SK hynix and OpenAI hold direct HBM talks for custom silicon

OpenAI Chief Executive Sam Altman held high-level discussions with SK Group leadership in Seoul on Nov. 2, 2026, advancing direct procurement talks for high-bandwidth memory to support custom artificial intelligence hardware.

The talks centered on securing dedicated allocations of specialized DRAM for OpenAI's hyperscale infrastructure project, known as Stargate. OpenAI projects monthly requirements reaching 900,000 DRAM wafers across the project, more than double current worldwide specialized capacity.

SK hynix and Samsung Electronics together control roughly 83% of global high-bandwidth memory (HBM) revenue. Direct supply agreements allow frontier model developers to bypass traditional accelerator merchant vendors when sourcing critical silicon.

Altman met with SK Group Chairman Chey Tae-won alongside executive leadership from SK hynix and SK Telecom. OpenAI Vice President of Hardware Richard Ho accompanied Altman to evaluate technical integration parameters, industry officials said.

The encounter followed an initial non-binding letter of intent signed between the two companies on Oct. 1, 2025. That framework established SK hynix as a primary memory supplier for OpenAI's custom hardware buildout.

Under that 2025 framework, SK hynix outlined production capabilities targeting up to 900,000 DRAM wafer starts per month. Reaching that run rate would require dedicating the majority of South Korea's planned advanced packaging lines.

SK hynix has already committed the entirety of its HBM production capacity through 2026 under pre-allocated customer contracts. Securing early reservation queues is vital for OpenAI to protect its deployment schedules.

Hyperscale operators are shifting from merchant graphics processors to proprietary application-specific integrated circuits (ASICs). Custom accelerators require bespoke memory configurations, shifting procurement negotiations directly to silicon fabricators.

Standard merchant accelerators pair memory stacks using standardized physical interfaces. Custom ASICs allow developers to dictate stack height, thermal tolerances and memory bus widths directly to the fabrication plant.

The Seoul discussions examined the technical timeline for fifth-generation HBM3E memory and next-generation HBM4 modules. HBM4 introduces a fundamental architectural shift by transitioning the underlying base logic die from memory processes to advanced logic nodes.

SK hynix plans to use foundry partners to fabricate custom base dies for HBM4. This architecture binds memory manufacturing closely to customer-specific compute engines, making merchant off-the-shelf procurement unfeasible.

OpenAI requires high-density memory to eliminate memory-wall bottlenecks — the throughput delay when moving data between storage and computing cores. Large transformer models require terabytes of continuous bandwidth to sustain inference efficiency.

Altman also met with Samsung Electronics leadership during his Seoul visit, reviewing parallel memory collaboration agreements. Samsung signed its own letter of intent with OpenAI in October 2025 covering high-performance DRAM.

Engaging both South Korean manufacturers allows OpenAI to hedge supply risks across the only two vendors capable of mass wafer stacking. Micron Technology currently targets roughly 100,000 monthly HBM wafer starts, leaving Korea as the dominant volume source.

The procurement strategy mirrors earlier captive component sourcing models used in consumer electronics. By contracting directly with foundries, model builders prevent middle-tier hardware margins from inflating data center deployment costs.

SK Telecom participated in the sessions to align domestic infrastructure hosting plans with the hardware schedule. The telecommunications affiliate holds a separate memorandum of understanding with OpenAI to develop regional data center sites in South Korea.

Constructing regional compute clusters ensures domestic latency advantages while providing testing grounds for custom silicon. The domestic sites will deploy hardware configurations tested through the direct SK hynix memory supply loop.

The capital scale required for the Stargate project exceeds standard supplier credit limits. Industry estimates place total hardware spending across multiple build phases at $500 billion through 2029.

Committing memory volumes at this scale alters SK hynix's capital expenditure planning for its Cheongju and Yongin fab complexes. Advanced cleanroom space must be scheduled years before deposition and packaging tools arrive.

Through-silicon via (TSV) etching equipment and advanced thermal compression bonders represent the primary tooling constraints. Tool lead times currently exceed 12 months, forcing memory makers to demand multi-year customer commitments before breaking ground.

SK hynix filed no regulatory disclosure on the Korea Exchange regarding the Nov. 2 meeting, and the companies have not published binding volumes, unit pricing or delivery schedules. The discussions remain governed by earlier non-binding intents while bilateral engineering teams validate custom package specifications.

Technical teams from OpenAI and SK hynix are scheduled to complete package specification reviews in February 2027 ahead of initial test wafer fabrication.