AINEWS
OpenAI named Samsung a 'joint production partner' for next gen chips days after Broadcom shipped the first. A July foundry deal suggests Samsung work runs through Broadcom's contract, not around it.
Broadcom is down 13.74% over the past month after OpenAI's Sept. 10 Samsung announcement. On the same Sept. 2 earnings call, chief executive Hock Tan said Broadcom shipped OpenAI's first custom AI inference chip, internally codenamed Jalapeño, during the quarter, and is already developing a third-generation version (247 Wall St). Both facts are true at once. An AI chip is not one thing; it is a stack of four layers, and OpenAI's announcement moved only one of them.
The four layers are the chip's logic design, the foundry that fabricates it, the high-bandwidth memory (HBM, the stacked DRAM that sits next to the accelerator), and the switch silicon that ties thousands of accelerators into a cluster. Broadcom's moat sits in layer one and layer four: the co-design work with OpenAI, and the Ethernet networking stack that makes a cluster behave like one machine. Hock Tan's Sept. 2 pitch to investors was a layer-one and layer-four pitch: planned 1.3 gigawatts of Jalapeño-class capacity in 2027, line of sight to more than 5 gigawatts in 2028, and a third-generation XPU (Broadcom's term for a customer-specific AI accelerator, distinct from a general-purpose Nvidia GPU) already in development (247 Wall St).
The foundry layer is where Samsung enters. OpenAI Korea general manager Harrison Kim said in Seoul on Sept. 10 that Samsung is a "joint production and research partner" for OpenAI's next-generation custom AI chips (TechTimes). Samsung is also a major high-bandwidth memory supplier, putting the partnership in play across more than one layer of the stack.
On July 25, 2026, Broadcom and Samsung signed a memorandum of understanding worth more than $200 billion, covering 2-nanometer and below foundry, memory, and advanced packaging through 2030 (TechTimes). If that agreement covers OpenAI's next chips, the Samsung foundry work would run through Broadcom's existing contract rather than a separate OpenAI-Samsung deal. Broadcom already holds the Samsung slot.
On Sept. 2, Broadcom projected AI semiconductor revenue of $115 billion in fiscal 2027 and $230 billion in fiscal 2028, with OpenAI on track to become Broadcom's second-largest XPU customer behind Anthropic (247 Wall St). Broadcom closed at $364.38 against an analyst price-target average of $533.41, on a trailing multiple around 47 times earnings. The $230 billion figure is a Broadcom forecast, and it assumes OpenAI stays a marquee customer. It does not bake in a Samsung second-source scenario where Broadcom's per-chip economics get renegotiated.
OpenAI's Jalapeño first-results post reports 1.5 to 1.9 times more AI work per watt at peak throughput, and 1.7 to 3.6 times lower end-to-end latency, than comparison systems across GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, with a 700-watt chip rating and sustained operation at or below 550 watts (OpenAI). ServeTheHome's Hot Chips 2026 live blog puts those numbers against the SemiAnalysis InferenceX benchmark, the industry standard for inference cost (ServeTheHome). Those are shipped numbers, not roadmap numbers.
The question the reader can carry forward: when a customer names a "joint production partner," does the foundry relationship run through the accelerator designer's existing contract, or has the designer just become the second source? The market is pricing the second answer at a 13.74% monthly drawdown. The contract structure points to the first. The next tell is whether OpenAI's third-generation XPU disclosure names Samsung as a design lead, which would break the layers-of-the-stack read. The public record so far does not support that. The same question is now on the table for every hyperscaler custom-silicon program: the foundry second-source move, by itself, does not have to mean the designer is out.