Smaller AI
Request→AI model→Result
A 3Stone architecture principle
Eco AI
Why Eco AI?
AI has a
physical footprint.
Artificial intelligence runs on physical infrastructure. That infrastructure needs electricity and, depending on how it is cooled, water.
Using AI intentionally means avoiding unnecessary demand—not giving up the capabilities that make AI useful.
If you’re going to use AI,
use Eco AI.
The idea
AI-Optional Architecture plus no proprietary foundation-model data centers equals Eco AI
AI-Optional
Architecture↓ PLUS ↓No Proprietary
Foundation-Model
Data Centers↓ EQUALS ↓ECO AI
Architecture↓ PLUS ↓No Proprietary
Foundation-Model
Data Centers↓ EQUALS ↓ECO AI
What does that mean?
Less unnecessary
AI processing.
Less unnecessary demand on AI infrastructure.
3Stone does not eliminate AI. It is designed to avoid requiring AI where conventional computing can complete the step.
AI where it matters.
Not everywhere.
A useful distinction
A smaller AI is still AI.
Eco AI can make a different choice: whether the step needs a model at all.
Eco AI
Request→
Does this step need AI?
→ResultYES Use AINO Use conventional computing
Eco AI isn’t about weaker AI.
It’s about unnecessary AI.
Two ways to build
Intelligence, applied deliberately.
AI-first architecture
Your requestAIAIAIYour result
A model is invoked throughout the path.
3Stone Eco AI
Your requestDoes AI add value here?Code when code will doAI when it mattersYour result
Eligible steps use conventional computing; AI remains available for work that needs it.
This is an architectural concept, not a claim about every private implementation used by another provider.
Evidence, not estimates
What providers disclose.
Physical infrastructure and product architecture are different questions. We report them separately and preserve the scope of each source.
OpenAI
OpenAI and its partners are building and procuring large-scale compute infrastructure through Stargate and related programs.
Google trains and runs Gemini across Google’s AI-optimized infrastructure, including its own TPU systems and data centers.
Anthropic
Anthropic says Claude is trained and run across partner infrastructure including AWS Trainium, Google TPUs and NVIDIA GPUs.
3Stone
3Stone does not operate proprietary foundation-model data centers. It operates the product layer that routes and combines external specialist capabilities.
View the evidence
Physical infrastructure disclosures
| Provider | Infrastructure disclosure | Published prompt energy | Published prompt water |
|---|---|---|---|
| OpenAI / ChatGPT | OpenAI and partners operate and procure model infrastructure. | About 0.34 Wh per average ChatGPT query. | About 0.000085 gal / 0.32 mL per average query. |
| Google / Gemini | Google operates model and data-center infrastructure. | 0.24 Wh per median Gemini Apps text prompt. | 0.26 mL per median Gemini Apps text prompt. |
| Anthropic / Claude | Claude runs across AWS, Google and NVIDIA infrastructure. | Not publicly disclosed per prompt. | Not publicly disclosed per prompt. |
| 3Stone | No proprietary foundation-model data centers in the current architecture. | 3Stone’s product layer is not directly comparable to a foundation-model provider’s per-prompt measurement. Underlying model execution still has an attributable footprint. | |
Publicly documented architecture
| Architecture behavior | OpenAI | Anthropic | 3Stone | |
|---|---|---|---|---|
| AI-optional processing | Not publicly established for the consumer product | Not publicly established for the consumer product | Not publicly established for the consumer product | Demonstrated |
| Deterministic paths | Documented in limited developer contexts | Documented in limited agent contexts | Documented in limited agent contexts | Demonstrated |
| Zero controller-model calls | Not assessed; architecture is proprietary | Not assessed; architecture is proprietary | Not assessed; architecture is proprietary | Demonstrated on eligible paths |
| Controlled escalation | Documented in limited developer contexts | Documented in limited agent contexts | Documented in limited agent contexts | Demonstrated |
| Duplicate-generation protection | Not publicly established | Not publicly established | Not publicly established | Demonstrated for supported media paths |
OpenAI
Environmental figures: Sam Altman, “The Gentle Singularity,” June 10, 2025. No methodology or measurement period accompanied the figures.
Environmental statement ↗Infrastructure disclosure ↗May 2025 median Gemini Apps text-prompt data. Google’s estimate includes accelerators, host systems, provisioned idle capacity and data-center overhead; it is point-in-time and not independently verified.
Methodology ↗Infrastructure disclosure ↗Anthropic
Anthropic describes its compute partners, but we found no comparable first-party per-prompt energy or water measurement.
Infrastructure disclosure ↗Sources last verified September 29, 2026. Provider figures use different scopes and methodologies and should not be treated as directly comparable.
The 3Stone layer
Powerful AI.
A more intentional architecture.
3Stone connects specialized AI capabilities with its own routing, orchestration, entitlement, execution, persistence, safety and product systems.
A better default
If you’re going to use AI,
use Eco AI.
Get the capabilities you need from a product designed to avoid unnecessary AI processing.
Try 3StonePlain answers
Eco AI, explained.
What is Eco AI?
Eco AI uses AI where it is needed, not where it is not. 3Stone combines AI-optional processing, bounded routing and capability-specific execution with conventional computing for eligible steps.
Why does AI use electricity and water?
AI inference runs on physical computing infrastructure that uses electricity. Depending on the data center and cooling system, operating that infrastructure can also consume water.
Is a smaller AI model the same as Eco AI?
No. A smaller model still performs model inference. Eco AI can avoid another model call entirely for eligible processing steps and use conventional software instead.
Does 3Stone operate foundation-model data centers?
3Stone does not operate proprietary foundation-model training or inference data centers. 3Stone operates the product layer that routes and combines external specialist capabilities. Underlying model execution still has an attributable physical footprint.