Samsung's memory gains on SK Hynix still hinge on Nvidia's supplier choices

In June, OpenAI made ChatGPT Enterprise and Codex available to every Samsung Electronics employee in Korea and every device-experience employee worldwide ^[10]. OpenAI called it "one of our largest to date" ^[11]. In September, OpenAI announced it was working with Samsung on next-generation chips ^[12].

The memory makers pitch technology. SK Hynix built its lead on more advanced core technologies and an early foothold with Nvidia ^[2]. Christopher Miller, the Tufts historian who wrote Chip War, said the contest is now judged on something else as well. "The question of technology leadership still matters, but now it is also assessed alongside capacity, because the entire memory industry faces constrained production capacity," he said ^[5]. Everyone outside the two Korean firms, Micron included, shares the remaining 17% of the memory chip market ^[1].

A buyer on Nvidia hardware gets a memory supplier chosen upstream. Ben Reynolds of Rhodium Group put the contest plainly: "Samsung's competitiveness with SK will depend, above all, on the extent to which it can poach share from SK as a supplier for Nvidia Vera Rubin" ^[14]. Samsung also sits on the customer side of Nvidia's allocation. Under a Korea-wide plan to deploy 260,000 Nvidia GPUs, it gets around 50,000, about a fifth ^[9][2].

Rest of World's speed-and-cost case gets half its support from its own sources ^[4]. Miller's point about tight capacity supports it ^[5]. In Reynolds' account, though, the memory race follows GPU allocation instead of setting it ^[13]. The report does not give the share figures behind "more than doubled," or any HBM prices or lead times ^[3].

Miller's second point bears on how buyers contract for memory. He said the leading DRAM makers, SK Hynix, Samsung and Micron, "are trying to use deals and long-term agreements with customers to make their HBM products more customized and shift from being a commodity business (as memory traditionally has been) to a more specialized product offering" ^[6].

One question for a capacity plan is who picks the memory: the accelerator vendor or the buyer. On Nvidia hardware, Reynolds' account puts that choice with Nvidia. He named a second route for Samsung: "Building relationships with alternative AI chip suppliers like AMD and Broadcom provide a secondary opportunity to build share" ^[15]. The other question is whether the timeline is covered by a committed allocation or agreement.

Vendor picks, allocation committed: memory risk sits inside the GPU delivery date. Vendor picks, nothing committed: tight memory capacity shows up as a slipping GPU date. Buyer has a say and has signed a long-term deal: the buyer is in the customized-HBM market Miller describes, with a reserved part and less room to move. Buyer has a say and has signed nothing: a specialized part is being bought on commodity terms, in a market Miller says is short of capacity ^[5][6].