Files
hips/HIPs/hip-0251-ai-compute-carbon-footprint.md
zeekay ad5a010f79 Three words for what a proposal is, and a generator that refuses
The same fact lived in nine hand-kept places and they disagreed about 58
proposals. This makes the file the only authority and everything else a
projection of it.

ONE VOCABULARY, in vocabulary.json, read by the lint, the generator and the
doc site. Three words, each answering "can I rely on this?" and nothing else:

    Draft    nothing implements it yet, or its own text says the work is ahead
    Final    the thing it specifies exists in the code
    Living   governance and process: amended in place, never finished

Eight spellings folded in. `Active` was the one doing real damage: on four Meta
and Process HIPs it meant living, on 48 others it meant shipped, and one word
doing two jobs is why nothing agreed. `Review` and `Last Call` described a
queue rather than a document, and a proposal under review is still one you
cannot build against.

STATUS IS EARNED, not declared. Every Standards Track proposal was checked
against the estate: 235 have their implementation, and eight do not and say so.
HIP-0095 carries a twelve-week plan ending at "mainnet launch" and has no code.
HIP-0525 specifies an all-machine meeting while apps/meet refuses a machine a
seat. HIP-0078 names a prover at a "planned location" that does not exist.
HIP-0114 is the cog.* procedure set, and that procedure set returns nothing.
HIP-0003, HIP-0102, HIP-0305 and HIP-0901 likewise: looked, found nothing.

SIXTEEN TOMBSTONES DELETED rather than statused. A dead document kept alive in
the index is one concept wearing two names. Each successor exists and carries
the text, and the 90 references that pointed at a tombstone now point at it --
whole markdown links at once, so link text cannot outlive its target. Three of
the sixteen were pure compatibility: HIP-0028 specified a migration off a
Bitnami chart, HIP-0029 dual-engine MySQL, HIP-0030 a schema-deprecation
policy. We ship no backwards compatibility, so we do not specify it.

Nine links were already broken and had been since the corpus was renumbered:
short-form ./hip-5.md paths that never existed. Two named the wrong proposal --
one cited "HIP-1: Hanzo Multimodal Models" where multimodal is HIP-0003.

THE GENERATOR REFUSES more often than it writes, because a short table looks
exactly as authoritative as a complete one. Rows must equal files; a shrinking
index must say by how much; a section whose input is missing is an error, not
an omission; nothing is written until all of it validates. Its self-test breaks
the corpus eleven ways and requires each to refuse AND leave README byte-
identical -- a guard that reports after half-writing has protected nothing.
The lint's second implementation of the same comparison is gone.

The site rendered 52 proposals with a fallback grey badge and counted them
toward none of its three stat tiles, because docs/lib/source.ts declared a
status union by hand: it admitted a `Stagnant` no file has ever carried and
omitted the one 52 files did. Five components each kept their own copy of the
colours. All six now read vocabulary.json.
2026-08-28 18:05:41 -07:00

8.1 KiB
Raw Permalink Blame History

hip, title, tags, description, author, discussions-to, status, type, created, requires
hip title tags description author discussions-to status type created requires
251 AI Compute Carbon Footprint
sustainability
carbon
compute
emissions
Methodology for measuring and reporting carbon emissions from AI training and inference. Hanzo AI Team (@hanzoai) https://github.com/hanzoai/hips/discussions Living Meta 2025-12-17 HIP-0200, HIP-0250

HIP-0251: AI Compute Carbon Footprint

Abstract

This HIP establishes the methodology for measuring, calculating, and reporting the carbon footprint of Hanzo AI's training and inference operations. It aligns with the GHG Protocol and provides AI-specific guidance for accurate carbon accounting.

Scope

Operational Boundary

Activity Scope Included
Model training Scope 2/3 ✅ Yes
Model inference Scope 2/3 ✅ Yes
Data storage Scope 2/3 ✅ Yes
Development compute Scope 2/3 ✅ Yes
Employee devices Scope 3 ✅ Yes
Cloud services Scope 3 ✅ Yes

Emissions Categorization

GHG Scope AI-Relevant Sources
Scope 1 On-site generation (if any)
Scope 2 Purchased electricity for owned compute
Scope 3, Cat 1 Cloud compute purchases
Scope 3, Cat 3 Fuel-related activities
Scope 3, Cat 11 Customer inference (if applicable)

Training Emissions

Training Calculation Methodology

Energy Consumption

Formula:

E_training = Σ(GPU_hours × TDP × PUE) / 1000

Where:

  • GPU_hours = GPU hours used
  • TDP = Thermal Design Power (kW)
  • PUE = Power Usage Effectiveness

Carbon Emissions

Formula:

CO2e_training = E_training × EF_grid × (1 - R%)

Where:

  • E_training = Energy consumption (kWh)
  • EF_grid = Grid emission factor (kgCO2e/kWh)
  • R% = Renewable energy percentage

