Harrison Kim kept the numbers sparse. The head of OpenAI’s Korean operations told an audience in Seoul on Tuesday that the company is working more closely with Samsung Electronics on semiconductors, and that joint production and research on next-generation chips is the most advanced part of the relationship. He declined to name the chip, its timeline, or its cost. The one figure he offered was a sizing: Samsung is “one of the largest enterprises using ChatGPT at scale globally.”
The Seoul appearance, reported on September 9, came the same week that OpenAI’s chief financial officer, Sarah Friar, laid out the company’s strategy in plainer terms at a Goldman Sachs conference in San Francisco. On one side, OpenAI is cutting prices to fill its user base. On the other, it is building its own silicon to take control of its largest single cost.
Friar told the Communacopia conference that OpenAI is pushing AI into vertical industries such as chip design, life sciences, and financial services, and that it is experimenting with charging customers for business results rather than raw usage. She said the company’s in-house chip, code-named Jalapeno, was designed with OpenAI’s own models and reached tape-out in nine months. Tape-out is the point at which a design is handed to a factory, and the speed suggests the silicon program is further along than the company had previously let on.
The pricing message was blunter. Friar said usage of the low-cost model Luna climbed roughly tenfold after OpenAI cut its price by 80 percent, and that running Luna in the cloud now undercuts open-weight rivals such as Zhipu’s GLM 5.3. She put Codex at 25 million users and said enterprise revenue grew 32 percent in June and July, well ahead of the company’s overall annualized revenue growth of 20 percent.
OpenAI has not said publicly which fabricator will produce Jalapeno at scale. A tighter Samsung partnership gives the company a manufacturing option as it races against rivals that already buy chips in enormous volume. Samsung, for its part, needs high-volume design wins to keep its foundry business competitive against TSMC, which builds most of the industry’s advanced AI accelerators.
The two fronts are connected. Cheaper models draw in customers and their workloads, which generate the data and revenue that fund custom silicon. Custom silicon, if it works, lowers the cost of serving those same workloads. Analysts have argued for years that OpenAI’s dependence on outside chips is a structural weakness; this week’s moves suggest the company is trying to turn that weakness into a margin story.
Behind the price cuts is a straightforward calculation. Every dollar OpenAI does not spend on serving its models is a dollar it can put toward the next training run, and the company’s largest suppliers have little incentive to make inference cheaper for a customer of its size. Building its own accelerator, and lining up a foundry to build it, is the long answer to a problem the price cuts only paper over in the short term.
OpenAI has been assembling the pieces for its own silicon for some time. Reuters reported last year that the company was working with Broadcom on custom accelerators, and it has quietly hired a chip design team led by Richard Ho, a former Google engineer who worked on its tensor processing units. Jalapeno is the first public confirmation that the program has produced a design fast enough to send to a factory.
None of it comes cheap. A chip that reaches tape-out in nine months is still years from proving itself inside a data center, and a price war, if Luna’s trajectory is any guide, burns cash before it returns any. Friar’s framing was that OpenAI can afford both: grow users now, pay less to serve them later.
For Samsung, the courtship matters beyond a single customer. The Korean company’s chip business has struggled to win leading-edge orders for AI accelerators, and a marquee design win from a company of OpenAI’s scale would send a signal to the rest of the market. The Seoul event was, in that sense, as much about Samsung’s foundry ambitions as OpenAI’s.
Analysts caution that a joint chip program with Samsung is not a guaranteed win. Samsung’s most advanced nodes have trailed TSMC’s on yield and power efficiency in recent years, and a design win does not by itself close that gap. OpenAI could still end up paying TSMC for most of its leading-edge silicon even as it builds a second source with Samsung. The Seoul remarks describe a deepening relationship, not a finished product.
Kim offered no forecast for how many chips or data centers a joint program would yield. He offered a different kind of statement: that OpenAI and Samsung are not merely transacting but jointly producing, a word that implies a longer commitment than a purchase order. For a company whose chief cost is compute, that commitment may be the part of the week that matters most.
