A sapiens runs on a single device. No data center, no training run, nothing leaving the building.
Talk to our infrastructure team
Google is the only frontier lab that has published what a single prompt costs to answer. Drag the volume and watch its own number climb. One of these columns cannot.
Priced at Google’s own published cost of 0.24 Wh, 0.26 mL and 0.03 g per prompt.
Runs on the device already in the room. No cost per query, at any volume.
—
Per-prompt cost: Google, “Measuring the environmental impact of AI inference,” 21 August 2025 — median Gemini Apps text prompt. Volume ceiling: OpenAI told Axios in July 2025 that ChatGPT answers 2.5 billion prompts a day, and has published nothing since. Neither OpenAI nor Anthropic discloses a per-prompt figure of its own, so Google’s is the only one available to apply. Household equivalence at 29 kWh per day, the US residential average per EIA. A sapiens draws only the power of the host device, which runs whether it is asked a question or not.
None of the figures below are ours. Each one is published by the company or the agency named beside it.
Its carbon footprint rose 18 per cent in 2025 alone, the largest annual increase it has ever reported, while electricity use climbed 37 per cent. Google is the most efficient hyperscale operator in the world and it is moving backwards.
Google 2026 Environmental Report21,121,000 tonnes of CO₂ equivalent in fiscal 2025, against 13,061,000 in fiscal 2020, the year of the pledge. Scope 2 emissions, the power the datacenters actually draw, grew almost ten-fold to 13 per cent of the total.
Microsoft 2026 Environmental Data Fact SheetThree Mile Island Unit 1 is under NRC review for a 2027 restart as the Crane Clean Energy Center, its full output contracted to Microsoft under a twenty-year agreement. This is no longer a forecast about future demand. It is a decommissioned reactor being reversed.
Constellation Energy; US Nuclear Regulatory CommissionUp to five million gallons a day for cooling, roughly 1.8 billion a year. Across the United States, data centers were already consuming 449 million gallons a day on 2021 figures, before the current buildout began.
Environmental and Energy Study InstituteEvery number on this page belongs to somebody else. We publish no lifecycle assessment and no carbon accounting of our own, because a sapiens has never had the footprint that would require one.
PJM is the largest grid operator in the United States, serving thirteen states. Its capacity prices are what generators are paid simply to be available, and they flow through to every customer in the region.
The intermediate year cleared at $269.92. Capacity prices are what generators are paid simply to be available, and they flow through to every customer in the footprint.
PJM capacity auction results
Roughly $9.3 billion recovered from customers across the PJM footprint in a single delivery year, according to analysis of the auction results.
Institute for Energy Economics and Financial Analysis
Roughly $18 a month in western Maryland and $16 in Ohio. Residential rates across affected territories are projected to rise 15 to 40 per cent by 2030.
Utility rate filings; ICF projectionsThree decisions, none of them made for energy reasons, all with the same consequence.
A sapiens holds knowledge as a structured model of interconnected concepts, encoded directly. There is no corpus to ingest, no gradient descent, no cluster running for weeks.
A sapiens runs on a mobile device processor. No round trip to a data center, no GPU cluster serving inference, no network transit per query.
Identical inputs produce identical outputs. No sampling, no repeated generation to find an acceptable answer, no verification pass to catch what the model invented.
Grid and reactor operations are two of the mission-critical sectors this platform serves. The operators who most need automation are being sold systems that require the very capacity they are struggling to balance.
| Characteristic | Mainstream AI | Sapiens |
|---|---|---|
| Deployment target | Data center cluster | Single edge device |
| Training corpus required | Web-scale | None — curated world model |
| Network dependency | Continuous connectivity | None — air-gap capable |
| Cooling water footprint | Attributable share of facility draw | None beyond the host device |
| Cost profile | Metered, vendor-set | Fixed at deployment |
| Retraining cycle | Periodic, full-cluster | Model is edited, not retrained |
The same architecture that removes the data center is the one that makes a system deployable in a control room. Efficiency is not a separate product line. It is what the rest of the platform already is.
Every CISA critical infrastructure sector a sapiens can serve, from grid and reactor management to space operations and defense.
Explore sectorsEdge hardware, embedded systems, robotics and command and control. No connectivity required and nothing phoning home.
Explore deploymentSynthetic Intelligence: cognitive world models built by engineers, deterministic by construction, with no corpus and no training run.
Explore the technology