AINEWS
Custom AI silicon won't break Nvidia's lock in alone. High bandwidth memory is the rate limiting step, and Samsung is one of the few companies that can build it at scale.
OpenAI's next training run will be limited less by the chips it designs than by the memory that sits next to them. That is the practical meaning of The Register's report that Samsung is dedicating advanced memory production and engineering collaboration to OpenAI, re-reported by WebProNews. The "reduce Nvidia dependence" framing is real, but the lever is memory, not chip design.
The memory in question is called high-bandwidth memory, or HBM. It is not the RAM in a laptop or phone. It is a stack of DRAM dies bonded to a base die and connected through thousands of vertical interconnects, designed to feed a processor data fast enough that the processor never sits idle. The current generation, HBM3E, is shipping in limited volume. The next generation, HBM4, is what OpenAI's reported roadmap and Samsung's production plans both target. Early public test results from Samsung's lines suggest HBM3E and HBM4 modules can deliver more than 1.5 terabytes per second of bandwidth per chip. The Register's reporting indicates OpenAI's follow-on accelerator, which the outlet has labeled "Habanero" (a publication placeholder rather than an official OpenAI name), is expected to use 12-layer HBM4 stacks.
OpenAI has been hiring chip engineers from Google and elsewhere for years, and the company showed its first inference chip, code-named Jalapeño, at Hot Chips last month. None of that chip-design work changes the underlying problem: a useful AI accelerator still has to be paired with specialized stacked memory, and only a small number of fabricators worldwide can produce the latest generations at scale. Move enough of those modules next to enough processors and you have a cluster. The question is who gets allocation when global supply is short.
OpenAI's strategy on this point is openly hybrid. The company still buys Nvidia H100 and Blackwell GPUs, will continue to do so for training, and is now layering Samsung memory and engineering collaboration underneath its custom silicon. The shape of that collaboration is not on the public record in detail; WebProNews's framing of "multi-year purchase agreements" reflects analyst speculation, not an OpenAI or Samsung disclosure. What is on the record is the direction: Samsung has announced plans to invest more than $40 billion in new production capacity, much of it aimed at AI-related technologies, and OpenAI is valued at more than $150 billion in recent funding rounds, with capital earmarked for infrastructure that could contain hundreds of thousands of processors.
There is a secondary tension inside Samsung itself. TechTimes reports that OpenAI confirmed the Samsung chip program and that Samsung has reserved roughly half of its best available production node for its own memory business, separate from the capacity it sells to foundry customers. That is the supply-competition sub-angle: Samsung's memory division and its foundry customers are now competing for the same advanced output, and OpenAI is one of the few buyers with the capital to make that allocation interesting. Samsung also operates fabs in South Korea, the United States, and other locations, so any shift in advanced memory allocation interacts with US-China trade restrictions and with the foundry diversification pushes that have defined chip policy since 2022.
The critique that should survive this story is also the honest one. Custom AI silicon timelines remain unproven. Nvidia's grip on training compute is intact for the foreseeable future, and the Blackwell ramp is the immediate constraint on OpenAI's largest clusters, not memory. HBM supply is tight industry-wide, and Samsung is one of three credible suppliers, which means Samsung's commitments to OpenAI come partly at the expense of other buyers. None of that changes the central fact. When OpenAI's next major model demands several times the compute used for GPT-4, the binding constraint will be how many HBM4 stacks Samsung, SK Hynix, and Micron can ship, not how many accelerators the design team can tape out. The deal does not let OpenAI escape Nvidia. It lets OpenAI reach the next training run.