One Almond vs. 38,000 ChatGPT Queries: Sam Altman’s AI Water Claim, Fact-Checked
Sam Altman has offered a striking answer to one of the most persistent criticisms of artificial intelligence.
How much water does ChatGPT really use?
According to the OpenAI CEO, approximately 38,000 ChatGPT queries use as much water as producing a single almond in California.
The comparison is memorable.
It is also more complicated than it initially sounds.
Altman's broader point—that the water footprint of one ordinary ChatGPT interaction can be extremely small compared with many everyday agricultural products—is plausible.
But the precise 38,000-to-one ratio cannot be reproduced using all of the numbers currently circulating alongside the claim.
Altman previously estimated that an average ChatGPT query uses about 0.000085 U.S. gallons of water, equivalent to roughly 0.32 milliliters.
Meanwhile, a 2019 U.S. Geological Survey-affiliated analysis of crop water use in California's Central Valley estimated approximately 3.56 liters of water per almond.
Divide 3.56 liters by 0.32 milliliters and the result is not 38,000.
It is roughly:
11,000 ChatGPT queries per almond.
There is, however, another peer-reviewed California almond study from 2019 that estimated a much larger total water footprint of about 12 liters per almond.
And 12 liters divided by Altman's roughly 0.32 milliliters per query produces approximately:
37,000 to 38,000 queries.
So Altman's comparison is not necessarily invented.
It appears remarkably close to one published almond-water estimate.
But the two sides of the comparison use different water-accounting methods, and OpenAI has not publicly provided enough methodology behind its 0.32-milliliter figure for outsiders to verify that they are truly comparable.
That distinction matters.
Because the real debate is not whether an almond or a ChatGPT prompt “wins” a water-consumption competition.
It is about how we measure the environmental footprint of an AI industry expanding at extraordinary speed.
What Sam Altman Actually Said
Altman discussed the issue during the September 1, 2026 premiere of Alex Heath's Sources podcast.
Asked about the growing backlash against AI and data-center development, he argued that concerns about extreme data-center water consumption have become a persistent public narrative that does not adequately reflect modern infrastructure.
He described the idea as a difficult-to-dispel “meme” and said that very large modern data centers can use amounts of water comparable with an office building.
He then introduced the almond comparison.
Altman said he was recalling the number from memory—explicitly acknowledging that it might not be exact—and estimated that the water associated with around 38,000 ChatGPT queries was comparable with the water used to produce one California almond.
He also joked that people routinely eat multiple almonds without feeling they have committed an environmental offense.
The argument was clear:
Public perception of AI's resource footprint may be dramatically out of proportion to the environmental attention given to many ordinary activities.
Where Does the 0.32 Milliliters Per ChatGPT Query Come From?
The number predates the 2026 podcast.
In June 2025, Altman published an essay titled The Gentle Singularity.
Among broader predictions about AI, robotics and falling intelligence costs, he included two unusually specific operational estimates.
According to Altman, an average ChatGPT query required approximately 0.34 watt-hours of electricity and 0.000085 gallons of water.
The water figure converts to approximately:
0.322 milliliters per query.
That is roughly one-fifteenth of a teaspoon.
At an individual level, it is tiny.
You could not practically measure 0.32 milliliters using an ordinary drinking glass.
But there is an important limitation.
Altman's post did not publish the detailed methodology behind the estimate.
It did not explain publicly which models were included in the “average.”
It did not specify the exact mix of short and long responses.
It did not provide a detailed geographic breakdown.
And it did not clearly define every component included in the water accounting.
Data Center Dynamics noted at the time that the figure was OpenAI's first notable public datapoint of this kind but that important methodological details remained unclear.
Therefore, 0.32 milliliters should be described as Altman's reported average, not as an independently audited universal physical constant for ChatGPT.
Why Every ChatGPT Query Cannot Literally Use the Same Amount of Water
There is no reason to expect every AI interaction to have an identical footprint.
Consider two prompts.
