OpenAI floated Samsung as a partner for its next chips, then walked the claim back within days. At a September 9 press conference in Seoul, OpenAI Korea chief Harrison Kim called joint work on next-generation chips a bright spot with Samsung Electronics. Reuters and the Seoul Economic Daily covered the event. Soon after, OpenAI Korea said there was “nothing new to announce,” as the Korea Times reported.
That soft denial does not erase the pressure behind the remark. Custom AI chips now need more factory capacity than one supplier can easily spare. OpenAI already builds its first custom inference chip, Jalapeño, with Broadcom, and TSMC makes it. Meanwhile, memory deals with Samsung and SK Hynix sit beside that factory story. The real signal is fear of running out of wafers.
In short, the Samsung Foundry chatter matters even after the walk-back. It shows how hard it is to scale private AI silicon when one foundry still owns the leading edge.
Chip Snapshot
Kim line: Joint production and research on next chips with Samsung
Walk-back: OpenAI: nothing new to announce; remarks meant existing ties
Jalapeño: Inference chip co-designed with Broadcom; made at TSMC
Memory path: Samsung and SK Hynix letters of intent for Stargate
What Kim Said About Samsung, Then What OpenAI Took Back
Kim’s quote was sharp enough to move headlines. “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 said, per Reuters and Tom’s Hardware. However, he named no process node and no product.
OpenAI Korea soon narrowed the claim. The company said Kim spoke about the existing relationship, which includes Samsung’s use of ChatGPT and a letter of intent from last October. “We have nothing new to announce on this front,” the statement said, per Reuters.
The Register and the Korea Times both carried a sharper line from OpenAI. According to both, it said it had not announced plans to make chips through Samsung. That is a hard brake on the factory reading of Kim’s words.
Still, the sequence teaches something useful. A senior local executive chose chip language in public. Corporate communications then pulled it back. Markets heard the first line, while careful readers must live with the second.
In hindsight, silence would have been safer for OpenAI. The fact that the remark landed at all shows how close chip supply sits to the firm’s growth story. When leaders talk about next chips in Seoul, listeners hear factories, not only memory.
Jalapeño Ties OpenAI to Broadcom and TSMC, Not Samsung Foundry
For context, OpenAI’s known custom chip path is already clear. On June 24, the company unveiled Jalapeño, its first custom AI chip for inference, as TechCrunch and CNBC reported. OpenAI and Broadcom designed it together, and TSMC manufactures it. That stack is the baseline for any Samsung talk.
Reuters and Tom’s Hardware both describe Jalapeño as a TSMC-made part. Kim, by contrast, spoke about “next-generation chips,” not the chip already in hand. Tom’s Hardware and eeNews Europe both say the second generation is well along in development. That timing makes a second factory more plausible for a later chip than for the one already locked to TSMC.
eeNews Europe draws the same line. OpenAI has called Jalapeño “the beginning of a multigenerational platform.” A later accelerator could, in theory, land at Samsung Foundry. Even so, neither firm has confirmed that contract.
Design speed also matters. OpenAI says Jalapeño went from initial design to manufacturing tape-out in just nine months, a pace VentureBeat also reported. That pace raises volume risk, because fast design cycles still hit a wall if the only top factory is fully booked.
We have nothing new to announce on this front.
OpenAI Korea statement, as reported by Reuters
Why a Samsung Foundry Second Source Is the Real Story
Even after the walk-back, the capacity math still bites. TSMC chairman C.C. Wei said in June that “customer demand has outstripped what we can supply,” per the Taipei Times, and Tom’s Hardware describes TSMC as fully booked. On top of that, OpenAI and Broadcom have planned 10 GW of custom AI accelerators across several years, according to Tom’s Hardware and eeNews Europe. Volume that large can outrun a single factory allotment.
Tom’s Hardware’s dual-source reading is frank. If OpenAI needs more wafers than TSMC can spare, paying to move a design onto Samsung Foundry buys both volume and backup. That is analysis, not a signed award. Still, it is the frame that makes Kim’s remark land.
