OpenAI patents 20‑HBM chip
- OpenAI filed a patent for a custom chip concept that stacks 20 HBM memory modules connected via EMIB bridges. (x.com) - The architecture is designed to dramatically increase memory bandwidth for large-model training and inference. (x.com) - The patent joins moves by Google and others building custom silicon to cut AI compute costs and scale data-center performance. (x.com)
Artificial intelligence chips spend much of their time waiting on memory, and an OpenAI patent application published on April 2 lays out a way to attach far more of that memory to one processor. (patents.justia.com) The filing, assigned to OpenAI OpCo, says embedded logic bridges could connect high-bandwidth memory chiplets to compute chiplets over longer distances than the roughly 6 millimeters allowed by the JEDEC memory standard. It was filed on October 1, 2024, and lists Clive Chan, Chian-min Richard Ho, Christopher Leary, Ravi Narayanaswami, Devin Persaud and Kaushik Vaidyanathan as inventors. (patents.justia.com; patents.justia.com) In the patent’s examples, OpenAI shows a compute chiplet surrounded by as many as 20 high-bandwidth memory stacks, a layout aimed at feeding more data to large-model training and inference systems. The bridges are described as active links with high-speed communication circuits that keep signals from fading over longer paths. (patents.justia.com; wccftech.com) High-bandwidth memory, or HBM, is stacked memory placed close to a processor so it can move data faster than conventional server memory. For AI systems, that matters because larger models and longer context windows push chips to shuttle huge amounts of data, not just perform arithmetic. (patents.justia.com) The filing is a patent application, not a product launch, and it does not mean OpenAI will ship this exact package. Patent records describe one hardware bottleneck OpenAI’s own filing names directly: machine-learning workloads “depend on access to large amounts of memory for efficient operations.” (patents.justia.com) OpenAI has already been moving beyond buying off-the-shelf accelerators alone. Reuters reported in February 2025 that OpenAI was working with Broadcom on its first in-house chip, with manufacturing planned at Taiwan Semiconductor Manufacturing Co., and said the effort was led by former Google chip executive Richard Ho. (economictimes.indiatimes.com) That effort became public on October 13, 2025, when OpenAI and Broadcom said they would develop and deploy custom processors starting in the second half of 2026. Reuters said the companies planned 10 gigawatts of custom chips, placing OpenAI alongside Google and Amazon in the push to build proprietary AI silicon. (nbcnews.com) Google has been patenting memory-heavy chiplet designs too. One Google patent application published on May 1, 2025 covers “multi-directional sharing and multiplexing” for high-bandwidth memory, another sign that chip packaging and memory access have become central design fights in AI hardware. (patents.google.com) The immediate takeaway from OpenAI’s filing is narrower than the online hype: it shows the company is spending patent effort on packaging, memory links and chiplet layouts, not just model software. If OpenAI’s future chips arrive on the 2026 timetable its partners have discussed, this application shows what problem they are trying to solve first. (patents.justia.com; nbcnews.com)
