Analysis
OpenAI and Samsung Electronics are expanding a partnership to co-develop OpenAI's next-generation AI accelerator chip, according to The Register, which cited comments from OpenAI Korea general manager Harrison Kim at a September 9 press conference. Kim said the companies have made "the most progress" on their joint production and research for the chip informally called Habanero, though neither side disclosed a dollar figure.
The deal follows OpenAI's disclosure of Jalapeno, its first-generation in-house inference chip, which the company claims outperforms Nvidia's current Blackwell GPUs on certain workloads -- a notable claim given OpenAI has spent years as one of Nvidia's largest customers rather than a rival chip designer. Habanero, the next chip in the line, is built around HBM4 and HBM4E high-bandwidth memory, a component Samsung is one of only three companies worldwide equipped to manufacture at scale, alongside SK Hynix and Micron.
The bear case is straightforward and the announcement itself doesn't address it: neither company disclosed a dollar figure or a signed-contract commitment, only comments at a press conference, and Samsung's advanced-node foundry yields have historically trailed TSMC's on comparable programs. Jalapeno and Habanero are also inference-only chips -- they don't touch the Blackwell and Rubin GPUs OpenAI still needs for frontier-model training, so even a fully realized partnership leaves OpenAI's largest compute expense, training, exactly as Nvidia-dependent as before.
“## Why OpenAI needs Samsung specifically The partnership solves two distinct supply-chain problems at once.”
Why OpenAI needs Samsung specifically
The partnership solves two distinct supply-chain problems at once. First, memory: HBM has been one of the tightest bottlenecks in AI hardware for two years, with SK Hynix and Samsung both effectively sold out of near-term HBM4 capacity to Nvidia, AMD and the hyperscalers already. Second, foundry capacity: TSMC remains the only foundry reliably producing at the most advanced nodes at scale, and every major AI chip designer competes for the same constrained wafer allocation. Samsung's foundry business, chronically the industry's distant number two, is offering OpenAI an alternative path to 2nm-class production that doesn't require waiting in TSMC's queue.
The custom-silicon trend OpenAI just joined
OpenAI is a late entrant to a strategy Big Tech has pursued for years:
- Google -- TPUs, now on their seventh generation, remain the most mature custom AI chip program among the hyperscalers, reducing Google's Nvidia dependence for its own workloads.
- Amazon -- Trainium and Inferentia chips power a growing share of AWS's internal AI training and inference, with Anthropic among the largest disclosed users.
- Meta -- MTIA chips handle a portion of Meta's recommendation and inference workloads, though the company remains a massive Nvidia GPU buyer overall.
OpenAI's move into custom silicon differs from its peers in one respect: it doesn't run its own cloud infrastructure at hyperscaler scale, meaning Jalapeno and Habanero have to prove out inside Microsoft Azure and OpenAI's own Stargate data-center buildout rather than a vertically integrated cloud business.
None of this changes OpenAI's Nvidia relationship in the near term -- Jalapeno and Habanero are inference-focused chips, not full replacements for the Blackwell and next-generation Rubin GPUs OpenAI still needs for frontier model training, and Nvidia remains the overwhelming majority of OpenAI's compute spend today. Samsung's foundry yields on advanced nodes have also historically lagged TSMC's, a gap a partnership announcement doesn't automatically close.
The next milestone to watch is whether Habanero chips actually tape out and ship at volume, and whether Samsung's involvement measurably eases the HBM shortage constraining every AI lab's compute buildout this year, or whether this remains, for now, a press-conference commitment ahead of the harder engineering work.