Meanwhile, Samsung already sits in OpenAI’s memory plan. Last October, Samsung and SK Hynix signed letters of intent to supply memory for the Stargate data center project, as Reuters and eeNews Europe report. Memory deals and foundry dreams are different businesses. Both answer scarcity, and mixing them up is how hype spreads.
Other AI buyers feel the same squeeze. When every large model shop wants leading-edge dies, the queue at TSMC becomes a strategy board. As a result, second sources stop looking optional and start looking like insurance.
Broadcom Pact Makes a Samsung Foundry Path Thinkable
On July 25, Samsung Electronics and Broadcom announced a deal worth more than $200 billion through 2030, according to Samsung’s newsroom and Fortune. The pact spans memory and foundry work. Samsung says it covers its 2 nm and smaller process nodes, along with advanced packaging that stitches dies together.
OpenAI already designs with Broadcom, and Broadcom now has a deeper Samsung manufacturing lane. That link does not prove OpenAI will build at Samsung. It does show why analysts keep drawing the line. In principle, a Broadcom-shaped OpenAI chip could move onto Samsung silicon if the customer pays for the port.
eeNews Europe stresses the limit: “OpenAI has not said that Samsung will manufacture Jalapeño Gen 2.” Samsung has not named OpenAI as a foundry customer either. For now, the story matters more as proof that OpenAI is widening its supplier base than as proof of a fab award.
For Samsung, a public link to OpenAI would be a prize. Samsung Foundry has chased leading-edge share for years, and a famous AI customer would help fill advanced lines. For OpenAI, even the rumor of a second address on the wafer map eases investor fear about a single point of failure.
Memory Versus Samsung Foundry: The Honest Counter
The strongest counter is the walk-back itself. OpenAI now says there is no new manufacturing plan. Until a purchase order appears, Kim’s line reads best as color on an old relationship. Hype that ignores the clarification is bad reporting.
A second counter is category mix-up. Samsung’s clearest OpenAI links today are memory and enterprise ChatGPT use. High-bandwidth memory is tight, and Samsung is a main volume player alongside SK Hynix and Micron. The Register argues that HBM supply, or even custom HBM base dies, may be the boring and correct reading of closer Samsung ties. That path can strengthen next chips without moving logic wafers off TSMC.
Process risk also cuts against a fast Samsung Foundry leap. The Register and DigiTimes both point to yield trouble on Samsung’s 3 nm process. Still, Samsung does make chips for AI startups such as Rebellions and Groq, according to The Register. EE Times Asia and KED Global reported those foundry deals earlier. Capability is not the same as an OpenAI commitment, though, and yield gaps still shape buyer caution.
Admittedly, those counters are real. They do not cancel the scarcity thesis. They only force precision about what is proven.
What the Samsung Foundry Episode Still Reveals
OpenAI is building a private silicon stack while still living inside a foundry market that one company dominates. Jalapeño proves the design muscle and TSMC’s grip on manufacturing. Samsung Foundry talk, even when walked back, exposes the anxiety about second sources.
Investors should separate what is documented from what is not. OpenAI and Samsung already share memory and software ties. By contrast, Samsung Foundry winning OpenAI logic production is not confirmed, and OpenAI denies announcing it.
Similarly, chip buyers outside OpenAI should watch the same pattern. When a model shop designs its own accelerators, it inherits the factory fight that chip firms already know. Supply chain talk becomes product talk, and a single remark in Seoul can drive a full news cycle.
For South Korea, the episode also showed how tightly OpenAI has woven itself into local business. South Korea has the most paying ChatGPT users outside the United States, according to Reuters and KED Global. Software adoption and chip talk now sit in the same briefing. That mix is why a research remark can sound like a fab deal.
The closing point is simple. Scarcity made the market hear a foundry deal in a research remark. The walk-back put the deal back in the rumor bin, yet the pressure that created the rumor remains. Until TSMC capacity loosens, every major AI model shop will keep shopping for a second address on the wafer map. Samsung is on that short list for a reason, even when the press note says nothing new.
