estimator
How much AI use equals one steak?
Oneis about
- By CO₂e
- 240 years
- By water
- 150 years
of AI use at this example usage.
Example: ChatGPT Free, 5 chats a day. Change it
Scenario range
By CO₂e: 42 years to 1,100 years. By water: 16 years to 1,400 years.
The low and high ends use the most and least efficient serving assumptions at the same usage. This is a range of scenarios, not a statistical confidence interval.
Assumption, about the scenario range
Scenario range. The low and high results vary utilization, serving allowance, cache misses and water factors using selected inputs. This is not a statistical confidence interval, and actual use can fall outside it.
Sources: Google 2025; Epoch AI 2025; University of Michigan / arXiv 2025; Lawrence Berkeley National Laboratory 2024
Order-of-magnitude estimate. AI text-serving electricity vs. meat production to retail. Not a full life-cycle or diet-savings comparison.
By CO₂e
Not to scale: the shorter bar is drawn wider than its value so it stays visible.
By water consumed
Assumption, about the meat figures
Why a 250 g steak?. We use a 250 g raw, boneless steak as an illustrative comparison, not a standard serving or dietary recommendation. Cooked weight and bone-in cuts would need different assumptions.
Why this beef carbon footprint?. We use a global beef-herd estimate of 99.48 kg CO₂e per kilogram of retail product, not a measurement of your steak. It covers meat production through retail, while the AI estimate covers operational electricity.
Why this beef water footprint?. We use an estimated global average of 550 liters of blue water consumed per kilogram of beef, based on production in 1996 to 2005. It excludes rainwater and grey-water pollution indicators, and is not a water-scarcity score.
Different studies, different boundaries. The meat greenhouse-gas and water figures come from different global studies and periods, and none measures a particular portion. Freshwater withdrawals per kg of meat are larger, but water withdrawn is not water consumed, so they are not compared with AI water consumption. The comparison shows scale; it does not mean one choice replaces the other.
Water accounting boundary. The AI estimate adds water consumed on site and at power plants, not water withdrawn. Published factors do not say how much of that is fresh water, and the meat figures are blue water from a different study, so the water comparison is approximate.
Sources: Science / AAAS 2018; Our World in Data 2018; Ecosystems / Springer 2012; UNESCO-IHE / University of Twente 2010; Our World in Data 2018; Communications of the ACM 2025; Lawrence Berkeley National Laboratory 2024
Your AI use
App
Plan
Models: OpenAI does not state one fixed model per plan. Codex is included, but not Codex Cloud. The plan only sets the starting example. Prices in US dollars per month, checked 21 October 2026.
Source, about the plan details
ChatGPT plan and model. Plan prices and included Codex access are taken from OpenAI's opened pages, but model use and allowances vary by account, task and product availability. Current plan terms or a verified per-request model would change the description, not turn example usage into a measured footprint.
Sources: OpenAI 2026; OpenAI Help 2026; OpenAI Help 2026
Adjust assumptionsUsage numbers, model size, electricity grid
How much you use it
Starting example for this plan. Change any number.
Assumption, about the usage numbers
Daily usage examples. The usage presets are editable example workloads, not measured subscriber averages or provider limits. Personal usage logs or direct counts would change them; plan price alone does not determine calls or tokens.
Calls in an agent session. Assumed: about 24 model calls per session, the median of one developer's logged Claude Code sessions. Different tasks, agents and measurement definitions would change that count.
Agent-use days per year. The annual example assumes 220 coding workdays, while chat and reasoning run on 365 days. Holidays, vacation and personal schedules would change the annual total.
Chat request length. The example chat uses a 3,800-token cached prefix, 200 fresh input tokens and 500 generated tokens. Actual conversation length, cache reuse and model routing would change the result.
Reasoning request length. The example reasoning request uses a 7,700-token cached prefix, 300 fresh input tokens and 4,000 generated tokens, including hidden reasoning. Actual reasoning and answer lengths depend on the task and model.
Coding-agent model call. One modeled call uses a 100,000-token context, with 95,000 reused tokens, 5,000 fresh tokens and 1,000 generated tokens. Real tasks trigger varying numbers of calls and token volumes, so task logs would change the estimate.
Sources: Simon P. Couch 2026; The Climate Brink 2026; OpenAI Help 2026; Google Jules 2026; Epoch AI 2025; NVIDIA Technical Blog 2026; University of Michigan / arXiv 2025
Model size scenario
Providers do not publish model sizes. This plan starts at medium.
