OpenAI AI Chip Patent Reveals EMIB Design With Up to 20 HBM Stacks

OpenAI is advancing its in-house chip and advanced-packaging efforts to reduce its reliance on external AI accelerators. Large-language-model training is constrained by memory capacity, bandwidth and die-to-die interconnect efficiency. The design uses Embedded Multi-die Interconnect Bridge (EMIB) technology to boost high-speed data transfers, with implications for future computing costs and model scalability.

As of July 20, 2026, the patent disclosures showed that OpenAI planned to use an EMIB packaging architecture integrating compute dies with up to 20 high-bandwidth memory (HBM) stacks. Compared with common existing configurations, the design focuses on expanding memory capacity and bandwidth, but no chip production timeline, foundry partner or investment amount has been announced.