Microsoft released an analysis saying the electricity used to send one question to a large language model can be far lower than widely cited figures. Assuming a 300-token response, it put the median at 0.31 watt-hours, with the middle 50 percent ranging from 0.16 to 0.60 watt-hours. It said earlier estimates may have overstated usage by 4 to 20 times by not reflecting batch processing and real-world GPU utilisation.

Applied Materials has introduced two new semiconductor process systems aimed at addressing uniformity challenges in high-stack 3D chip structures as demand for AI chips grows. The company unveiled the Centris Spectral SiN atomic layer deposition system and the Producer Selectra molybdenum etch system. It said the tools support transitions to vertical chip architectures such as GAA transistors and high-stack 3D NAND by improving deposition and etch uniformity in deep, narrow structures.