Our flagship program, SETSUDEN, investigates low-bit AI models, energy-frugal learning, continual and edge learning, and rigorous methods for measuring their real energy cost.
Zeus AI Lab — Energy-Frugal Learning Research
Learning doesn’t need lightning. ZEUS AI LAB — Research Division of Zeus, Inc.
Measure learning in joules, not just FLOPs.
Zeus AI Lab is the research division of Zeus, Inc., exploring ways to measure the energy used by AI learning in joules — and reduce it.
power trace — making the energy cost of learning visible.
Research
Research
Research Areas
One flagship program, supported by three research pillars. We ask not only “How fast?” but also “What quality can we reach, and at what energy cost?”
Flagship Program Patent pending
SETSUDEN
Following the electricity constraints Japan faced after 2011, measuring power use and cutting what was unnecessary became a widely practiced form of energy conservation.
We bring that mindset into AI research. SETSUDEN explores how far the compute and energy required for learning can be reduced while preserving capability and quality.
Three Pillars
- 01
Low-bit LearningLow-bit, Resource-Frugal Learning
We study AI models that use low-bit representations, with the goal of reducing the compute and energy required for learning. We aim to make continued adaptation practical across diverse environments, including CPUs and low-power devices.
- 02
MeasurementMeasurement Methodology
We study reproducible ways to evaluate the energy efficiency of AI learning using hardware counters and external power meters. The goal is to compare systems not only by speed, but by the quality reached per joule consumed.
- 03
MethodologyHuman-led, AI-assisted Research
Research judgment, acceptance or rejection of hypotheses, and design responsibility remain with humans. Multiple AI systems support review, verification, and the search for counter-evidence. We maintain preregistration and audit records to improve reproducibility.
Principles
Principles
Research Principles
- 01
Measure FirstMeasure before we claim.
We do not draw conclusions from estimates alone. We combine hardware counters and external power meters, and confirm results with physical measurement whenever practical.
- 02
FalsifyDefine falsification criteria first.
Before an experiment, we define predictions and stopping criteria, then fix the conditions and analysis method before testing.
- 03
Whole-SystemCount the whole system.
We look beyond the core learning loop and account for surrounding processing and evaluation when judging total energy use.
- 04
Negative ResultsKeep negative results.
Mechanisms that fail — and the reasons why — are research assets that inform the next design.
- 05
ReproduciblePublish in stages.
After reviewing IP, confidentiality, and third-party licensing, we release the materials needed for reproducibility in stages.
Research Status & Disclosure
Status
Research Status & Disclosure
We filed a patent application in Japan for the core technology of SETSUDEN in October 2026. Validation and further IP protection for related work are ongoing. We share the direction of the research while disclosing technical details and quantitative results in stages.
Research Status
Validation in progress
This page describes research directions. It does not guarantee commercial performance or general energy savings. Quantitative results will be released after validation and the necessary IP protection.
IP & Disclosure Policy
Protect first. Publish openly.
We value reproducible research while also respecting the intellectual property needed to translate research into practice. Unpublished algorithms, implementations, experimental conditions, and performance figures will be disclosed in stages after the necessary protection and validation. The core technology of SETSUDEN is patent pending. Technical details will be released as the application is published and validation is completed.
Publications & Artifacts
Outputs
Research Outputs
- 01
Patent PendingSETSUDEN: Patent application on reducing power consumption in machine learning training
Filed with the Japan Patent Office in October 2026. Details will be made available after publication of the application.
- 02
Research Outputs In preparationSETSUDEN: Research on Energy-Frugal AI Learning
Quantitative results and technical details will be released in stages after validation and the necessary IP protection.
- 03
Reproducibility Package Under considerationMeasurement Data & Reproducibility Materials
The scope of release will be determined after reviewing intellectual property, confidentiality, and third-party licensing requirements.
Mission
Democratizing Compute
A world in which AI can be trained only by those with massive clusters separates those who can participate from those who cannot.
We aim for AI that can continue learning on CPUs and other accessible hardware.
By reducing energy requirements, we hope to expand the places where learning AI can exist — including regions constrained by power infrastructure or capital.
Zeus has provided free IT education in Myanmar since 2013, and that work comes from the same belief.
Generate genius.
Genius can emerge anywhere. Compute should not stand in the way.
Lab Overview
Overview
About the Lab
- Name
- Zeus AI Lab — Research Division of Zeus, Inc.
- Established
- 2026
- Location
- 4F-BC, 2-15-17 Nishi-Shimbashi, Minato-ku, Tokyo 105-0003, Japan
- Lab Director
- Masanori Katsura
- Research Areas
- Energy-frugal AI learning / Low-bit AI models / Continual learning & edge AI / Energy measurement methodology
Flagship Program: SETSUDEN — patent pending - Operated by
- Zeus, Inc. (founded 2010 / AI, Web3 and systems development)
Contact
Research collaboration, media & careers.
For research collaboration, access to measurement environments,
media inquiries, or career opportunities, please contact Zeus, Inc.