x402GAL · water estimate
x402GAL Water Intensity Methodology
2026-09 edition · sha256:7f27acc35d4e67bd50b60e894c30c51932d2318c6bc20ca8f38413d03122b6f0
How x402GAL estimates the water footprint of AI usage, with every coefficient published.
This is a modeled estimate from published coefficients, not a measurement.
The estimate covers water used at the data center only, and does not include water used to generate the electricity, which can be several times larger.
What a public tier uses
The text-only estimate maps each public tier onto a published energy row and a published facility water rate. It does not add a new coefficient.
| Public tier | Energy row | Facility water row |
|---|---|---|
| frontier | gpt4_class | hyperscaler |
| local | small_open | unknown |
| own_agents | unknown | unknown |
The own_agents tier uses unknown rows on both axes, so its band is the widest, and it will narrow as published coefficients for agentic workloads appear.
frontier uses gpt4_class and the hyperscaler facility water rate (0.2 L/kWh). local uses small_open and the unknown facility water rate (0.45 L/kWh). own_agents uses the unknown energy row and the unknown facility water rate.
Usage rows
Each usage answer picks a token row. Where you are based does not change the estimate. It only sets which city the map opens on.
| Text use per day | Tokens in | Tokens out |
|---|---|---|
| A few prompts | 2,000 | 4,000 |
| Steadily through the day | 10,000 | 20,000 |
| It runs my workflow | 40,000 | 80,000 |
| I run agents or automations | 200,000 | 400,000 |
| Images or video | Items per day | IT energy per item |
|---|---|---|
| Rarely | 0.5 | 0.003 kWh |
| Sometimes | 5 | 0.003 kWh |
| Regularly | 25 | 0.003 kWh |
The text and media estimate adds image and video energy from the table above. The text-only estimate does not. The two results can differ.
A frontier model on a hyperscaler uses 0.6x to 1.6x. Every other published pairing uses 0.3x to 3x.
The text and media estimate rounds the center to 1 significant figure. The text-only estimate shows the range as well as that center.
Estimates from your own token counts
The same formula and the same coefficients apply when an assistant passes token counts from a usage record. The result is only as accurate as the counts supplied.
The tier describes the model and where it runs, not whether the user runs agents. Hosted models from OpenAI (GPT), Anthropic (Claude), Google (Gemini), xAI (Grok), Meta (Muse Spark), Microsoft (MAI) and Amazon (Nova) are frontier, including through coding agents like Claude Code or Cursor. Open-weight models (Llama, Gemma, Phi, Nemotron, DeepSeek, Qwen, Kimi, Mistral) are frontier when used through a hosted API and local when run on the user's own hardware. own_agents is only for an unknown model on unknown infrastructure.
Count cache creation tokens, also called cache writes, in the input tokens. Cached reads are excluded because there is no published coefficient for them.
Include reasoning or thinking tokens in the output tokens. Providers bill them as output, and they can be a large share of agent usage.
Adding the low end of each entry and adding the high end is a conservative combined range. It is wider than a statistical range.
The model
Based on The Green Grid's WUE metric.
W_site = WUE × [(T_in/1000)*e_in + (T_out/1000)*e_out + e_overhead] × F_boundary
WUE is water-use effectiveness in liters per kWh of IT energy. e_in and e_out are kWh per thousand input and output tokens. e_overhead is fixed energy, used by the text and media estimate for images and video. F_boundary is 1.0 when energy is already at the IT boundary.
| Model tier | e_in (kWh / 1K tokens) | e_out (kWh / 1K tokens) |
|---|---|---|
| gpt4_class | 0.0002 | 0.0006 |
| small_open | 0.0001 | 0.0003 |
| unknown | 0.0003 | 0.001 |
| Facility water row | WUE (L / kWh) |
|---|---|
| hyperscaler | 0.2 |
| unknown | 0.45 |
| Mode | Multiplier | Meaning |
|---|---|---|
| facility | 1 | Water used at the data center. This is what the text-only estimate and the text and media estimate return. |
| source | 4 | Adds water used to generate the electricity. Not included in the text-only estimate or the text and media estimate. |
| lifecycle | 8 | Adds hardware manufacturing. Informational only. Not included in either estimate. |
The hashed 2026-09 data calls this mode 'site.' The next edition will use 'facility.'
