Samsung moves up OpenAI's chip stack

OpenAI will design its next-generation AI accelerators with Samsung Electronics, The Register reported, citing remarks by OpenAI Korea general manager Harrison Kim at a press conference on Wednesday.

The relationship is not new. Samsung and OpenAI signed a letter of intent in October 2025, under which Samsung joined the Stargate programme as a strategic memory partner supplying high-performance, low-power memory for OpenAI's data centre build.

That release carried the number that has anchored every subsequent conversation: OpenAI's projected demand of as many as 900,000 DRAM wafer starts per month, to be met by Samsung and SK hynix scaling advanced-memory production.

Nine hundred thousand wafer starts a month is not an incremental order. It is a demand signal large enough to shape the capacity plans of the two firms that dominate high-bandwidth memory, and it commits capital years before the compute it supports is built.

The move from supplying memory to helping design the accelerator is a different kind of commitment. Memory is a component bought against a specification. A custom accelerator is a joint engineering programme with a multi-year cycle, and it ties the two firms together far more tightly than a supply contract does.

The strategic logic on OpenAI's side is the one every large buyer of AI silicon has reached. Custom parts reduce dependence on a single dominant GPU supplier and give the buyer control over performance per watt and cost per token, which are now the numbers that decide whether a model is economic to serve.

For Samsung the relationship runs in both directions: supplier of memory, prospective foundry and packaging partner, and a large enterprise customer of the systems it helps build. Those roles are usually held by different companies.

What is not yet public is the part that decides whether this matters: which process node, which packaging, and on what schedule. Until those are known, this is a direction rather than a product.