Everyone is asking Pharaoh, OpenAI has developed a self-designed chip called "Jalapeño" (Mexican chili), which directly outperforms Nvidia's GB300. Is this going to take Huang's job? Pharaoh directly said that OpenAI finally figured it out: instead of paying Nvidia hundreds of billions in "compute tax" every year, it's better to make their own knife. How powerful is this "Mexican chili"? With 700 watts power consumption, it beats Nvidia's 1400 watts, delivering 1.5 to 1.9 times the throughput per watt compared to GB200/GB300, and reduces end-to-end latency by 1.7 to 3.6 times. The independent semiconductor research institute SemiAnalysis tested it and said: it has defeated all Nvidia, AMD, and Google chips we can measure. The inference cost is expected to drop by 50%, with a total cost of ownership per chip per hour of only $1.56, on par with H100's $1.55, while Nvidia's next-generation Vera Rubin costs as high as $3.61. OpenAI didn't act on a whim; nine months ago, they teamed up with Broadcom to start development, using TSMC's 3nm process and HBM4 memory, designed specifically for AI inference. The real kicker is that during the chip design process, OpenAI's own large models participated in architecture optimization. The chip is tailored at the transistor level to OpenAI's model characteristics, rather than pursuing generality like Nvidia. OpenAI plans small-scale deployment by the end of 2026 and large-scale volume in 2027. Like Pharaoh's pyramids, it wasn't built in a day, but the foundation is already laid