Samsung and SK join OpenAI’s Stargate initiative to advance global AI infrastructure

OpenAI just widened its Stargate coalition, and this time the new members aren't chipmakers in the usual sense — they're the memory giants. Samsung Electronics and SK, parent of SK Hynix, are signing on to help scale production of advanced memory chips and to help stand up next-generation data centers on Korean soil. That's a notable shift for an initiative that started out mostly as a story about Nvidia GPUs and American real estate.

Memory has quietly become the bottleneck nobody outside the industry talks about. Training and running today's largest models eats through high-bandwidth memory at a pace that's outstripped supply for the better part of two years. Samsung and SK Hynix between them control the overwhelming majority of the world's HBM output, so bringing them formally into Stargate isn't a courtesy invite — it's OpenAI trying to lock in the one resource that could throttle its expansion plans regardless of how much compute it secures elsewhere.

Building data centers in Korea also changes the geography of Stargate in a useful way. Up to now the project's biggest headline sites have been American, tied to Oracle and SoftBank money and Texas power grids. A Korean footprint gives OpenAI a foothold closer to Asian markets and closer to the fabs actually producing the memory it needs, cutting down on the logistics of shipping chips halfway around the planet before they even reach a rack.

For Samsung and SK, there's an obvious upside too. Neither company wants to be a pure component supplier watching Nvidia and TSMC capture the margin story of the decade. Getting named as infrastructure partners in one of the most watched AI buildouts on earth is as much a marketing coup as an engineering one, and it gives them leverage in future pricing conversations with every other AI lab shopping for HBM.

What's left unsaid, as usual with Stargate announcements, is the actual dollar figure and timeline. OpenAI has a habit of unveiling partners before locking down specifics, and this one reads the same way — big names, vague scale, details to follow.