The Environmental Footprint of an Intelligence That Does Not Need a Data Center
EMPHOS Group, May 30, 2026, 7 min read
Microsoft is building nuclear reactors. Amazon has bought a power plant. Google's emissions are up 48 percent over four years. The frontier AI companies are not being secretive about what is coming. They are openly investing in dedicated power generation because their architectures require it. The question is not whether the current trajectory is environmentally sustainable. The question is whether it is architecturally necessary.
It is not.
This is what the comparison looks like when you run the numbers honestly.
One hundred million tasks a day
The relevant scale for this comparison is 100 million agentic tasks per day. That is roughly four percent of OpenAI's current daily ChatGPT query volume. Well into frontier product territory, but not unreasonable for any AI assistant that reaches mainstream adoption. At an average of 25 inference operations per agentic task, that is 2.5 billion underlying inference operations per day. This is the workload all four systems compared here must service.
The Heinrich figures are projections. Heinrich is not yet running at this scale. They are derived from the architecture's known properties. Coordinate lookups against a numpy float64 WaveField, CPU only execution, 5 ms per inference operation, 1.4 PUE, water use efficiency worse than what frontier providers actually run today. Every assumption is biased toward making Heinrich look more resource intensive than it likely will be. The figures for Claude, GPT, and Gemini come from the providers' own disclosures and from peer reviewed studies.
Under promise, over deliver. The numbers below are the conservative case.
Power
Heinrich at 100 million tasks per day draws 6 to 10 megawatts of continuous electricity. That is the scale of a moderate manufacturing facility. The daily electricity consumption is 144 to 240 megawatt hours.
Claude at the same task volume draws 375 to 625 megawatts. GPT draws 450 to 750 megawatts. Gemini, the most efficient of the three, draws 300 to 500 megawatts. Their daily consumption runs from seven thousand to eighteen thousand megawatt hours.
The midpoint figures mean Heinrich uses roughly 30 to 75 times less electricity than competing agentic systems at the same task volume. Even against the most efficient frontier system, Heinrich is 30 to 50 times more energy efficient.
This is not optimisation. It is what happens when each inference is a coordinate lookup against a half gigabyte numpy array instead of a forward pass through 400 billion parameters streamed through high bandwidth memory. The physical work per task is categorically different, and the power bill reflects that.
Water
Data centers do not run cool on air alone. At scale, they consume water. For evaporative cooling, for chilled water systems, for keeping the silicon below thermal limits. The figure is rarely on the marketing page.
Heinrich at 100 million tasks per day consumes 50,000 to 100,000 litres of cooling water. Roughly 20 to 50 Olympic swimming pools per year.
Claude at the same task volume consumes 2 to 5 million litres per day. GPT consumes 2.5 to 50 million litres per day, depending on which generation of the model is doing the work. Gemini, with Google's industry leading cooling efficiency, consumes about 650,000 litres per day. Six to thirteen times more than Heinrich on its worst assumption.
GPT at 100 million agentic tasks per day, on its higher end, would consume up to 7,300 Olympic swimming pools of cooling water per year. Heinrich, on its conservative high end, consumes around 50. Both numbers are real water leaving real watersheds.
Physical footprint
The space the hardware actually occupies follows the same pattern. Heinrich at 100M tasks per day requires roughly 100 racks. A single data center floor. About 600 to 800 square feet of compute.
Claude requires 3,000 to 4,500 racks. GPT requires 3,750 to 5,500. Gemini requires 2,500 to 3,750. Each frontier deployment occupies an entire purpose built data center. 20,000 to 45,000 square feet of compute floor, before cooling infrastructure, power distribution, and aisle clearance are accounted for. The capital cost of constructing those facilities runs into the hundreds of millions of dollars before the first server is racked.
Heinrich fits on a floor that already exists.
Why the gap is structural, not engineering
It is worth being clear about why these numbers look the way they do. The advantage is not that Heinrich is better tuned or that EMPHOS engineers are more careful. It is that Heinrich is doing a different category of physical work for every inference.
A frontier agentic LLM stores knowledge as hundreds of billions of floating point weights distributed across a neural network. Every step of every agentic task, and an agentic task is many steps, executes a forward pass through every layer of that network. That requires streaming hundreds of billions of parameters through compute units per generated token, which requires high bandwidth memory at multiple terabytes per second, which requires HBM3 chips, which require GPUs or TPUs, which require warehouse scale infrastructure and the cooling and power that goes with it. The hardware profile follows from the architecture. There is no version of an LLM that does not need this.
Heinrich stores knowledge as harmonics in a WaveField at 44.1 kHz. The WaveField is the knowledge, not a representation of it. Each inference resolves a coordinate against that field. Standard DDR5 memory at 400 to 500 GB per second per socket is more than sufficient. A dual socket Xeon server with 128 threads handles approximately two million inference operations per day. A standard data center floor handles 100 million agentic tasks per day. No GPU. No TPU. No HBM3. No accelerator factory.
Frontier providers are getting more efficient. Gemini went from 8 watt hours per query to 0.24 watt hours over twelve months. Heinrich's advantage does not close as that efficiency improves. An optimised forward pass through 400 billion parameters is still doing fundamentally more work than a coordinate lookup. The architectural gap is the gap. It does not narrow because the inefficient architecture got better at being inefficient.
Edge deployment
There is one dimension where the comparison is not close. It is absent. Heinrich runs on the edge. The HAVEN Ear ships under 50 megabytes and executes agentic tasks entirely on device. No data leaves the user. No round trip to a data center per step. No per task cloud cost.
No frontier agentic system can do this. Claude is cloud only. GPT is cloud only. Gemini Nano, the most aggressive edge optimised frontier model, ships at roughly 1.8 gigabytes and runs only on flagship Android devices. And is not the same model that handles serious agentic work.
The implication is that as agentic AI moves into wearables, hearing aids, embedded sensors, and the long tail of devices that do not have a reliable data connection, the dominant frontier architecture cannot follow. The numpy WaveField can. The forward pass through 400 billion parameters cannot be miniaturised. It can only be moved farther from the user.
What this means
The environmental conversation about AI has, until now, taken one assumption for granted. That powerful AI requires industrial scale infrastructure. The carbon footprint, the water footprint, the rare earth footprint, the land use footprint of the data centers. Those have been treated as the unavoidable cost of capability.
Heinrich is the test case for whether that assumption holds. The architecture says it does not. The benchmark says it does not. The simulated footprint at 100 million agentic tasks per day says it does not. The physical product shipping at under 50 megabytes says it does not.
The right question is no longer how much faster we can build data centers. It is whether the architecture that demands them is the only architecture available.
It is not. Heinrich is here. The wager is that this is not a margin advantage. It is a different category of system, with a different relationship to the physical world, and the consequences for energy, water, land, and access reach far beyond a single product.
Engineered for Presence.
Stay in the loop
EMPHOS publishes twice a week. Product updates, research, and the thinking behind the build.
Explore Haven, HEINRICH Intelligence, The EMPHOS Vision, All Posts
EMPHOS Group, Chilliwack, BC, Canada, info@emphosgroup.com