OpenAI has found ways to run its models on far fewer Nvidia chips, cutting inference costs by more than half, even as Nvidia (NVDA) shares rose. According to BeInCrypto, The Information reported OpenAI engineers achieved the savings via new software optimization methods, while OpenAI and Broadcom (AVGO) on June 24 unveiled Jalapeño, OpenAI’s first custom chip.
OpenAI said early tests show better performance per watt, with first deployments at a gigawatt scale by the end of 2026 and Microsoft as lead partner. BeInCrypto also cited Meituan training its 1.6 trillion parameter LongCat-2.0 model on domestic chips without Nvidia hardware.