Anthropic is in talks with SK Hynix to secure memory chip supplies for custom AI semiconductors, a move that underscores the intensifying scramble for components needed to train and run large language models.
Why memory matters
High-bandwidth memory, or HBM, has become a bottleneck in AI development. The chips are stacked vertically to shuttle data quickly between memory and processors, a critical step for training models that swallow billions of parameters. SK Hynix, one of the few companies that produces HBM at scale, has seen demand surge as AI companies expand their computing infrastructure.
The talks suggest Anthropic is planning to build its own accelerators rather than rely entirely on off-the-shelf chips from established suppliers. That strategy would give the company more control over performance and power consumption, but it also means locking in supply lines for specialized memory.
Anthropic's custom silicon push
The AI startup, best known for its Claude chatbot, has been quietly developing custom semiconductors for years. Building its own hardware lets Anthropic tailor chips to the specific demands of its models, potentially squeezing out more efficiency than general-purpose AI processors offer. But the approach requires heavy upfront investment and a reliable supply chain for components like HBM.
Securing memory directly from SK Hynix would reduce Anthropic's exposure to shortages that have plagued the broader AI chip market. Other companies in the sector have faced long lead times and price volatility for HBM, which is also used in high-end data center GPUs.
What SK Hynix brings
SK Hynix has been ramping up capacity for its fifth-generation HBM, called HBM3E, as customers like Anthropic and other AI firms rush to lock in allocations. The South Korean chipmaker has already signed multiyear deals with several large tech companies, though it has not disclosed those clients by name.
For Anthropic, the partnership would mean priority access to advanced memory that is now a critical enabler of AI training clusters. The company's custom chip design, if successful, could eventually reduce its dependence on general-purpose AI processors from other vendors.
The talks are ongoing, and no agreement has been finalized. Both companies declined to comment on the negotiations.
