The meeting was brief, and the agenda was described in the thinnest of terms: cooperation in artificial intelligence and semiconductors. But when Samsung Electronics Chairman Lee Jae-yong sat down with OpenAI Chief Executive Sam Altman at the company’s San Francisco headquarters this week, industry executives had little doubt about the real subject — high-bandwidth memory, and who will supply it.
Lee, the grandson of Samsung’s founder, has made the memory business the centerpiece of his push to reshape the conglomerate. Samsung dominates conventional DRAM and NAND, but in the segment that now matters most — HBM, the stacked memory that feeds AI accelerators — it has been chasing SK Hynix for two years and has yet to catch it.
The meeting with Altman is the most direct signal yet that Samsung intends to press its case in the market that counts. OpenAI operates some of the largest AI training clusters in the world and has signed compute agreements with several infrastructure providers. Every one of those clusters runs on HBM, and the memory supplier that wins the next generation of contracts will lock in years of revenue.
SK Hynix holds the pole position. The company supplies the bulk of the high-bandwidth memory in Nvidia’s accelerators, including the HBM4 that powers the Rubin generation, and last week deepened its position when Nvidia and SK Group signed letters of intent for a partnership valued at more than $500 billion, with memory supply at its core. Samsung watched that deal from the outside.
Samsung’s problem has been qualification, not capability. Its HBM3E stacks took longer than SK Hynix’s to pass Nvidia’s demanding validation process, and Samsung lost early supply windows as a result. The company has said it is investing heavily to correct the course, and executives have signaled that its HBM4 offering — built on a more advanced process and design — will close the gap.
The stakes for Samsung are enormous. AI memory has become the most profitable slice of the semiconductor industry, with prices climbing as demand outruns supply. Every point of market share in HBM translates into billions of dollars of revenue, and the memory upcycle is expected to run through at least next year as data center operators keep building.
OpenAI is an attractive customer for reasons beyond volume. The company’s compute needs are growing at a pace that alarms its own investors, and securing memory supply directly — or through its infrastructure partners — hedges against a market where allocation, not price, has become the constraint. A relationship with Samsung would also give OpenAI a second source beyond the SK Hynix supply chain that feeds most of its current clusters.
People familiar with the discussions said the two sides also touched on broader cooperation, including AI chips and potential joint projects in Samsung’s foundry business. Samsung is the only company in the world that designs and manufactures memory, logic and foundry services under one roof, a combination that gives it a negotiating position few suppliers can match.
The meeting fits a pattern of courtship between AI labs and memory makers. Altman has said repeatedly that compute — chips, memory, power — is the binding constraint on AI progress, and he has been unusually active in securing supply chains. OpenAI has talked with chip designers and manufacturers about dedicated AI hardware, and memory has become part of those conversations.
For Samsung, the timing is delicate. The company is also a competitor in AI hardware: it builds its own accelerators, though they are aimed at a different market segment. Analysts said that dual role is less of an obstacle in memory, where Samsung can supply OpenAI without competing with it directly.
The company’s leadership has been clear about the priority. Lee has reshuffled executives and increased investment in advanced memory, and Samsung’s share price has trailed SK Hynix’s as investors priced in the gap between the two. Closing it has become a matter of corporate identity as much as strategy.
SK Hynix is not standing still. Its long-term agreements with Nvidia, and the co-development of next-generation memory stacks, put it inside the design process for future accelerators. Samsung’s path to relevance runs through the customers Nvidia does not serve exclusively — the AI labs and cloud providers building their own compute.
Whether the Altman meeting produces a contract will become clear over the next several quarters, when HBM4 supply agreements for 2027 are finalized. Samsung has said its next-generation memory will be in production on schedule, and its customers have shown willingness to qualify a second source in a market this tight.
The two companies declined to comment on the substance of the talks. Executives who have dealt with both said the meeting itself is the story: Samsung’s chairman does not travel to San Francisco for small talk, and OpenAI does not host chip suppliers for courtesy calls. The AI memory race just gained another front.
