Samsung Electronics' foundry division has sharply improved the production yield of Nvidia's Grok3, an AI inference accelerator (chip for running trained models) built on a 4-nanometer process. The yield is estimated to have risen two to three times from April, or from the second quarter of this year, to above 80%—a level generally viewed as sufficient for stable mass production. Nvidia announced at Hot Chips 2026 in California that Grok3 had entered full-scale mass production. In Nvidia benchmarks, the chip processed a 100,000-token context on Google's Gemma4 31B model while generating 3,400 output tokens per second, the fastest result recorded on that model and up to four times faster than competing platforms. Produced at Samsung's Pyeongtaek S5 foundry line, Grok3 reflects broader improvements in Samsung Foundry's 4nm manufacturing stability, utilization and profitability. Samsung expects AI and high-performance computing to account for more than 30% of foundry revenue this year, while leading-edge processes are expected to represent more than half of sales in the second half. The company has also entered the design stage for 2nm projects and reportedly raised prices for new 4nm orders from U.S. and Chinese customers by up to 15%.