Assumption, about the model size scenarios
Hypothetical model size. Small, medium and large are hypothetical parameter, layer and attention-width scenarios, not disclosed specifications of subscription models. Verified architecture details for a selected request would change the chosen scenario.
Sources: OpenAI / arXiv 2020; Epoch AI 2025; Google 2025
Electricity grid of the data center
Country changes CO₂e only; water uses the same reference factors for every country.
Assumption, about the grid figure
United States grid. The US preset uses Ember's 2025 lifecycle intensity for total generation, 0.384 kg CO₂e per kWh. EPA's eGRID2023 direct-emissions factor is lower, about 0.35 kg CO₂ per kWh, because it leaves out upstream emissions.
Location-based grid emissions. Lifecycle CO₂e per Ember: an annual country-level estimate for total electricity generation, including upstream and supply-chain emissions. Actual data-center emissions vary by location, time and electricity contracts, and a provider's market-based claims are not used.
Sources: Ember 2025; US Environmental Protection Agency 2025; GHG Protocol 2015
DetailsWater, CO₂e and electricity totals, car distance, published estimates
Totals
- Water consumed
- 930 mL
- scenario range 99 mL to 8.6 L
Assumption, about the water figure
On-site water. The central value is LBNL's estimated US data-center site water use, 0.36 L per kWh (2023). The range runs from 0.15 L (AWS 2024 fleet average) to 1.08 L (implied by Google's 0.26 mL per 0.24 Wh prompt). It is applied to all serving electricity, which slightly overcounts facility overhead.
Electricity-generation water. The central value is Li et al.'s US-average water consumed by power plants, 3.1 L per kWh. The range runs from 1.48 L (Li et al., Ireland) to 4.52 L (LBNL's 2023 estimate for US data-center electricity). The real value depends on which power plants supply the data center and how they are cooled.
Water accounting boundary. The AI estimate adds water consumed on site and at power plants, not water withdrawn. Published factors do not say how much of that is fresh water, and the meat figures are blue water from a different study, so the water comparison is approximate.
Scenario range. The low and high results vary utilization, serving allowance, cache misses and water factors using selected inputs. This is not a statistical confidence interval, and actual use can fall outside it.
Sources: Amazon 2024; Google 2025; Lawrence Berkeley National Laboratory 2024; Communications of the ACM 2025; Ecosystems / Springer 2012; Epoch AI 2025; University of Michigan / arXiv 2025
- CO₂e, electricity only
- 100 g
- scenario range 23 g to 590 g
- Like driving about 410 m in a typical US gasoline car.
Assumption, about the CO₂e figure
United States grid. The US preset uses Ember's 2025 lifecycle intensity for total generation, 0.384 kg CO₂e per kWh. EPA's eGRID2023 direct-emissions factor is lower, about 0.35 kg CO₂ per kWh, because it leaves out upstream emissions.
Location-based grid emissions. Lifecycle CO₂e per Ember: an annual country-level estimate for total electricity generation, including upstream and supply-chain emissions. Actual data-center emissions vary by location, time and electricity contracts, and a provider's market-based claims are not used.
Scenario range. The low and high results vary utilization, serving allowance, cache misses and water factors using selected inputs. This is not a statistical confidence interval, and actual use can fall outside it.
Car tailpipe comparison. The car comparison uses EPA's typical US gasoline-car tailpipe CO₂ per mile converted to kilometers. It excludes fuel production and vehicle manufacture, and does not describe an individual car or another country's fleet.
Sources: Ember 2025; US Environmental Protection Agency 2025; GHG Protocol 2015; Google 2025; Epoch AI 2025; University of Michigan / arXiv 2025; Lawrence Berkeley National Laboratory 2024; US Environmental Protection Agency 2026
- Electricity used
- 270 Wh
- scenario range 61 Wh to 1.5 kWh, at the medium model size
Assumption, about the electricity figure and the scenario range
Inference energy formula. The formula counts dense linear-layer work plus causal attention separately for fresh input and generated tokens. It is a simplified arithmetic model, so memory traffic, batching, model architecture and software can change actual electricity use.
H100 reference chip. We use an NVIDIA H100 SXM reference with about 989 trillion dense BF16 operations per second and up to 700 watts of rated power. A provider's actual chips, precision and software would change this reference estimate.
