified computing approach. But the supplier-side confirmation from Chey — combined with Amodei's separate announcement that Anthropic has signed supply deals with both Samsung Electronics and SK Hynix — suggests the company is methodically assembling the pieces of a custom silicon program.

Why Build Your Own Chips?

The economics are straightforward. Anthropic raised $65 billion at a $965 billion valuation in May, and it's pouring capital into computing infrastructure — including up to five gigawatts of Amazon capacity, five gigawatts of Google TPU capacity with Broadcom, and GPU clusters in SpaceX's Colossus data centers. At that scale, the margin between buying off-the-shelf GPUs and designing task-specific accelerators can run into billions of dollars annually.

Anthropic would join Google (TPUs), Amazon (Trainium, Inferentia), and reportedly Meta in the club of AI companies designing their own chips. The approach gives companies more control over performance, cost, and supply chain security at a time when Nvidia's GPU allocation decisions effectively determine who can train frontier models and when.

The HBM angle is particularly strategic. SK Hynix controls roughly 50% of global HBM production, and CEO Jensen Huang's Nvidia has already locked down a multi-year HBM4 partnership with the Korean supplier worth over $500 billion. Anthropic's approach to SK Hynix signals it wants its own seat at the memory supply table rather than competing for whatever Nvidia leaves on it.

The Broader Context

Chey's revelation came during a high-profile visit by South Korean President Lee Jae Myung to Silicon Valley, where $950 billion in AI infrastructure initiatives were announced. SK Group alone signed $750 billion in deals, including the Nvidia partnership. Samsung separately signed a memorandum with Broadcom covering up to $200 billion across memory, foundry, and advanced packaging.

For Anthropic, which is reportedly eyeing an IPO, vertical integration into hardware could reshape its cost structure and competitive positioning. The company has already weathered government scrutiny over its Mythos and Fable models, and its Opus 5 release last week emphasized cost efficiency — matching the same $5/$25 per-million-token pricing as its predecessor while delivering near-Fable performance.

Building custom chips is a multi-year, multi-billion-dollar bet. But for a company that just raised $65 billion and sees compute as its primary input cost, the math may already work.