New Software Cut AI Energy Use by Half Without Hardware Upgrades
New York, Saturday, 1 August 2026.
Zenith Flow Innovations demonstrated HXS, a software layer that reduces GPU energy consumption by up to 50.5% and boosts Microsoft’s Phi-4 throughput by 167% without physical infrastructure upgrades.
Zenith Flow Innovations Demonstrates HXS Efficiency
On July 29, 2026, Zenith Flow Innovations founder Brayon Michael Pieske demonstrated the HXS compute-efficiency layer at an AI glasses event in New York [1]. The software layer ran a 90-billion-parameter AI vision model offline on a smartphone, showcasing significant reductions in energy usage per GPU card without hardware upgrades [1]. Testing conducted on a single NVIDIA H100 80GB GPU using the vLLM 0.26.0 stack across five open AI models showed energy consumption reductions per card ranging from 17.2% to 27.4% [1]. Specific testing on Microsoft’s Phi-4 model demonstrated a 50.5% reduction in energy use, with board power dropping from 306 watts to 152 watts [1]. This reduction is calculated as 50.327 percent savings in power consumption [1]. Additionally, the throughput increased by 167%, rising from 109.8 to 293.7 requests per second at identical latency targets [1]. The throughput increase is derived from 167.486 percent growth in processing capacity [1]. Peak operating temperatures for the tested models declined by up to 15.0 degrees Celsius while maintaining identical model output accuracy [1].
Escalating Energy Demands in Data Centers
This technological advancement arrives as global data center electricity demand is projected to reach 945 TWh by 2030, according to the International Energy Agency [2]. U.S. electricity consumption for data centers is expected to account for 6.7–12% of the total by 2028, based on Department of Energy estimates [2]. AI workloads are driving data center power requirements to 120 kW per rack, rendering traditional air cooling systems obsolete [2]. Modern AI applications require power densities exceeding 50 kW per rack, with advanced training workloads reaching up to 120 kW per rack [2]. Securing grid power for new data centers in major markets, such as Northern Virginia, currently requires lead times of 3+ years [3]. Global data center electricity consumption is projected to rise from approximately 415 TWh to approximately 945 TWh by 2030, with peak power demand potentially hitting 130 GW by 2028 [3].
Comparative Efficiency Metrics
Implementing liquid cooling delivers 30–40% total energy savings compared to traditional air cooling, while integrating renewable energy can reduce operational costs by 25–50% [2]. California Energy Commission research indicates that advanced direct-to-chip liquid cooling systems can reduce cooling energy consumption by 60–80% compared to traditional air cooling [2]. In the networking sector, Nokia achieved a 75% reduction in power consumption while increasing system throughput by nearly 4x in the transition from its FP4 chipset to the FP5 generation [3]. European Internet Exchange NL-ix reported that implementing Nokia’s FP5-equipped routers reduced power consumption per Gbit from 0.9115 Watts to 0.1065 Watts [3]. This reduction represents a 88.316 percent decrease in power usage per unit of data [3]. Power efficiency enables higher performance density within existing power constraints and improves system reliability by reducing the number of components [3].
Future Deployment and Availability
The HXS technology will be presented in Silicon Valley during the early week of August 2026, with an expected window of August 3–7, 2026 [1]. It is being offered to one partner via potential exclusive license or acquisition [1]. HXS is a software-based compute-efficiency layer designed to reduce AI infrastructure energy consumption and cooling costs without requiring new hardware [1]. It is compatible with general server computing but currently verified only for AI-serving workloads [1]. Trust Carbon Infrastructure, the developer of HXS, was named one of five global winners of the 2025 DPI for People and Planet Innovation Challenge [1]. Verification of AI-serving performance metrics for HXS is cryptographically signed and independently timestamped [1]. Energy and cooling costs are becoming central constraints for companies expanding AI infrastructure [1]. HXS is designed to help existing hardware handle more work while lowering the energy required to support that demand [1].