OpenAI Says Its Own Chip Now Beats Nvidia’s Best on Energy Efficiency
At the Hot Chips conference, OpenAI published the first benchmarks for “Jalapeño,” the inference chip it designed in-house, claiming 1.5x to 1.9x more AI work per watt than Nvidia’s Blackwell systems and up to 3.6x lower response latency. Independent analysts who verified some of the runs noted the comparison is flattering — Jalapeño uses newer memory than the Nvidia systems it was tested against — but said it still edges out Nvidia’s next-generation platform. It is the clearest sign yet that OpenAI intends to stop renting all of its computing power from its biggest supplier.