The Environmental Cost of Artificial Intelligence
A comprehensive analysis of energy consumption, carbon emissions, and water usage across 30 AI models — from lightweight text classifiers to frontier reasoning systems.
Impact Overview
Energy consumption varies by orders of magnitude across AI task categories. Video generation dominates resource usage, while text classification remains remarkably efficient.
Top Energy Consumers
Max Wh per prompt (log scale)
Model Distribution
Energy flows across AI model categories
Data Explorer
Sort, filter, and search across all 30 AI models. Energy values are per-prompt. Click column headers to sort.
Model Comparison
Select up to 4 models to compare side-by-side
No models selected yet
Use the selector above or try a quick preset to get started
Key Findings
Critical insights from analyzing energy, carbon, and water data across 30 AI models and 10+ task categories.
1,888,880x Energy Gap
The most energy-intensive AI task (CogVideoX video generation at 944 Wh) consumes nearly 1.9 million times more energy per prompt than the most efficient (LTX-Video at 0.0005 Wh).
Text Models: Wide Spectrum
Text generation spans from 0.016 Wh (Llama 3.1 8B) to 39.2 Wh (GPT-o3) — a 2,450x range within the same task category, driven by model size and reasoning depth.
MoE Architecture Wins
Mixtral 8x22B (141B params, MoE) uses only 0.07 Wh — dramatically less than dense models of similar capability, demonstrating mixture-of-experts as a key efficiency strategy.
Small Models, Big Impact
Task-specific fine-tuned models (0.06 Wh) and Llama 3.1 8B (0.016 Wh) prove that smaller, specialized models can deliver high accuracy at a fraction of the environmental cost.
Hidden Water Cost
Video generation consumes up to 1,000 mL of water per prompt for cooling — equivalent to a full water bottle. Even text models like Mistral Large 2 use 45 mL per prompt.
Reasoning Tax
Chain-of-thought and agentic reasoning (GPT-o3 at 39.2 Wh, Agentic AI at 4.32 Wh) impose a significant energy premium — the cost of "thinking harder" is measurable.
Carbon Emissions
Video generation produces up to 380 gCO2e per prompt — equivalent to driving 1.5 km in a car.
Water Consumption
Data center cooling for a single video generation prompt can consume up to 1 liter of water.
Measurement & Verification
Building the foundation for trusted AI environmental data
Data Sourcing & Model Assumptions
We believe in radical transparency. Every number on this dashboard has a source, every calculation has an assumption, and every assumption has a known limitation. This section documents our methodology so you can evaluate, challenge, and improve it.
Read the Full Mālama AICo2 Methodology (PDF)The Measurement Problem
There is no industry-standard methodology for measuring AI's environmental footprint. Companies, in the words of the Federation of American Scientists, "report whatever they choose, however they choose." The true carbon footprint of major AI providers may be up to 662% higher than publicly reported figures.
Current metrics like Power Usage Effectiveness (PUE) — a 20-year-old standard — measure only facility efficiency, not how efficiently IT equipment actually uses delivered power. As the FAS describes it: "Like a car that reports how much fuel reaches the engine but not the miles per gallon of that engine."
This dashboard aggregates the best available data from multiple independent sources, but we acknowledge that all estimates carry significant uncertainty. Our goal is not to present definitive numbers, but to make the scale of AI's environmental impact visible and to push for better measurement infrastructure.
Our full estimation framework — including the dual-phase architecture, five-stage pipeline, task-class parameterization, and cryptographic telemetry roadmap — is documented in the Mālama AICo2 Methodology.
Data Sources
Model Assumptions
Click each assumption to see the evidence basis and known limitations. Confidence levels indicate our assessment of reliability.
How We Calculate
Energy (Wh/prompt)
Step 1: Measure GPU power draw during inference using hardware power sensors (NVML/RAPL) or software tools (Code Carbon, ML.Energy).
