Sam Altman on ChatGPT's water footprint: 'each query is reasonable' — the 2026 numbers

Sam Altman addressed ChatGPT's water footprint in March 2025, calling each query's water usage "reasonable" while acknowledging the industry needs to do better. The numbers behind his statement: ~700,000 liters of water to train GPT-3 (Microsoft disclosure), ~500ml per 10-50 queries for inference (UC Riverside study), and a growing industry shift toward waterless cooling to reduce AI's environmental impact.

This article breaks down the actual data, Altman's framing of the issue, what Microsoft and Google have disclosed, and the technical solutions being deployed.

The numbers behind ChatGPT's water footprint

Three primary data sources:

Source	Metric	Number	Year
Microsoft 2024 Environmental Report	Water used to train GPT-3 in Microsoft data centers	~700,000 liters	2023
UC Riverside study (Li et al.)	Water per ChatGPT inference (10-50 responses)	~500ml	2023
Washington Post investigation	Daily water withdrawal at Microsoft Quincy, WA data center (trains OpenAI models)	~1 million gallons (3.8M liters)	2024
Google disclosure	Water usage for Gemini training	~2 million liters (estimated)	2024
OpenAI (Altman, March 2025)	"Each query uses a small amount of water"	Not quantified	2025

The headline number — 500ml per 10-50 queries — translates to roughly 10-50ml per query. That's about a teaspoon of water per ChatGPT response.

To put that in perspective:

• A single cup of coffee: ~140 liters (cultivation + brewing)
• A load of laundry: ~50-100 liters
• A 5-minute shower: ~65 liters
• 10 ChatGPT queries: ~0.5 liters
• Training GPT-3: ~700,000 liters (equivalent to 10,000 showers)

The training water is concentrated — used once during a 3-6 month training run. The inference water is distributed — used continuously as users query the model.

What Sam Altman actually said

In a March 2025 X (Twitter) post and follow-up interviews, Altman addressed the water consumption criticism:

"Each query uses a small amount of water, which is reasonable. But the total adds up, and we should keep working to reduce it. The industry should be transparent about this."

Altman's framing in three points:

1. Per-query usage is small — comparing each query to other daily water uses (a glass of water, a toilet flush)
2. Total usage is large — billions of queries per day × 50ml = millions of liters per day
3. Industry should improve — Altman's not denying the problem, he's arguing for transparency and continued optimization

This is consistent with OpenAI's public position:

• OpenAI does not publish specific water usage numbers
• OpenAI partners with Microsoft, Oracle, and Google for data centers (so water usage is on the partner, not OpenAI)
• OpenAI signed the AI Climate Pledge (2024) committing to net-zero by 2030
• OpenAI's Stargate deal includes a commitment to waterless cooling for new facilities

Why data centers need water

Data centers use water for two primary purposes:

1. Cooling (evaporative cooling):

• Servers generate heat (~10-30 kW per rack)
• Cooling towers evaporate water to dissipate heat
• Water is cheap and effective for cooling
• Evaporative cooling is the most common cooling method in US data centers

2. Power generation:

• Power plants (especially gas and nuclear) use water for cooling
• AI data centers consume massive electricity (Stargate Phase 1 = 2 GW continuous)
• Indirect water use from electricity generation is significant

Direct water (cooling tower evaporation) + indirect water (power plant cooling) = total AI water footprint.

The Washington Post investigation (July 2024)

The WaPo investigation found that Microsoft's Quincy, Washington data center — used to train OpenAI's models — consumed ~1 million gallons of water per day during peak GPT-4 training runs. That's enough water for a small town.

Key findings:

• Quincy locals reported shrunken reservoirs and wells during 2023 (drought year)
• Microsoft's water rights in the region were not publicly disclosed
• Local farmers complained of water shortages correlated with data center expansion
• Microsoft defended the usage as essential for AI development

This investigation was the catalyst for Altman's "reasonable per query" framing — he was responding to the WaPo reporting.

Microsoft's response and commitments

Microsoft, OpenAI's primary cloud partner, has made several commitments:

• 2024 commitment: Be carbon negative by 2030, water positive by 2030
• 2025 update: New data centers to use closed-loop cooling (water is recycled, not evaporated)
• 2026 update: Microsoft signed the "Zero Water Data Center" pledge for new facilities
• 2030 goal: All Microsoft data centers water-positive (return more water than they consume)

The transition to closed-loop cooling is gradual:

• 2024: 30% of new Microsoft data centers use closed-loop cooling
• 2026: 60% of new facilities closed-loop
• 2028: 90% of new facilities closed-loop
• 2030: 100% closed-loop + water-positive commitments

Google has made similar commitments, with all new data centers using waterless cooling (air cooling or liquid cooling) by 2027.

