AI Efficiency Layer Cuts Energy Use and Expands Server Capacity on Existing Hardware
PALO ALTO, Calif., Aug. 01, 2026 (GLOBE NEWSWIRE) — T ust Ca bo I f ast uctu e, ope ated by Ze ith Stream I ovatio s LLC, a ou ced sig ed measu eme ts showi g that its pate t-pe di g HXS compute-efficie cy laye educed e e gy use du i g AI se vi g o existi g NVIDIA H100 ha dwa e whereas i c easi g capability with out cha gi g mannequin outputs.
B ayo Michael Pieske p ese ts T ust Ca bo at a AI glasses eve t held i New Yo okay o July 29, 2026. Du i g the eve t, he i fo mally demo st ated the tech ology to atte dees, i cludi g a 90-billio -pa amete AI visio mannequin u i g e ti ely offli e o his sma tpho e—a functionality that might o mally equi e se ve a d GPU i f ast uctu e.
The tech ology might be p ese ted i Silico Valley du i g the ea ly week of August a d is bei g offe ed to o e pa t e th ough a pote tial unique lice se o acquisitio .
Testi g compa ed HXS ope ati g o a d off u de ide tical hundreds usi g o e NVIDIA H100 80GB g aphics ca d a d the vLLM 0.26.0 se vi g stack. Ac oss ope ly printed fashions f om Alibaba, Meta, DeepSeek, Google a d Mic osoft, e e gy use pe ca d fell betwee 17.2% a d 27.4%.
The fashions i cluded Alibaba’s Qwe 2.5-72B-I st uct, Meta’s Llama 3.3 70B, DeepSeek’s R1-Distill 70B, Google’s Gemma 3 27B a d Mic osoft’s Phi-4. Peak ope ati g tempe atu es decli ed by as a lot as 15.0 deg ees Celsius. The fashions a e ope ly printed by thei espective labs; othi g i the testi g implies e do seme t, pa t e ship o fu di g by a y of them.
I a deepe co figu atio i volvi g Mic osoft’s Phi-4, e e gy use decli ed 50.5%, with boa d powe falli g f om 306 watts to 152 watts. O the identical ha dwa e, th oughput i c eased f om 109.8 to 293.7 equests pe seco d on the similar late cy ta get, ep ese ti g a 167% capability i c ease. No late cy pe ce tile pe fo med wo se tha befo e, a d the mannequin p oduced ide tical esults with ze o e o s.
The e e gy savi gs va y acco di g to wo kload, a d each the 50.5% eductio a d the la gest capability i c ease we e measu ed o Mic osoft’s Phi-4. O ga izatio s evaluati g the tech ology ought to validate the esults usi g thei ow t affic befo e elyi g o the measu eme ts.
“E e gy a d cooli g prices a e becomi g ce t al co st ai ts fo compa ies expa di g AI i f ast uctu e,” stated B ayo Michael Pieske, fou de . “HXS is desig ed to assist existi g ha dwa e ha dle mo e wo okay whereas lowe i g the e e gy equi ed to suppo t that dema d.”
HXS is ot restricted to AI se vi g. The identical compute-efficie cy laye ca apply to se ve s a d othe fo ms of computi g, though o ly the AI-se vi g esults a e cu e tly sig ed a d measu ed. Adoptio does ot equi e the pu chase of ew ha dwa e.
The di ecto of BCG, JICA, a d Cadt ust. The awa ds ce emo y passed off du i g COP30
HXS is a compute-efficie cy laye developed by the workforce behi d T ust Ca bo I f ast uctu e, the ca bo -ve ificatio platfo m amed o e of 5 international wi e s of the 2025 DPI fo Folks a d Pla et I ovatio Challe ge, out of 540 sta tups f om 73 cou t ies. HXS is a sepa ate tech ology, measu ed o AI se vi g; it does ot ve ify ca bo . The fashions amed a e ope ly printed by thei labs; o e do seme t o pa t e ship is implied.
Ve ificatio : The AI-se vi g measu eme ts a e c yptog aphically sig ed a d i depe de tly timestamped; certified pa ties ca ve ify the whole recordsdata u de ag eeme t. Outcomes depe d o every ope atio ’s t affic; validate o you ow wo kload.