Reference Values

GPU Power Consumption

GPU TDP (W) Typical Utilization Effective (W)
H100 SXM 700 80% 560
H100 PCIe 350 80% 280
A100 SXM 400 80% 320
A100 PCIe 300 80% 240

Data Center PUE

Provider/Region PUE Source
Hyperscaler average 1.1-1.2 Provider reports
Colocation average 1.3-1.5 Industry benchmarks
On-premise average 1.5-2.0 Industry benchmarks

Training Tracking

Required Metrics

For each training run:

Metric Collection Method
GPU type Cluster configuration
GPU hours Job scheduler logs
Data center location Cluster metadata
Time period Job timestamps

Aggregation

Period Aggregation
Per run Individual training job
Weekly Development activity
Monthly Reporting period
Annually Annual report

Inference Emissions

Inference Calculation Methodology

Per-Request Emissions

Formula:

CO2e_request = E_request × EF_region / 1000

Where:

E_request = (GPU_power × latency_seconds) + memory_energy + network_energy

Estimation Approach

For large-scale inference:

CO2e_inference = Total_GPU_hours × GPU_power × PUE × EF_avg / 1000

Efficiency Metrics

Metric Definition Target
CO2e/1K tokens Emissions per 1,000 tokens Track and reduce
CO2e/request Emissions per API request Track and reduce
CO2e/MAU Emissions per monthly active user Track and reduce

Geographic Distribution

Track inference by region:

Region % of Requests Grid Factor Weighted Factor
US-West X% 0.35 Calculated
US-East Y% 0.42 Calculated
EU Z% 0.28 Calculated
Asia W% 0.50 Calculated

Data & Emission Factors

Grid Emission Factors

Source Coverage Update Frequency
EPA eGRID US regions Annual
EEA EU countries Annual
IEA Global Annual
Provider-specific Cloud providers As published

Cloud Provider Data

Provider Data Available Source
AWS Region carbon intensity AWS Customer Carbon Footprint Tool
GCP Carbon-free energy % Google Cloud Carbon Footprint
Azure Emissions reporting Microsoft Sustainability Calculator

Data Quality

Level Definition Use
Primary Measured data Preferred
Secondary Provider-reported Acceptable
Tertiary Industry average Gap-filling

Reporting Standards

Internal Reporting

Monthly Dashboard

Metric Display
Total compute emissions CO2e (tonnes)
Training vs inference split % breakdown
YoY change % change
Efficiency trend CO2e/request over time

Quarterly Report

Section Contents
Summary Total emissions, trends
Training Major training runs, emissions
Inference Volume, efficiency
Initiatives Reduction progress

External Reporting

Annual Disclosure

Report Contents
ESG Report Summary metrics, targets
CDP Response Detailed methodology
Model Cards Per-model training emissions

Model Card Emissions

For each model release:

training_emissions:
  total_co2e_tonnes: X
  gpu_hours: Y
  energy_kwh: Z
  data_centers: [list]
  renewable_percentage: W%
  methodology: "HIP-251"

Reduction Strategies

Training Efficiency

Strategy Impact Implementation
Efficient architectures 10-50% reduction Architecture research
Mixed precision 30-50% speedup Training configuration
Gradient checkpointing Memory vs compute Based on model size
Curriculum learning 10-30% reduction Training methodology

Inference Efficiency

Strategy Impact Implementation
Model quantization 2-4x efficiency INT8/INT4 deployment
Speculative decoding 2-3x speedup Inference optimization
Batching Improved utilization Request aggregation
Caching Variable Response caching

Infrastructure

Strategy Impact Implementation
Green regions 30-90% reduction Region selection
Renewable PPAs Up to 100% reduction Energy procurement
Efficient hardware 20-50% per generation Hardware refresh
Cooling optimization PUE improvement Data center ops

Targets

Absolute Targets

Year Target Baseline
2025 Establish baseline Measure all emissions
2027 -30% vs baseline Reduction
2030 Net zero Reduction + offsets

Intensity Targets

Year CO2e/1K tokens CO2e/request
2025 Baseline Baseline
2027 -50% -50%
2030 -80% -80%

Renewable Energy

Year Renewable %
2025 50%
2027 80%
2030 100%

Verification

Internal Verification

Activity Frequency
Data validation Monthly
Calculation review Quarterly
Methodology audit Annual

External Verification

Activity Frequency Standard
Third-party audit Annual ISO 14064-3
CDP verification Annual CDP methodology
  • HIP-200: Responsible AI Principles
  • HIP-250: Sustainability Standards Alignment
  • HIP-260: Efficient Model Practices
  • HIP-270: AI Supply Chain Responsibility
  • HIP-290: Evidence Locker Index

Changelog

Version Date Changes
1.0 2025-12-17 Initial draft

Copyright and related rights waived via CC0.