Generation compute utilization. We assume generation reaches 10% of peak compute throughput. This is a rough scenario, and batching, latency, model design and context length would change it.
Input-processing compute utilization. We assume input processing reaches 40% of peak compute throughput. This is a scenario value, and workload shape or measured utilization for the actual serving system would change it.
Generation power draw. We assume generation draws 70% of the chip's rated power; compute utilization is not the same as power draw. Provider power measurements would replace this illustrative setting.
Input-processing power draw. We assume input processing draws 100% of rated power, based on observed behavior in other systems rather than this provider. Measurements of the target system would change the value.
Serving-system allowance. Google measured 1.72 times fleet-average active-accelerator energy for comprehensive serving. We use it as an illustrative allowance for other services; provider-specific measurements would change it.
Cache-miss share. We assume 10% of calls cannot reuse the modeled prefix and must process the full input. Cache expiry, routing and prefix changes would alter this share; provider cache telemetry would replace it.
Scenario range. The low and high results vary utilization, serving allowance, cache misses and water factors using selected inputs. This is not a statistical confidence interval, and actual use can fall outside it.
Sources: OpenAI / arXiv 2020; NVIDIA Technical Blog 2026; NVIDIA 2026; Epoch AI 2025; University of Michigan / arXiv 2025; Microsoft Azure / ASPLOS 2024; Google 2025; Simon P. Couch 2026; The Climate Brink 2026; Lawrence Berkeley National Laboratory 2024
- A year of this use takes about 1/240 of the CO₂e of one steak.
- A year of this use takes about 1/150 of the water of one steak.
How this compares with published estimates
This tool, central value at the medium model size: 0.15 Wh per chat message, 80 Wh per coding-agent session, no coding day because no sessions are set.
| Published | Value | This tool |
|---|---|---|
| Google 2025Median Gemini Apps text prompt, measured | 0.24 Wh | Per chat message: 0.15 Wh |
| Sam Altman 2025Average ChatGPT query, stated without method | 0.34 Wh | Per chat message: 0.15 Wh |
| Epoch AI 2025GPT-4o query, modeled | about 0.3 Wh | Per chat message: 0.15 Wh |
| Simon P. Couch 2026Median Claude Code session, modeled from logs | 41 Wh | Per coding-agent session: 80 Wh |
| The Climate Brink 2026Median agent session, modeled (250 to 1,200 Wh) | 600 Wh | Per coding-agent session: 80 Wh |
| Simon P. Couch 2026Median coding day, modeled from logs | 1.3 kWh | Per coding day: no sessions set |
| The Climate Brink 2026Average day of his own agent use, modeled (1.2 to 5.9 kWh) | 3.0 kWh | Per coding day: no sessions set |
Source, about the published estimates
Published energy comparisons. Published figures use different models, dates, workloads and accounting boundaries. They provide context rather than a directly interchangeable per-request factor, and new comparable measurements could change the context.
Sources: Google 2025; Sam Altman 2025; Epoch AI 2025; Simon P. Couch 2026; The Climate Brink 2026
how this is calculated
Method and sources
Show the formula, every input and all sources
Formula
N = active parameters, L = layers, D = attention width
C = cached prefix tokens, U = fresh input tokens, T = C + U
O = generated tokens, including hidden reasoning
F_prefill = 2·N·U + 4·L·D·[U·C + U(U+1)/2]
F_decode = 2·N·O + 4·L·D·[O·T + O(O−1)/2]
s_prefill = F_prefill / (peak FLOP/s × compute use while reading)
s_decode = F_decode / (peak FLOP/s × compute use while generating)
chip J = s_prefill × rated W × power share while reading
+ s_decode × rated W × power share while generating
call Wh = chip J × serving allowance / 3600
cache miss: C = 0, U = T
expected = (1 − miss share) × cached call + miss share × cold call
year kWh = 365 × (chats × chat Wh + reasoning × reasoning Wh) / 1000
+ workdays × sessions × 24 × agent call Wh / 1000
water L = kWh × (cooling water + power-plant water per kWh)
CO₂e kg = kWh × grid CO₂e per kWh
average day = year / 365The formula counts dense linear-layer work plus causal attention separately for fresh input and generated tokens. It is a simplified arithmetic model, so memory traffic, batching, model architecture and software can change actual electricity use.Sources: OpenAI / arXiv 2020; NVIDIA Technical Blog 2026
Model size scenarios
Small, medium and large are hypothetical parameter, layer and attention-width scenarios, not disclosed specifications of subscription models. Verified architecture details for a selected request would change the chosen scenario.Sources: OpenAI / arXiv 2020; Epoch AI 2025; Google 2025
| Scenario | Active parameters | Layers | Attention width | Per chat message | Per reasoning question | Per coding-agent model call | Per agent session |
|---|---|---|---|---|---|---|---|
| Small | 10B | 40 | 5,120 | 0.039 Wh | 0.29 Wh | 1.1 Wh | 27 Wh |
| Medium | 40B | 64 | 8,192 | 0.15 Wh | 1 Wh | 3.3 Wh | 80 Wh |
| Large | 100B | 100 | 12,288 | 0.37 Wh | 2.6 Wh | 8 Wh | 190 Wh |
Electricity per request at the central scenario, including cache misses.
