AI Efficiency Layer Cuts Energy Use and Expands Server Capacity on Existing Hardware

Trust Carbon Infrastructure, operated by Zenith Flow Innovations LLC, has announced significant advancements with its patent-pending HXS compute-efficiency layer. This technology reportedly reduces energy use during AI serving on existing NVIDIA H100 hardware while simultaneously increasing capacity without altering model outputs.

Brayon Michael Pieske demonstrated the technology at an AI glasses event in New York on July 29, 2026. He showcased a 90-billion-parameter AI vision model functioning entirely offline on a smartphone, a capability typically reliant on server and GPU infrastructure. Further presentations are planned in Silicon Valley during the first week of August, with discussions of an exclusive license or acquisition by a selected partner.

Testing of the HXS layer involved comparing its operations under identical loads using one NVIDIA H100 80GB graphics card and the vLLM 0.26.0 serving stack. Results indicated that energy usage per card decreased by 17.2% to 27.4% across various models, including Alibaba’s Qwen2.5-72B-Instruct and Meta’s Llama 3.3 70B, among others. In peak conditions, operating temperatures were reduced by as much as 15.0 degrees Celsius.

In a detailed configuration utilizing Microsoft’s Phi-4 model, energy use was found to decrease by 50.5%, with board power dropping from 306 watts to 152 watts. Throughput also improved significantly, increasing from 109.8 to 293.7 requests per second while maintaining the same latency target, resulting in a 167% capacity increase. The model produced identical outcomes without any errors, and no latency percentile performed worse than prior measurements.

Pieske commented on the technology’s relevance in today’s market, stating, “Energy and cooling costs are becoming central constraints for companies expanding AI infrastructure,” adding that HXS is engineered to help existing hardware handle increased workloads while lowering the energy required. He noted that the HXS technology is versatile and applicable beyond AI serving, but its current measurements are limited to that domain.

The HXS compute-efficiency layer is part of the Trust Carbon Infrastructure initiative, recognized as one of the five global winners in the 2025 DPI for People and Planet Innovation Challenge among 540 startups from 73 countries. While the measurements related to AI serving are independently verified and timestamped, organizations interested in the technology are encouraged to validate the results through their own operational traffic.