The request, according to Business Today, went directly to two Korean semiconductor giants: Anthropic has submitted proposals to Samsung Electronics and SK Hynix to supply custom AI chips, known as ASICs, for its training and inference workloads. People familiar with the matter said the company has been evaluating custom-silicon options for more than a year, and the formal submissions mark the first time its interest has been made concrete.
Anthropic has run its models on NVIDIA GPUs since its founding, renting compute from cloud partners and buying capacity wherever it could find it. The dependence is common across the industry — NVIDIA controls the overwhelming share of the AI accelerator market — but it is also a strategic vulnerability: allocation, pricing and availability all sit in someone else’s hands. The requests to Samsung and SK Hynix are Anthropic’s formal entry into the custom-chip race that Google, Amazon, Microsoft and Meta have already joined.
Custom silicon is the industry’s quiet answer to NVIDIA’s dominance. Google’s TPUs power much of its AI business; Amazon builds Trainium and Inferentia chips for AWS; Meta has developed its own accelerators; Microsoft has invested in custom silicon through its partnerships. Each program represents the same logic: a company that spends billions on compute wants to control its supply, tune the hardware to its models, and escape the pricing power of a dominant supplier.
For Anthropic, the calculus is sharper than for most. The company has committed enormous sums to compute, its models are among the most expensive to train, and its margins depend on inference costs that are largely determined by hardware efficiency. A custom chip designed around Anthropic’s model architecture could cut serving costs substantially — or fail, leaving the company dependent on the supplier it was trying to escape. The requests are therefore both an engineering program and a hedge.
The choice of partners says something about the state of the market. Samsung brings foundry capacity, memory expertise and a history of ASIC manufacturing; SK Hynix brings the high-bandwidth memory that has become the binding constraint on AI chip performance and the most sought-after component in the industry. A custom chip needs both logic and memory, and the two Korean companies together control a critical share of both. For Anthropic, going to Korea rather than to the established Taiwanese foundry ecosystem is a supply-chain decision with political overtones: a way to diversify manufacturing outside the narrow geography where advanced chips are made today.
SK Hynix’s position is particularly interesting. The company’s HBM products have made it the indispensable supplier of the AI buildout, and it has been riding a demand wave that has pushed its profits to record levels. Taking on an ASIC project for Anthropic would diversify it beyond memory into logic — a bigger market, but one where it faces entrenched competition. The request puts SK Hynix in an unusual position: asked to build the chips that would compete with the NVIDIA platforms its own memory serves.
The timing reflects a broader shift in how AI labs think about silicon. For years, the frontier labs treated hardware as an externality — a constraint to be priced, not a variable to be engineered. The custom-chip wave represents the opposite view: that the models themselves should shape the hardware, and that the companies that control both will have an advantage over those that control only one.
There are also practical hurdles. Designing a competitive ASIC takes years, and the gap between a chip that works and a chip that beats NVIDIA at the system level is enormous. The custom programs at Google, Amazon and Meta have consumed billions of dollars over many years; none has yet displaced NVIDIA at the frontier. Anthropic’s requests are best understood as an insurance policy and a negotiating lever: a credible alternative supplier that can be invoked in conversations with NVIDIA, even if the chips never ship at scale.
The industry is watching for the follow-through. Samsung and SK Hynix have not publicly confirmed the requests, and negotiations of this kind often stall over volumes, pricing and the ownership of intellectual property. But the direction is clear. The arrangement in which AI labs simply rented capacity from whoever had it is giving way to a more deliberate one, in which the largest buyers commission their own silicon, choose their suppliers by strategic logic rather than convenience, and treat compute as a competitive weapon rather than a utility bill.
Anthropic’s move, if it closes, would put a third frontier lab firmly in the custom-chip camp — and put the two Korean giants at the center of a supply chain that was, until recently, a one-company market. It would also give Anthropic an answer to the question its investors ask most often: how a company spending billions on compute from a single supplier plans to control its own economics. The proposals are in Seoul. The answers will come back in silicon.