Request shapes
| Request | Cached prefix | Fresh input | Generated | Kind |
|---|---|---|---|---|
| Chat message | 3,800 | 200 | 500 | Assumption |
| Reasoning question | 7,700 | 300 | 4,000 | Assumption |
| Coding-agent model call | 95,000 | 5,000 | 1,000 | Assumption |
Token counts. Generated tokens include hidden reasoning.
Scenario range
The low and high results vary utilization, serving allowance, cache misses and water factors using selected inputs. This is not a statistical confidence interval, and actual use can fall outside it.Sources: Google 2025; Epoch AI 2025; University of Michigan / arXiv 2025; Lawrence Berkeley National Laboratory 2024
The low end sets every input below to its low column at once, the high end to its high column. Model size is your choice in the estimator and is not part of the range.
| Input | Low end | Central | High end | Sources |
|---|---|---|---|---|
| Compute use while generating | 30 % of peak | 10 % of peak | 3.3 % of peak | Epoch AI 2025; University of Michigan / arXiv 2025; Microsoft Azure / ASPLOS 2024 |
| Compute use while reading input | 90 % of peak | 40 % of peak | 13 % of peak | Microsoft Azure / ASPLOS 2024; NVIDIA Technical Blog 2026 |
| Serving-system allowance on chip energy | 1.3 × | 1.72 × | 2.4 × | Google 2025 |
| Calls that miss the cache | 2 % of calls | 10 % of calls | 30 % of calls | NVIDIA Technical Blog 2026; Simon P. Couch 2026; The Climate Brink 2026 |
| Cooling water at the data center | 0.15 L per kWh | 0.36 L per kWh | 1.08 L per kWh | Amazon 2024; Google 2025; Lawrence Berkeley National Laboratory 2024 |
| Water at the power plant | 1.48 L per kWh | 3.1 L per kWh | 4.52 L per kWh | Lawrence Berkeley National Laboratory 2024; Communications of the ACM 2025 |
The AI estimate adds water consumed on site and at power plants, not water withdrawn. Published factors do not say how much of that is fresh water, and the meat figures are blue water from a different study, so the water comparison is approximate.Sources: Communications of the ACM 2025; Lawrence Berkeley National Laboratory 2024; Ecosystems / Springer 2012
Fixed inputs
| Input | Value used | Kind | Sources |
|---|---|---|---|
| Reference chip peak compute (H100, dense BF16) | 989 trillion FLOP per second | Assumption | NVIDIA 2026 |
| Reference chip rated power (H100 SXM) | 700 W | Assumption | NVIDIA 2026 |
| Power draw while generating | 70 % of rated power | Assumption | Microsoft Azure / ASPLOS 2024; Epoch AI 2025 |
| Power draw while reading input | 100 % of rated power | Assumption | Microsoft Azure / ASPLOS 2024 |
| Model calls per coding-agent session | 24 calls | Assumption | Simon P. Couch 2026 |
| Coding-agent workdays per year | 220 days | Assumption | Own assumption |
| Beef steak, raw and boneless | 0.25 kg | Assumption | Own assumption |
| Beef greenhouse gases (beef herd, global mean) | 99.48 kg CO₂e per kg | Sourced | Science / AAAS 2018; Our World in Data 2018 |
| Beef blue water consumption (global average) | 550 L per kg | Sourced | Ecosystems / Springer 2012; UNESCO-IHE / University of Twente 2010 |
| Lamb portion, raw | 0.2 kg | Assumption | Own assumption |
| Lamb and mutton greenhouse gases (global mean) | 39.72 kg CO₂e per kg | Sourced | Science / AAAS 2018; Our World in Data 2018 |
| Sheep meat blue water consumption (global average) | 522 L per kg | Sourced | Ecosystems / Springer 2012; UNESCO-IHE / University of Twente 2010 |
| Pork portion, raw | 0.2 kg | Assumption | Own assumption |
| Pig meat greenhouse gases (global mean) | 12.31 kg CO₂e per kg | Sourced | Science / AAAS 2018; Our World in Data 2018 |