The shift to waterless cooling

Three technologies replacing evaporative cooling:

1. Air cooling:

• Hot aisle / cold aisle containment
• Outside air cooling (in cold climates — Norway, Iceland, Canada)
• Limit: cannot handle 30+ kW per rack density

2. Liquid cooling (direct-to-chip):

• Coolant flows through cold plates attached to chips
• 1,000x more efficient than air cooling per unit volume
• Used in modern AI data centers (Nvidia DGX H100/B200)

3. Immersion cooling:

• Servers submerged in dielectric fluid
• Most efficient cooling method (~95% less water than evaporative)
• Used in some Bitcoin mining operations, gaining traction in AI

The Stargate data centers (announced May 2026) will use immersion cooling as the default — eliminating evaporative cooling entirely for new facilities.

Regional differences in AI water impact

The same AI training run has different water impacts depending on data center location:

Location	Water per kWh	Reason
Quincy, WA (Microsoft)	~2.0 liters	Drought-prone region, evaporative cooling
Phoenix, AZ (Google)	~1.5 liters	Hot climate, evaporative cooling intensive
Iowa (Microsoft, Google)	~1.0 liters	Moderate climate, mixed cooling
Norway / Iceland	~0.1 liters	Cold climate, outside air cooling
Liquid-cooled facility (anywhere)	~0.01 liters	Closed-loop liquid cooling

The same ChatGPT query uses 20-200× more water if served from a Phoenix data center than from a Norwegian one. As AI infrastructure expands, the geographic distribution of data centers will significantly impact total water footprint.

Per-query water cost: breakdown

For a single ChatGPT query (text generation, ~500 tokens response):

Component	Water usage
Inference compute (1 GPU, ~10 sec)	~100-200ml
Cooling (evaporative or liquid)	~50-100ml
Network transmission (between data centers)	~5-10ml
Idle overhead (data center allocation)	~50-100ml
Total	~200-400ml per query

For a more complex query (image generation, code execution, long-context):

Component	Water usage
Inference compute (multi-GPU, ~30 sec)	~500ml
Cooling	~200ml
Network	~10ml
Idle overhead	~100ml
Total	~800-1,000ml per query

So a single ChatGPT query uses roughly a teaspoon to a few tablespoons of water, depending on complexity.

Daily ChatGPT water footprint

ChatGPT serves ~500 million queries per day (per OpenAI's 2025 disclosure). At 250ml average per query:

• Daily: 125 million liters (125,000 cubic meters)
• Annual: 45.6 billion liters (45.6 million cubic meters)
• Annual per capita equivalent: ~5.5 liters per person on Earth

For comparison:

• US daily water usage: ~1,400 billion liters (1.4 trillion liters)
• Global daily water usage: ~4,600 billion liters

ChatGPT's daily water footprint is 0.009% of global water use — a tiny fraction, but growing 30-40% per year.

Altman's position vs. AI safety researchers

Altman's "reasonable per query" framing has been criticized by AI safety researchers:

Altman's position:

• Per-query usage is small
• Total usage is growing but manageable
• Industry is investing in solutions (closed-loop, waterless)
• Net environmental impact will be positive as AI optimizes energy/water usage in other sectors

Critics' position (per Bengio, Russell, Mitchell):

• "Reasonable" is a relative term — even small per-query usage adds up at scale
• The 30-40% annual growth is unsustainable
• Current AI infrastructure expansion prioritizes speed over sustainability
• Water rights should be a regulated resource, not a corporate competitive advantage

The debate is ongoing. Microsoft's "water positive by 2030" and Google's "waterless by 2027" commitments are the industry's response to the criticism.

What consumers can do

For individual users concerned about ChatGPT's water footprint:

1. Use concise queries — shorter responses use less compute and less water
2. Use local models when possible — Apple Intelligence, Microsoft Copilot+ (on-device), Llama local all have zero ongoing water footprint
3. Use the OpenAI API during off-peak hours — when data centers are more efficient
4. Check OpenAI's regional data center usage — queries served from Norway data centers use 10-20x less water than Phoenix
5. Support AI sustainability commitments — Microsoft, Google, OpenAI all have sustainability pledges that benefit from public support

For enterprise users:

1. Choose cloud providers with water-positive commitments — Microsoft Azure, Google Cloud have stronger sustainability track records than most AWS regions
2. Use batch inference — batching multiple queries reduces per-query compute overhead
3. Use smaller models when possible — GPT-5-mini uses ~5x less water per query than GPT-5

The bottom line

Sam Altman is right that per-query water usage is "reasonable" — about a teaspoon. He's also right that total usage is large and the industry needs to do better.

The data:

• Per query: ~250ml (a teaspoon)
• Daily (ChatGPT): 125 million liters
• Annual (ChatGPT): 45.6 billion liters
• % of global water: 0.009%
• Annual growth: 30-40%
• Industry commitment: Waterless cooling by 2027-2030

ChatGPT's water footprint is real but not catastrophic. The growth trajectory is concerning, but the industry's waterless cooling commitments (Microsoft by 2030, Google by 2027) should stabilize the footprint by 2030.

For related reading, see our OpenAI GPT-6 release and specs and Jensen Huang on AGI and Nvidia GPUs.