NVIDIA’s HGX B200 platform marks a significant leap in both performance and environmental sustainability, delivering a 24% reduction in embodied carbon emissions intensity compared to its predecessor, the HGX H100. This improvement is based on newly published Product Carbon Footprint (PCF) summaries that analyze the full lifecycle of the systems, from raw material extraction to manufacturing. The HGX B200 and HGX H100 are both eight-GPU accelerated computing platforms designed for AI training, inference, and high-performance computing. They leverage NVIDIA’s NVLink and NVSwitch technologies to enable high-bandwidth communication between GPUs. However, the B200 introduces substantial advancements that drive both efficiency and sustainability. At the core of the improvement is the new NVIDIA Blackwell B200 GPU, which features 180 GB of HBM3E memory—more than double that of the H100. It also includes a second-generation Transformer Engine supporting FP4 precision alongside FP8, along with fifth-generation NVLink and NVSwitch, delivering up to 1.8 TB/s per GPU and 14.4 TB/s aggregate bandwidth. These enhancements result in a 2.3x increase in FP16 throughput over the H100. This performance boost is also more energy efficient. For AI inference workloads, the HGX B200 can be up to 15 times more energy efficient than the H100—equivalent to a 93% reduction in energy use for the same task. This translates into a dramatic drop in operational carbon emissions. For example, processing one million inference tokens using the DeepSeek-R1 model on the HGX B200 results in a 90% reduction in operational carbon emissions compared to the H100, based on 2023 IEA emission factors weighted by regional data center energy mix. Embodied carbon intensity—the emissions generated during manufacturing—also declined from 0.66 gCO2e per exaflop (FP16) for the H100 to 0.50 gCO2e per exaflop for the B200, a 24% improvement. This reduction is driven by lower emissions in key material categories, particularly thermal components, integrated circuits, and memory. The PCF data was gathered using primary supplier data for over 90% of materials by weight, supplemented by industry-standard life cycle assessment databases such as ecoinvent 3.10, Sphera LCA, and imec.netzero. The analysis follows ISO 14040, 14044, and 14067 standards for life cycle assessments and carbon footprinting. NVIDIA continues to prioritize transparency and sustainability, committing to reduce the environmental footprint of its products with each new generation. The HGX B200 exemplifies this mission—delivering transformative AI performance while significantly lowering both embodied and operational carbon emissions. As AI workloads grow, such advancements are critical to building a more sustainable future for computing.