OpenAI가 자체 AI 칩 로드맵(Stargate 및 광범위한 컴퓨팅 전략의 일부)을 위해 Samsung Foundry를 제2 공급업체로 선정하면서 단일 TSMC 의존성을 줄이고 있다.

Every OpenAI chip built so far has come out of a single factory: Taiwan Semiconductor Manufacturing Company's leading-edge fabs. That changed on Wednesday, when OpenAI Korea General Manager Harrison Kim confirmed at a Seoul press conference that Samsung Electronics is now a joint production and research partner for the next generation of custom AI chips OpenAI is developing. The statement makes Samsung Foundry the first viable second-source option for OpenAI's custom silicon program — one whose 2nm yield at its Taylor, Texas facility has reportedly improved to approximately 80%.

The announcement arrived the day after Samsung shipped the world's first HBM4E high-bandwidth memory samples to major global customers — a distinction that cements Samsung's position on both sides of the AI chip equation: it supplies the memory that goes into AI accelerators, and it is now building the logic chips those accelerators sit alongside.

The most important structural element in the announcement is not the bilateral relationship between OpenAI and Samsung. It is the three-way web that connects them through Broadcom — and a $200 billion memorandum of understanding that makes Samsung Foundry's entry into OpenAI's chip supply chain a contractually paved path rather than an aspirational goal. Broadcom is OpenAI's primary chip co-designer. The two companies jointly created Jalapeño, OpenAI's first custom AI accelerator, unveiled in June 2026. On July 25, 2026, Broadcom and Samsung signed a memorandum of understanding at the AI Summit in San Francisco, committing to more than $200 billion in collaboration across memory, foundry manufacturing, and advanced packaging through 2030. The foundry portion of that agreement specifically covers Samsung's 2-nanometer and below process technologies for Broadcom-designed chips.

The engineering consequence is significant: because Broadcom co-designs OpenAI's chips and now has a signed foundry agreement with Samsung for 2nm logic manufacturing, future OpenAI chips designed with Broadcom's assistance could be physically manufactured by Samsung under that existing contract — without requiring a separate, direct production agreement between OpenAI and Samsung Foundry. OpenAI's relationship with Samsung on logic silicon runs through Broadcom as an intermediary, and the contractual infrastructure for that path already exists. Young Hyun Jun, Vice Chairman and CEO of Samsung's Device Solutions Division, said at the time of the MOU signing: "By expanding our collaboration with Broadcom across these critical technologies, we look forward to delivering greater value to customers while advancing the AI infrastructure of the future."

Samsung Foundry's Texas facility has been in development for years and delayed repeatedly, but as of September 2026 it is closer to commercial relevance than at any prior point. The Taylor fab's 2nm production was confirmed for a 2027 start by Samsung Foundry Vice President Margaret Han at the company's SAFE Forum in May 2026. The process is Samsung's SF2P+ node — a Gate-All-Around transistor architecture that the company has been refining since its 3nm runs, and which now covers Broadcom-designed networking chips and Tesla's AI5 automotive processor as confirmed customers at the 2nm node.

Yield — the fraction of manufactured chips that meet specification — is the metric that separates a competitive foundry from an expensive experiment. When Samsung began producing Exynos chips on its first-generation SF2 process in late 2025, analysts estimated yield around 50%, below the 60% threshold generally considered commercially viable. As of early September 2026, yield at the Taylor fab has reportedly improved to approximately 80%, according to industry sources. That figure, if accurate, would place Samsung's 2nm yield inside competitive range of TSMC's mature process performance — though independent verification is not available.

The Taylor facility has now received confirmed customers on its 2nm process, with Tesla's AI5 chip entering production and Broadcom's networking silicon on the roadmap. Samsung has invested approximately $37 billion in the Taylor facility, partly supported by US CHIPS Act incentives. An OpenAI logic chip award would be the most prominent AI-inference anchor customer the fab has won — a win that would not just add revenue but likely accelerate the yield learning curve and attract additional fabless customers evaluating second-source options to TSMC.

TSMC currently controls approximately 95% of advanced AI chip foundry manufacturing. Every major AI lab that has announced custom silicon — Google's TPU, Meta's MTIA, Amazon's Trainium, Microsoft's Maia, and now OpenAI's Jalapeño — has done so on TSMC's process. A commercially viable Samsung Foundry 2nm option does not threaten TSMC's existing position in the near term, but it changes the negotiating structure for every future chip design cycle. Customers who can credibly threaten to move tape-outs to Samsung extract different pricing conversations than customers who cannot.

Samsung's announcement of being the first to ship HBM4E samples — which occurred on May 29, 2026, well before rivals SK Hynix and Micron — is not merely a memory speed headline. It is the other half of OpenAI's hardware strategy. High-bandwidth memory is the component inside every AI accelerator that prevents compute cores from sitting idle. During AI inference — the process of generating responses to user requests — a chip must load enormous weight matrices and key-value caches from memory on every pass through the transformer layers. Jalapeño, OpenAI's first custom accelerator, carries six stacks of HBM4 totaling 216 gibibytes at 15.4 terabytes per second of bandwidth. The chip was specifically designed so each compute core slice is paired with a local HBM slice, eliminating the cross-chip memory traffic that throttles general-purpose GPUs.

HBM4E is the generational step beyond what Jalapeño uses. Samsung's 12-layer HBM4E runs at a stable 14 gigabits per second per pin, scalable to 16 gigabits per second — more than 20% faster than HBM4's qualified 11.7 gigabits per second speed. Capacity rises to 48 gigabytes per stack, a more than 30% increase over HBM4's 36-gigabyte configuration. Energy efficiency improves 16% and thermal resistance improves more than 14% versus HBM4, according to Samsung's HBM4E specifications.

Any successor to Jalapeño would almost certainly use HBM4E rather than HBM4. Samsung is currently the only company shipping HBM4E samples. SK Hynix, which had originally guided HBM4E samples for the second half of 2026, was reportedly ahead of schedule as of May 2026 but had not publicly confirmed shipments as of the Samsung announcement. Micron's 2026 HBM4 capacity was already committed, with HBM4E still on its roadmap. In the market for next-generation AI memory, Samsung has a credible first-mover window — and a signed LOI to supply memory to OpenAI's Stargate infrastructure. Samsung's projected memory demand from Stargate alone: up to 900,000 DRAM wafers per month.

Kim's statement at Wednesday's press conference was deliberately limited. "One of the areas where we have made the most progress and gained the most recognition with Samsung Electronics is our joint production and research on the next-generation chips we are developing," he told reporters, declining to elaborate on technical specifications, process nodes, timelines, or which types of chips are involved. Samsung, for its part, declined to confirm or comment on customer-specific details — its standard response when asked about chip production relationships. Neither company disclosed whether Samsung Foundry has a signed production contract, what process node a potential OpenAI chip would use, or whether the chip in question is logic, memory, or a packaged combination of both.

The deliberate vagueness is itself informative. OpenAI has been careful throughout its custom chip program to limit the competitive intelligence available to rivals, and the pattern from the Jalapeño program — in which the company confirmed a direction long before it revealed architecture — suggests the Seoul disclosure follows the same playbook. What Wednesday confirmed is that the Samsung relationship now extends beyond memory supply to joint chip development. What it left unconfirmed is everything specific about what that chip will be.