OpenAI and Broadcom recently released the first public benchmark data for their jointly developed AI inference chip, Jalapeño. The chip demonstrated AI operations per watt that are 1.5 to 1.9 times higher than comparable NVIDIA systems, with response times 1.7 to 3.6 times faster. Small-scale deployment of the chip is planned for later this year, with mass production ramp-up expected in 2027. Rich Privorotsky, head of trading at Goldman Sachs, analyzed that Jalapeño is not an "NVIDIA killer," but its emergence signifies a shift in the AI hardware market from high concentration to diversification. The inference side is becoming the main battleground for custom chips, which could lead to a gradual erosion of NVIDIA's long-term profit structure. OpenAI has explicitly stated that the chip is a complement to partners like NVIDIA, not a replacement, and NVIDIA's moat in training chips and the CUDA ecosystem remains unchallenged in the short term.