OpenAI buys HBM4 directly
OpenAI struck direct HBM4 sourcing deals—an exclusive arrangement with Samsung for up to 800 million gigabits of 12‑layer modules starting in H2 2026. (x.com) It is also taking HBM from SK Hynix for a custom 'Titan' AI chip (designed with Broadcom) and for AMD MI450 accelerators under a reported 6GW agreement. (x.com) OpenAI and others have identified memory as the top bottleneck for training and inference after reported HBM price rises of about 60% in late 2025 and roughly 90% in early 2026. (x.com)
High-bandwidth memory is the stacked memory sitting next to an artificial-intelligence chip, and OpenAI is now trying to lock up that supply itself instead of leaving it to chip vendors. (news.samsung.com) (bloomberg.com) Samsung began mass production of HBM4 on Feb. 12, 2026, and said its 12-layer parts ship in capacities from 24 gigabytes to 36 gigabytes, with transfer speeds of 11.7 gigabits per second. Separate reports in March said OpenAI secured an exclusive Samsung arrangement for up to 800 million gigabits of 12-layer HBM4 starting in the second half of 2026. (news.samsung.com) (msn.com) OpenAI is also building out a second path for memory through SK hynix, according to reports tying that supply both to a custom “Titan” chip designed with Broadcom and to AMD’s Instinct MI450 systems. OpenAI and AMD said on Oct. 6, 2025 that they had signed a multi-year agreement to deploy 6 gigawatts of AMD GPUs, with the first 1 gigawatt of MI450 deployments due in the second half of 2026. (tech.yahoo.com) (openai.com) (amd.com) HBM matters because modern AI chips are often limited less by raw compute than by how fast they can pull model weights and activations from memory. Brad Lightcap, OpenAI’s chief operating officer, said at the Hill and Valley Forum on March 24 that “right now, it’s memory” when asked about infrastructure bottlenecks. (trendforce.com) (bloomberg.com) HBM4 is the next version of that memory, and the main change is a wider data path: 2,048 bits instead of 1,024 bits on earlier generations. That lets chip designers move more data per cycle, which is why Samsung, SK hynix and Micron have treated HBM4 as the key memory product for 2026 AI servers. (trendforce.com 1) (trendforce.com 2) The supply picture has tightened across the industry. TrendForce said in September 2025 that SK hynix and Micron were preparing HBM4 production ramps into 2026, while Samsung said in February 2026 that it had already started commercial HBM4 shipments to customers. (trendforce.com) (news.samsungsemiconductor.com) OpenAI’s move follows a broader shift in AI infrastructure, where model companies are reaching deeper into the hardware stack instead of buying finished systems and taking whatever memory allocation comes with them. OpenAI has already committed to AMD capacity, is developing custom silicon with Broadcom for internal use, and is now reported to be sourcing the memory layer directly as well. (openai.com) (tech.yahoo.com) (msn.com) That leaves 2026 looking less like a race for graphics processors alone and more like a race for the memory attached to them. OpenAI’s reported Samsung and SK hynix deals suggest the companies with guaranteed HBM allocations will be the ones that can actually turn planned AI capacity into running systems. (bloomberg.com) (openai.com) (news.samsung.com)