| Pig meat blue water consumption (global average) | 459 L per kg | Sourced | Ecosystems / Springer 2012; UNESCO-IHE / University of Twente 2010 |
| Chicken portion, raw | 0.2 kg | Assumption | Own assumption |
| Poultry meat greenhouse gases (global mean, chicken proxy) | 9.87 kg CO₂e per kg | Assumption | Science / AAAS 2018; Our World in Data 2018 |
| Chicken meat blue water consumption (global average) | 313 L per kg | Sourced | Ecosystems / Springer 2012 |
| Typical US gasoline car, tailpipe | 0.24855 kg CO₂ per km | Sourced | US Environmental Protection Agency 2026 |
Electricity grid presets
- United States: 0.384 kg CO₂e per kWh · Ember 2025; US Environmental Protection Agency 2025
- Germany: 0.334 kg CO₂e per kWh · Ember 2025; Umweltbundesamt 2026
- Ireland: 0.263 kg CO₂e per kWh · Ember 2025
- France: 0.041 kg CO₂e per kWh · Ember 2025
- Sweden: 0.035 kg CO₂e per kWh · Ember 2025
- Poland: 0.591 kg CO₂e per kWh · Ember 2025
- India: 0.671 kg CO₂e per kWh · Ember 2025
Lifecycle CO₂e per Ember: an annual country-level estimate for total electricity generation, including upstream and supply-chain emissions. Actual data-center emissions vary by location, time and electricity contracts, and a provider's market-based claims are not used.Sources: Ember 2025; GHG Protocol 2015
Meat comparison
| Portion | CO₂e | Blue water | Withdrawals per kg, context only |
|---|---|---|---|
| 250 g steak | 25 kg | 140 L | 1,451.2 L |
| 200 g lamb portion | 7.9 kg | 100 L | 1,802.8 L |
| 200 g pork portion | 2.5 kg | 92 L | 1,795.8 L |
| 200 g chicken portion | 2 kg | 63 L | 660 L |
The meat greenhouse-gas and water figures come from different global studies and periods, and none measures a particular portion. Freshwater withdrawals per kg of meat are larger, but water withdrawn is not water consumed, so they are not compared with AI water consumption. The comparison shows scale; it does not mean one choice replaces the other.Sources: Science / AAAS 2018; Ecosystems / Springer 2012; Our World in Data 2018
Scope
This estimate excludes training, hardware manufacture, network electricity, your device, tool execution, and image, video and voice generation. Including those activities would produce a broader footprint than this text-inference estimate.
Plans, prices and models were checked on 21 October 2026. They change often.
Sources
- NVIDIA H100 GPU · NVIDIA, 2026 · opened 2026-10-21H100 SXM product specifications: about 989 trillion dense BF16 operations per second and up to 700 W.
- Measuring the environmental impact of delivering AI at Google Scale · Google, 2025 · opened 2026-10-21Google's measured energy, water and emissions for the median Gemini text prompt, and its 1.72 factor from chip energy to full serving energy.
- How much energy does ChatGPT use? · Epoch AI, 2025 · opened 2026-10-21Modeled energy per ChatGPT query and typical chip utilization; an estimate, not provider metering.
- Characterizing Power Management Opportunities for LLMs in the Cloud · Microsoft Azure / ASPLOS, 2024 · opened 2026-10-21Measured power draw of language models on A100 servers, separately for reading input and for generating output.
- The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization · University of Michigan / arXiv, 2025 · opened 2026-10-21Measured GPU energy for model inference, compared with estimates based on rated chip power.
- Scaling Laws for Neural Language Models · OpenAI / arXiv, 2020 · opened 2026-10-21The rule of thumb for the compute per token of a transformer model.
- Co-Designing AI Model Attention for Fast, Interactive Long-Context Inference · NVIDIA Technical Blog, 2026 · opened 2026-10-21Attention compute for reading input and generating output, prefix caching, and newer attention designs.