Where the defaults come from
Each published coefficient is listed with the source that states that number. A row with no direct source is an x402GAL assumption.
| Coefficient | Published value | Source |
|---|---|---|
| gpt4_class e_out | 0.0006 kWh / 1K tokens | Epoch AI GPT-4o estimate, about 0.0006 kWh per 1K output tokens. |
| gpt4_class e_in | 0.0002 kWh / 1K tokens | x402GAL assumption. |
| small_open e_in | 0.0001 kWh / 1K tokens | x402GAL assumption. |
| small_open e_out | 0.0003 kWh / 1K tokens | x402GAL assumption. |
| unknown e_in | 0.0003 kWh / 1K tokens | x402GAL assumption. |
| unknown e_out | 0.001 kWh / 1K tokens | x402GAL assumption. |
| hyperscaler facility water rate | 0.2 L/kWh | x402GAL assumption, set within the 0.15 to 0.27 L/kWh range reported for 2024 and FY25: AWS 0.15 (2024 Sustainability Report), Meta 0.19 (2024, per the 2025 Environmental Data Index), Microsoft 0.27 (FY25). |
| unknown facility water rate | 0.45 L/kWh | x402GAL assumption, set above LBNL's hyperscale median of 0.32 L/kWh, because a site of unknown type is assumed to be less efficient than a hyperscaler. |
| facility mode multiplier | 1 | Defines the facility boundary. The hashed 2026-09 data calls this mode 'site.' The next edition will use 'facility.' |
| source mode multiplier | 4 | x402GAL assumption. No published basis yet. |
| lifecycle mode multiplier | 8 | x402GAL assumption. No published basis yet. |
| frontier on a hyperscaler band | 0.6x to 1.6x | x402GAL assumption. |
| every other published pairing band | 0.3x to 3x | x402GAL assumption. |
| text use token rows | Usage rows table | x402GAL assumption. |
| images or video rows | 0.003 kWh per item | x402GAL assumption. |
| F_boundary | 1.0 when energy is already IT energy | x402GAL assumption. Green Grid WUE is defined on IT energy. |
The hashed reference list also includes the LBNL 2024 United States Data Center Energy Usage Report. The model notes also name TokenPowerBench and IEEE Access Pythia measurements. They are not tied to a row above.
Checking the hash
The hash is SHA-256 of the exact string below: the specification line, the formula, the references, and the three default tables, in that key order. Change any coefficient and the hash changes.
{"spec":"Green Grid WUE v1, boundary-aware split","formula":"W_site = WUE × [(T_in/1000)*e_in + (T_out/1000)*e_out + e_overhead] × F_boundary","refs":["https://www.thegreengrid.org/system/files/store/WUE_v1.pdf","https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf","https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use"],"MODEL_DEFAULTS":{"gpt4_class":{"e_in":0.0002,"e_out":0.0006},"small_open":{"e_in":0.0001,"e_out":0.0003},"unknown":{"e_in":0.0003,"e_out":0.001}},"WUE_DEFAULTS":{"hyperscaler":0.2,"unknown":0.45},"MODE_MULT":{"site":1,"source":4,"lifecycle":8}}Expected hash: 7f27acc35d4e67bd50b60e894c30c51932d2318c6bc20ca8f38413d03122b6f0
What this is not
It is not a single liters-per-token figure. It does not reduce, replace, or balance any water use. A pledge link is an intention. Verified replenishment is not yet live.
Estimates only. Nothing here is a measurement, an offer, or a water right.