Step 2: Double the GPU measurement to account for CPU, RAM, networking, storage, and cooling overhead (~50% rule).
Step 3: Apply PUE multiplier (1.1-1.2) for data center infrastructure overhead.
E_total = E_gpu × 2.0 × PUECarbon (gCO2e/prompt)
Step 1: Convert energy to kWh.
Step 2: Multiply by location-based grid carbon intensity (gCO2e/kWh) from EPA eGRID or Electricity Maps.
Caveat: Market-based accounting with RECs can show near-zero emissions even when actual grid mix is carbon-heavy.
CO2 = E_total × GridIntensityWater (mL/prompt)
Step 1: Estimate cooling energy from total energy and PUE breakdown.
Step 2: Apply Water Usage Effectiveness (WUE) ratio — liters of water per kWh of IT energy.
Caveat: Air-cooled facilities use minimal water; evaporative cooling facilities use significantly more. Few providers disclose WUE.
H2O = E_cooling × WUEWe Welcome Better Methods
This dashboard is a living document. We recognize that our methodology has significant gaps — particularly around water consumption, agentic AI workloads, and the variance between estimated and actual data center operations.
If you have access to better data, more rigorous measurement approaches, or can identify errors in our assumptions, we want to hear from you. The goal is not to be right — it is to be progressively less wrong and to build toward a future where AI's environmental impact is measured with trusted, objective, real-time data rather than estimates and extrapolations.
This is exactly why we advocate for hardware-level sensor integration inside data centers — to replace estimation with measurement. See the Agents & Sensors section below for our vision.
Submit a ContributionThe Compounding Cost of Autonomous AI
As AI evolves from single-query chatbots to autonomous agents that plan, execute, and iterate, the energy footprint doesn't just grow — it compounds. Traditional "energy per query" metrics fail to capture the true cost of agentic workflows.
From Episodic to Continuous
A single agentic workflow can involve multiple model calls, data retrieval, validation loops, and downstream integrations. Unlike a simple chatbot query, agents create persistent, background compute demand.
As Ampere Computing's research (April 2026) warns: infrastructure demand from agentic AI grows in non-linear ways. Automated decisions generate follow-up processes, and workflows branch into additional tasks. This multiplicative effect is easily underestimated.
The FAS policy memo acknowledges this shift directly: "When we move from chatbots to agentic AI systems that plan, act, remember, and iterate autonomously, traditional 'energy per query' metrics no longer capture the full picture."
Energy Escalation Ladder
Autonomous agent loops can invoke 10-50+ model calls per task. At enterprise scale with thousands of concurrent agents, energy costs compound exponentially. Current measurement infrastructure cannot track this.
From Estimation to Observation
Every number on this dashboard is an estimate. The real goal is to replace estimation with measurement — deploying hardware-level sensors inside data centers to create an environment of trusted, objective, verifiable data that can dynamically and drastically reduce AI's environmental impact.
Hardware-Level Sensor Integration
IoT sensors deployed at rack, server, and cooling system level — measuring actual environmental impact, not estimates.
Today: Estimation & Opacity
- ×Software-based GPU power measurement (CodeCarbon, NVML)
- ×Rule-of-thumb extrapolation: GPU = ~50% of total energy
- ×Average PUE values applied uniformly across facilities
- ×Regional grid carbon intensity averages (not real-time)
- ×Water consumption rarely disclosed, often unknown
- ×Self-reported data with no independent verification
- ×True emissions may be 662% higher than reported (FAS)
Future: Measurement & Trust
- ✓Hardware-level sensors at rack, server, and GPU level
- ✓Actual energy measurement per workload, per inference
- ✓Real-time PUE calculated per facility, per hour
- ✓Live grid carbon intensity from Electricity Maps / sensors
- ✓Flow sensors on cooling systems measuring actual water use
- ✓Blockchain-verified, tamper-proof measurement records
- ✓AI-powered dynamic optimization reducing impact in real-time
Sensor Capabilities for Data Centers
Real-Time Energy Metering
Hardware-level power sensors at the rack, server, and GPU level — measuring actual energy consumption per workload, not estimates derived from GPU draw alone.
