OpenAI will work with Samsung Electronics to develop its next-generation AI accelerators, a collaboration that could address critical shortages in both memory and manufacturing capacity, according to a report published by The Register on September 9. Harrison Kim, OpenAI's Korea general manager, revealed the partnership at a Wednesday press conference, though the company later clarified the comments referenced an existing relationship from last year rather than a fresh agreement. The tie-up arrives weeks after OpenAI's first-generation "Jalapeño" accelerators outperformed Nvidia's Blackwell GPUs across multiple inference tasks.
The collaboration centers on joint production and research for chips OpenAI is currently developing, with Samsung positioned to supply high-bandwidth memory (HBM) or fabricate logic chips on its 2 nm process technology. OpenAI's Jalapeño chip uses HBM4, the same memory technology deployed in AMD and Nvidia's latest GPUs, while the successor design—nicknamed Habanero by The Register—will likely adopt the faster HBM4E standard. Only three manufacturers produce HBM at scale: SK Hynix, Micron, and Samsung Electronics. Samsung operates as the world's second-largest leading-edge foundry, trailing only TSMC, which currently fabricates the vast majority of cutting-edge silicon for AMD, Nvidia, and most hyperscalers designing custom chips. Nvidia's Ampere generation relied on Samsung's 8 nm process for most PCIe-based cards, though HBM-equipped accelerators came exclusively from TSMC.
According to The Register, OpenAI faces intense competition from larger, higher-volume chip designers for both logic and memory silicon, putting the ChatGPT maker in "a tough spot." The report notes that securing TSMC capacity means competing with practically every chip company and hyperscaler, while the HBM supply remains similarly constrained. OpenAI's next chip will probably use custom base dies to shift memory controllers or logic from compute silicon to the HBM stack itself—the same approach Nvidia disclosed with NVHBM—and these custom base dies require coordination with the HBM supplier. Samsung has also proven it can handle sophisticated AI accelerators, with chip designers including Rebellions, Tenstorrent, and Groq building advanced designs at its fabs in 2026.
The partnership offers OpenAI a path around bottlenecks that have made TSMC the sole viable option for most designers, The Register explains. Multi-die architectures and widespread HBM adoption pushed chip companies toward TSMC after reported yield challenges hit Samsung's 3 nm process technology, but Samsung's capabilities have grown more compelling as proof points emerge. Rebellions' Rebel100 accelerator demonstrates Samsung can manufacture multi-die designs with HBM3E using sophisticated advanced packaging techniques to stitch components together. Manufacturing at Samsung could unlock larger silicon allocations than OpenAI would receive from TSMC alone, giving the company breathing room as it scales production. Forging relationships with memory suppliers early also positions OpenAI to secure ample HBM4E supply despite ongoing shortages, even if the compute chiplets themselves end up fabricated by TSMC. The dual-supplier strategy reduces OpenAI's dependence on a single foundry while ensuring access to the specialized memory its designs require. For fabless chip designers entering a market dominated by giants with massive volume commitments, diversifying manufacturing partnerships isn't just strategic—it's existential.