Viro AI — Energy & Impact Methodology
This page explains how Viro estimates the energy use of AI queries, how those estimates translate into renewable energy matching, and how remaining profits support climate restoration projects. It also outlines our scope, assumptions, data sources, and limitations.
​
Our goal is to be transparent, improve literacy, and incorporate community feedback as the methodology evolves.
1. Scope & System Boundary
Viro currently measures operational energy from inference — the electricity used when generating responses to user prompts.
​
Not yet included (out of scope, for now):
• embodied emissions of hardware (manufacturing + end-of-life)
• data center construction
• training runs of foundation models
• long-distance networking infrastructure
• water use associated with cooling
These components matter. As credible and comparable research emerges, we intend to expand the boundary.
2. Token-Based Energy Estimation
Inference energy scales with token throughput (input + output), not raw request count. Viro estimates energy by mapping: tokens → compute → watt-hours (Wh)
​
Token counts are retrieved directly from model APIs. Longer prompts and outputs imply higher operational energy.
Example intuition:
-
"What’s the weather tomorrow?”
→ small model, short output, low token count -
“Explain quantum mechanics aboard the Millennium Falcon.”
→ larger model, long output, high token count
Routing small tasks to small models reduces token throughput and energy use.
3. Data Sources & Research References
We draw on multiple sources, including:
-
Google (Gemini inference research)
-
Microsoft (data center & AI efficiency publications)
-
Meta (LLM inference benchmarking)
-
OpenAI model documentation
-
Academic & community benchmarking studies
Corporate research can reflect organizational interests. We include independent studies where possible and treat all estimates as approximate rather than precise. Assumptions are conservative.
4. Renewable Energy Matching
Viro matches estimated inference energy with verified renewable energy generation, including:
-
Solar
-
Wind
-
Hydro
-
Battery storage
Scope notes:
-
matching is Wh-based
-
currently global-average
-
not yet region- or time-specific
-
no offsets or carbon credits
​
We treat renewable procurement as structural fossil displacement, not carbon accounting.
5. Climate Restoration Funding
All remaining profits support climate restoration projects. Our first partner is The Ocean Cleanup, focused on removing plastic from oceans and rivers.
We categorize impact into two buckets:
• Renewables → avoid future emissions
• Restoration → repair ecological systems
As Viro scales, we intend to diversify across:
• biodiversity
• reforestation
• coastal restoration
• water systems
• soil health
• carbon removal
6. Model Routing & Efficiency
Viro treats compute as a resource.
We route:
• small queries → small, efficient models
• complex queries → larger models
Most consumer AI tools route all requests to the largest model available. We view routing as a major lever for reducing inference energy without compromising capability.
7. Eco-friendly Terminology
We avoid describing AI as “eco-friendly.”
AI has a real footprint. Our goal is not to claim zero impact, but to redirect AI usage toward renewable energy generation and ecological restoration.
We use “climate-positive” to signal alignment and improvement, not elimination.
More questions?
We welcome critique and contribution from the broader community.
If you work in:
• AI energy research
• carbon accounting
• lifecycle assessment
• policy & regulation
• sustainability
• open-source modeling
• climate computation
We’d love to connect, contact nick@viro.app.