Carbon Emissions Tracking
Live grid carbon intensity integration combined with actual energy measurements to calculate real-time, location-based carbon emissions — not market-based credits.
Water Consumption Monitoring
Flow sensors on cooling systems measuring actual water consumption per rack, correlated with compute workloads to attribute water usage to specific AI tasks.
Thermal & Airflow Analysis
Temperature, humidity, and airflow sensors creating a real-time thermal map of the data center — identifying cooling inefficiencies and hot spots for optimization.
The Measurement-to-Optimization Pipeline
Deploy Sensors
IoT sensors installed at rack, server, and cooling system level across the data center.
Measure Everything
Real-time energy, water, thermal, and airflow data streamed continuously — not sampled or estimated.
Verify On-Chain
Measurement data anchored to blockchain (Cardano) for tamper-proof, immutable audit trails — creating trusted, objective records.
AI-Powered Analysis
Machine learning models analyze sensor data to identify inefficiencies, predict failures, and recommend optimizations.
Dynamic Optimization
Automated workload scheduling, cooling adjustments, and resource allocation based on real-time environmental data.
Verified Reporting
Standardized, auditable environmental reports generated from actual measurements — replacing self-reported estimates.
Malama dMRV: Trusted Environmental Data
Digital Measurement, Reporting & Verification
The Malama dMRV (digital Measurement, Reporting, and Verification) infrastructure — already proven in carbon credit verification for agriculture and land management — provides the technological foundation for trusted data center monitoring.
By combining IoT sensors for real-time environmental data collection, AI-powered analysis for pattern recognition and optimization, and blockchain verification (Cardano) for tamper-proof audit trails, the platform creates an environment where environmental claims are backed by verifiable, objective measurements — not self-reported estimates.
Extending this infrastructure into data centers means every energy reading, every water flow measurement, and every carbon calculation is recorded immutably on-chain. Operators, regulators, and the public can independently verify environmental impact claims, creating the accountability that the FAS, CMU, and industry researchers have called for.
Read the Full Mālama AICo2 MethodologyThe Impact: What Trusted Data Enables
Dynamic Workload Scheduling
Route AI inference to facilities with the lowest real-time carbon intensity, shifting compute to when and where clean energy is available.
Cooling Optimization
Real-time thermal data enables predictive cooling adjustments — reducing water consumption by matching cooling output to actual heat load, not worst-case estimates.
Hardware Utilization
Identify idle capacity and underutilized servers that consume energy without productive output. Google found idle machines are a significant hidden cost.
Regulatory Compliance
Automated, verifiable reporting that meets emerging EU AI Act requirements and anticipated US mandatory disclosure frameworks (FAS/CMU recommendations).
Carbon-Aware Computing
Schedule training runs and batch inference during periods of high renewable energy availability, verified by real-time grid data rather than annual averages.
Accountability & Trust
Replace the current 662% reporting gap with blockchain-verified measurements that investors, regulators, and the public can independently audit.
Contribute to Better Data
This dashboard is a living document. If you have access to better data, more rigorous measurement approaches, or can identify errors in our assumptions — we want to hear from you. The goal is to be progressively less wrong.
What are you contributing?
Submission Guidelines
- Include DOI or URL for any referenced papers
- Specify hardware and inference conditions if submitting energy data
- Note confidence level and known limitations
- Peer-reviewed sources are prioritized but not required
What Happens Next
Submit
Send your data, correction, or methodology proposal via the form above.
Review
Our research team evaluates submissions against existing sources and methodology.
Integrate
Accepted contributions are incorporated into the dashboard with full attribution.
Publish
Updated data and methodology changes are reflected in the next version release.