- The Gentle Singularity · Sam Altman, 2025 · opened 2026-10-21States that an average ChatGPT query uses about 0.34 Wh; no method is given.
- Electricity use of AI coding agents · Simon P. Couch, 2026 · opened 2026-10-21Claude Code energy estimated from one developer's request logs: median session of 24 requests and about 592,000 tokens, 41 Wh.
- The real energy use of agentic AI · The Climate Brink, 2026 · opened 2026-10-21Energy of one person's AI agent use, modeled from token logs.
- Making AI Less 'Thirsty' · Communications of the ACM, 2025 · opened 2026-10-21Water consumed per kWh at data centers and at power plants, including US and Ireland averages.
- 2024 United States Data Center Energy Usage Report · Lawrence Berkeley National Laboratory, 2024 · opened 2026-10-21US data center water use per kWh, on site and at the power plants that supply them.
- 2024 Amazon Sustainability Report · Amazon, 2024 · opened 2026-10-21AWS reports 0.15 L of water per kWh at its data centers in 2024.
- eGRID Summary Data, eGRID2023 · US Environmental Protection Agency, 2025 · opened 2026-10-21US national emission rates for electricity generation, without upstream emissions.
- Entwicklung der spezifischen Treibhausgas-Emissionen des deutschen Strommix in den Jahren 1990–2025 (Climate Change 16/2026) · Umweltbundesamt, 2026 · opened 2026-10-21Germany's preliminary emission factors for 2025 electricity, with and without upstream chains.
- Yearly Electricity Data, global generation dataset · Ember, 2025 · opened 2026-10-21Life-cycle CO₂e intensity of total electricity generation by country, 2025.
- Scope 2 Guidance · GHG Protocol, 2015 · opened 2026-10-21Explains location-based and market-based accounting of electricity emissions.
- Introducing ChatGPT Go, now available worldwide · OpenAI, 2026 · opened 2026-10-21ChatGPT Go and Plus prices and what each plan includes.
- About ChatGPT Pro tiers · OpenAI Help, 2026 · opened 2026-10-21ChatGPT Pro tiers at $100, $200 and $500 and their allowances.
- Using Codex with your ChatGPT plan · OpenAI Help, 2026 · opened 2026-10-21Which ChatGPT plans include Codex and Codex Cloud; usage limits vary by plan.
- Plans & pricing · Anthropic, 2026 · opened 2026-10-21Claude plan prices, model access and relative usage limits.
- Use Claude Code with your Pro or Max plan · Anthropic Support, 2026 · opened 2026-10-21Claude Code is included with Pro and Max and shares their usage limits.
- Google AI Pro & Ultra: get access to Gemini 4 Argon and more · Google, 2026 · opened 2026-10-21US prices of Google AI plans and the models and agents each one includes.
- Limits and Plans · Google Jules, 2026 · opened 2026-10-21Daily Jules task limits by plan.
- Reducing food's environmental impacts through producers and consumers, Data S2 · Science / AAAS, 2018 · opened 2026-10-22Global mean greenhouse gases per kg of retail meat, production through retail, for beef herd, lamb and mutton, pig meat and poultry meat. Poultry stands in for chicken.
- Greenhouse gas emissions per kilogram of food product (Poore and Nemecek, 2018) · Our World in Data, 2018 · opened 2026-10-22Chart of the same greenhouse-gas means per kg of food product.
- A Global Assessment of the Water Footprint of Farm Animal Products · Ecosystems / Springer, 2012 · opened 2026-10-22Table 1: global average blue water consumed per kg of beef (550 L), sheep (522 L), pig (459 L) and chicken (313 L) meat, 1996 to 2005. Rainwater and pollution volumes are not included.
- Greenhouse Gas Emissions from a Typical Passenger Vehicle · US Environmental Protection Agency, 2026 · opened 2026-10-21EPA's typical US passenger car emits about 400 g of CO₂ per mile from the tailpipe.
- The green, blue and grey water footprint of farm animals and animal products, Volume 2: Appendices · UNESCO-IHE / University of Twente, 2010 · opened 2026-10-22Appendix tables that match the blue-water values to boneless beef and sheep cuts and to fresh pork cuts.
- Freshwater withdrawals per kilogram of food product · Our World in Data, 2018 · opened 2026-10-22Freshwater withdrawals per kg of food product. Water withdrawn is not water consumed, so these are context only.