Nvidia announced on the 24th (local time) at the semiconductor conference "Hot Chips 2026," held at Stanford University in the U.S., that its AI inference accelerator "Nvidia Grok3 LPX" had entered full-scale production.
Grok3 LPX is a chip developed based on the technology of Groq, an AI semiconductor company Nvidia indirectly acquired last year. It specializes in boosting the speed at which AI agents—which repeatedly perform complex computations—generate responses for users.
Samsung Electronics' foundry division handles production of the Grok3 LPX. Nvidia CEO Jensen Huang revealed in his keynote at "GTC 2026" this past March that "Samsung is making the Grok3 LPX." Samsung Electronics also formalized the collaboration at the time by unveiling a 4-nanometer wafer for the Grok3 LPX at its exhibition booth.
Currently, the Grok3 LPX is produced on Samsung Electronics' S5 line at its Pyeongtaek campus in Gyeonggi Province using the company's 4-nanometer process. S5 is Samsung Electronics' only dedicated foundry production line, built at the Pyeongtaek plant 2 (P2).
The industry is paying attention to how quickly Samsung's foundry raised the production yield of the Grok3 LPX. Yields are said to have recently climbed above 80%. Typically, when yields exceed 80% in advanced semiconductor processes, production is considered to have entered a stable mass-production stage.
Some analysts say this is the result of Samsung Electronics' continuous improvement of its 4-nanometer process. Samsung Electronics is known to have raised yields to over 90% on HBM base dies, which use the same 4-nanometer family process. As experience in producing AI semiconductors accumulates, manufacturing know-how that reduces process variability and raises the proportion of quality products has been applied to Grok3 LPX production, the explanation goes.
Recently, Samsung Electronics' Pyeongtaek 4-nanometer foundry line has maintained high utilization rates, buoyed by expanding demand for AI chips. Accordingly, Samsung Electronics is known to have recently raised its 4-nanometer foundry prices by up to 15%.
CEO Jensen Huang said, "Inference is the growth engine of AI," adding, "Through the Grok3 LPX, we will implement ultra-high-speed token generation, taking AI throughput, efficiency, and response speed up another notch."
