Originally published at emphosgroup.com
The AI industry will consume more electricity this
year than the entire country of the Netherlands. That
number will double within three years. By 2030 AI
data centers are projected to consume between 85 and
134 terawatt-hours of electricity annually — comparable
to the total electricity consumption of a mid-sized
nation, spent entirely on computation.
The response from the industry has been consistent:
renewable energy procurement, carbon offset programs,
commitments to net zero by 2030, and press releases
about more efficient chips. These are real efforts
made by organizations that understand the problem.
They are also insufficient — not because the
organizations are insincere, but because they are
treating symptoms of a structural condition they
believe they cannot change.
EMPHOS Group is changing it. Not by making the
existing architecture more efficient. By replacing
the architecture entirely.
WHERE THE ENERGY ACTUALLY GOES
To understand why the industry's response is
insufficient, you have to understand where the
energy goes.
A large language model stores knowledge as numerical
parameters — billions of floating-point weights
distributed across a matrix. Every query requires
multiplying the input representation by those weights,
applying activation functions, and passing the result
through dozens of layers of computation. This is not
a process that can be made arbitrarily efficient. The
minimum energy required to perform a matrix
multiplication of a given size is bounded by physics.
Better chips reduce the energy per operation. The
number of operations required by the architecture
does not change.
GPT-3 has 175 billion parameters. GPT-4 has an
estimated 1.76 trillion. The models are getting larger
because larger models are more capable. The energy
cost scales with the parameters. The capability scales
with the parameters. The industry is locked into a
trade-off it cannot escape: more capability requires
more energy, and the world wants more capability.
Renewable energy procurement does not break this
trade-off. It changes the carbon intensity of the
energy consumed without changing the amount consumed.
THE NUMBERS BEHIND THE CRISIS
BLOOM 176B consumes 3.9 watt-hours per query,
measured directly by Luccioni et al. in 2022 on a
16-GPU cluster. GPT-4o consumes 0.34 watt-hours per
query by OpenAI's own disclosure. Llama 3.1 70B
consumes approximately 0.93 watt-hours per query.
At 1 billion queries per day — a conservative estimate
for a widely deployed AI system — GPT-4o consumes
340,000 kilowatt-hours daily. At the IEA's 2023 global
average grid intensity that is 136 tonnes of CO₂ per
day from inference alone. BLOOM at the same scale
produces 1,560 tonnes per day. These are not annual
figures. They are daily.
Training costs sit on top of this. GPT-3's training
run consumed 1,287 megawatt-hours. GPT-4's estimated
training cost is 16,200 megawatt-hours — the annual
electricity consumption of approximately 1,500 average
homes, spent once to produce one version of one model.
The industry knows these numbers. The response has
been to manage the narrative around them rather than
to solve the underlying cause.
WHAT EMPHOS IS BUILDING INSTEAD
Heinrich AI stores knowledge as frequency coordinates
in a layered signal field. Retrieving knowledge is
Goertzel correlation — a single-frequency signal
processing operation that runs in microseconds on
any CPU. No GPU. No matrix multiplication. No
dedicated AI silicon. No data center.
The measured energy per query is 0.00003 watt-hours
— approximately 11,000 times less than GPT-4o. This
measurement was taken on April 13, 2026 on a standard
Windows laptop with no optimization applied. The
methodology and sources are documented in EMPHOS
Group's Environmental and Resource Efficiency Report.
Since that measurement was taken the knowledge field
has grown from 128 concepts to 1.75 million. The CPU
usage is still 0.2%. The RAM is still 78 megabytes.
The energy per query has not changed. This is not
coincidence. It is the architecture.
Heinrich has no training run. Knowledge is added by
writing frequency coordinates to the field — a process
that costs fractions of a millisecond per concept.
The total training energy expenditure of Heinrich AI
to date is effectively zero in any meaningful
comparison to the systems it is being measured
against.
THE ENVIRONMENTAL GRANTS WE ARE PURSUING
EMPHOS Group is a small company building genuinely
novel technology in Chilliwack, British Columbia.
We are pursuing environmental innovation funding
through three programs.
Innovate BC supports British Columbia companies
developing technology with economic and environmental
impact. Heinrich's combination of novel architecture,
proven efficiency measurements, and clear product
roadmap positions EMPHOS as a strong candidate for
clean technology innovation funding.
The NRC Industrial Research Assistance Program
provides direct technical and financial support to
Canadian small and medium enterprises conducting
research and development. EMPHOS's R&D — the Heinrich
AI architecture, the ingestion pipeline, the HSR
pipeline, the HAVEN Ear hardware specification — is
exactly the kind of foundational technology
development IRAP was designed to support.
The federal Strategic Innovation Fund targets
transformative projects with significant environmental
and economic benefit at scale. A deployment of
Heinrich AI at the scale of a single major language
model deployment would save approximately 50,000
tonnes of CO₂ per year compared to the equivalent
LLM deployment.
PRIVACY AS AN ENVIRONMENTAL ARGUMENT
There is an environmental dimension to privacy that
rarely gets discussed.
Every voice assistant that sends audio to a server
generates a data center workload for every user
interaction. The energy cost is paid at the server,
not at the device. The user experience feels local.
The environmental cost is not.
Heinrich runs entirely on device. HAVEN Ear — the
personal intelligence device EMPHOS is building
around Heinrich — has no cloud dependency. Your voice
never leaves your ear unit. The energy cost is local
— and at 14 milliwatts for the full ear unit, it is
negligible.
Privacy by architecture is not just a user benefit.
It is an environmental position. A world where
personal AI runs locally at milliwatt power levels
is a fundamentally different world from one where
every personal AI interaction routes through a data
center. EMPHOS is building toward the first world.
WHAT THIS IS NOT
This is not a claim that Heinrich can replace every
AI application that exists today. Large language
models do things Heinrich does not yet do. Those
capabilities have value. The energy cost of those
capabilities is real and the industry should be
honest about it.
This is a claim that for the applications where
structured knowledge retrieval, honest uncertainty
reporting, and on-device inference matter — personal
intelligence, accessibility tools, real-time
translation, the hearing aid — Heinrich is not one
option among several. It is the only architecture
that delivers those capabilities at the power budget
required.
And it is a claim that the architectural alternative
exists, is proven, is measured, and is being built
right now — not as a research project, not as a
theoretical proposal, but as a production system
with 1.75 million knowledge nodes, 746 passing tests,
and a hardware roadmap targeting production in
Q4 2027.
WHAT COMES NEXT
The field continues to grow. The efficiency numbers
continue to hold. The hardware design is complete
at concept level. The patent disclosure is filed.
The investors are in conversation.
The AI industry's energy problem is not going to be
solved by the organizations most invested in the
current architecture. It is going to be solved by
building something genuinely different and proving
that it works.
That proof is running right now, on a laptop in
Chilliwack BC, at 0.2% CPU, growing at over a
million nodes per day.
Engineered for Presence.
——
EMPHOS Group · Chilliwack, BC, Canada
emphosgroup.com
Top comments (0)