HXS cuts Phi-4 serving energy 50.5%

Trust Carbon Infrastructure says its HXS efficiency layer cut Microsoft Phi-4 serving energy 50.5%, from 306 watts to 152 watts, while lifting throughput from 109.8 to 293.7 requests per second on one H100 at the same latency target. The company’s signed-measurement announcement also reports 17.2%–27.4% lower energy per card across five open models, but the files are available only to qualified parties under agreement and the deepest result comes from a short, single-card test. Operators following the shift from token abundance to inference rationing should treat HXS as a benchmark candidate, not capacity planning truth, until the method and workload reproduce independently.