OpenAI denies Samsung foundry deal for custom AI processors

OpenAI has not announced plans to manufacture custom chips through Samsung Electronics, the artificial intelligence developer said Sept. 11, 2026, walking back comments from its regional executive about foundry production.

The statement addressed remarks by Harrison Kim, head of OpenAI Korea, who said Sept. 9 in Seoul that both sides were conducting joint research and production on next-generation chips under development.

A spokesperson for OpenAI told technology outlet Tom's Hardware that the executive referred only to existing ties. Those include enterprise adoption of ChatGPT and a preliminary framework signed last year, the company said.

OpenAI and Samsung Electronics have not disclosed contract values or wafer allocations for proprietary processors. Neither company has filed documentation verifying an active commercial fabrication order for OpenAI-designed silicon.

The existing relationship stems from an Oct. 1, 2025 letter of intent signed at Samsung headquarters in Seoul. Four group affiliates, including Samsung Electronics and Samsung SDS, joined the agreement to support AI infrastructure.

That agreement projected memory demand for OpenAI data center initiatives could reach up to 900,000 dynamic random-access memory wafers per month. Samsung agreed to explore supplying high-performance memory to meet those targets.

The 2025 framework covered memory supply, cloud software and floating data center concepts developed by Samsung Heavy Industries. The document did not include commitments for contract chip fabrication or dedicated logic capacity.

OpenAI relies on Taiwan Semiconductor Manufacturing Co. to build its first-generation custom inference processor, called Jalapeño. The chip accelerates artificial intelligence inference — the process of running trained models to generate responses.

TSMC fabricates the processor using advanced nodes and integrates memory using chip-on-wafer-on-substrate packaging, known as CoWoS. The Taiwanese company manufactures virtually all commercial accelerators deployed by OpenAI and Nvidia.

Severe packaging constraints at TSMC have pushed tech firms to evaluate secondary fabricators. TSMC controls roughly 90% of global advanced packaging capacity for artificial intelligence chips, creating an industry-wide delivery bottleneck.

Hyperscale demand has pushed packaging lead times past five months for top-tier accelerators. Designers fear single-source dependence on TSMC could cap deployment speeds as data center projects expand globally.

Samsung has marketed a turnkey manufacturing service to capture artificial intelligence customers. The Suwon-based company can fabricate logic chips, produce high-bandwidth memory — known as HBM — and assemble both on advanced packaging substrates.

Contract chip makers that lack internal memory production must procure DRAM stacks from external vendors before packaging. Samsung provides all three stages within its Device Solutions division, promising reduced shipping friction and tighter integration.

Samsung Foundry continues to trail TSMC in leading-edge market share. Samsung accounted for roughly 11% of global foundry revenue in the second quarter, while TSMC held more than 60% of the contract manufacturing market.

Samsung is pitching its 2-nanometer gate-all-around process — a transistor architecture that wraps gates around channel sheets to curb leakage. Commercial mass production remains targeted for the second half of 2026.

Securing anchor foundry customers for the 2-nanometer node remains critical for Samsung. Low yield rates on early gate-all-around iterations have previously prompted major mobile and computing clients to keep orders at TSMC.

Memory competition in South Korea adds another layer of complexity. Rival chipmaker SK hynix holds more than 50% of the global market for fifth-generation high-bandwidth memory, supplying the bulk of Nvidia server modules.

Samsung has expanded capital spending to qualify its latest high-bandwidth memory with major artificial intelligence designers. The company aims to bundle its memory with packaging services even when customers choose alternative foundries.

OpenAI has sought diverse infrastructure partners for its global computing expansion. The company has engaged cloud providers, utility operators and governments to secure gigawatts of power and server capacity for future models.

The developer maintains enterprise relationships across South Korea alongside its hardware talks. Samsung Electronics deployed custom ChatGPT tools internally across engineering units earlier this year under the cooperation framework established in 2025.

The public filings do not confirm whether OpenAI has taped out test silicon or completed physical design files for Samsung fabs. Neither party has scheduled public signing ceremonies for semiconductor manufacturing.

TSMC is scheduled to deliver production silicon for OpenAI's first-generation inference chip in late 2026. Samsung Foundry plans to present updated customer yields for its 2-nanometer line at its annual technology forum in November.

The clarification leaves OpenAI dependent on TSMC for its near-term proprietary silicon requirements. Samsung continues discussions with cloud and software developers to fill leading-edge capacity at its Pyeongtaek fabrication complex.