One asks:
“What is the capital of France?”
The other asks a sophisticated reasoning model to analyze thousands of words, use tools, conduct research and produce a long technical report.
Those tasks can require very different amounts of computation.
Water use can also change according to the hardware executing the request, utilization levels, cooling technology, local weather and the electricity supplying the facility.
A query processed in a cool region at a highly efficient facility powered largely by low-water electricity generation may have a different footprint from the same computational workload processed elsewhere.
That means “water per query” is an average abstraction.
Useful for scale.
Not universal for every interaction.
Imrul
The Almond Math Reveals a Fascinating Mismatch
Now consider the almond side.
Business Insider reported Altman's recent comments and compared them with a 2019 study involving U.S. Geological Survey researchers.
That study used satellite-derived evapotranspiration estimates and agricultural production data to examine crop water use in California's Central Valley.
For almonds, the researchers calculated approximately:
3.56 liters of water per almond.
Using Altman's 0.32-milliliter estimate:
3.56 liters equals 3,560 milliliters.
3,560 divided by approximately 0.322 gives roughly:
11,064 queries.
Business Insider itself said it could not independently verify Altman's exact comparison, although it concluded that available numbers broadly supported his larger argument that an almond can represent orders of magnitude more water than one chatbot prompt.
This is a crucial difference.
One almond being equivalent to roughly 11,000 average queries is still an extraordinary comparison.
But it is not 38,000.
So Where Does 38,000 Come From?
This is where the story becomes much more interesting.
Another peer-reviewed study, published in Ecological Indicators, examined the broader water footprint of California almonds.
Researchers Julian Fulton, Michael Norton and Fraser Shilling estimated an average total water footprint of approximately 12 liters per individual California almond.
Now perform the same calculation.
12,000 milliliters divided by approximately 0.322 milliliters per ChatGPT query gives roughly:
37,300 queries.
Allow for rounding and you arrive almost exactly at Altman's:
38,000 queries per almond.
That suggests the number may have a legitimate numerical origin.
But we still have a problem.
The 12-liter almond study and Altman's ChatGPT estimate may not be measuring water in the same way.
Not All “Water Footprint” Numbers Mean the Same Thing
This is perhaps the most important part of the entire debate.
The 12-liter almond figure represents a water footprint, not simply irrigation water physically pumped onto an orchard.
The researchers divided the footprint into categories commonly known as blue, green and grey water.
Blue water generally represents surface water and groundwater.
Green water represents rainwater stored in soil and used by the crop.
Grey water is different again: it is an accounting concept representing the volume of freshwater theoretically required to dilute pollutants to certain water-quality standards.
The study estimated approximately 10,240 liters of total water footprint per kilogram of California almond kernels, which translates to roughly 12 liters for a 1.2-gram almond.
That does not mean a farmer necessarily pours exactly 12 liters of irrigation water onto a tree for every almond harvested.
The 3.56-liter figure, by contrast, came from a satellite-based estimate of actual crop evapotranspiration divided by almond production.
The two studies therefore answer different environmental questions.
Both can be scientifically useful.
They are not interchangeable.
This Makes the 38,000 Comparison Directionally Useful but Scientifically Imperfect
If Altman's 38,000 figure is based on the roughly 12-liter total almond water footprint, the arithmetic works surprisingly well.
But that creates the next question:
Does his 0.32-milliliter ChatGPT estimate include the equivalent categories on the AI side?
For example, does it include only direct water consumed at the data center?
Does it include water associated with electricity generation?
Does it include construction?
Semiconductor manufacturing?
Supply-chain water?
The podcast transcript records Altman describing the comparison as broad, total water accounting rather than merely water operating inside a data center.
But the detailed underlying methodology has not been publicly released.
Without equivalent system boundaries on both sides, “38,000 queries versus one almond” should be treated as an illustrative comparison rather than a settled scientific ratio.
The Most Defensible Version of the Claim
Based on the publicly available evidence, a careful version would be:
One California almond can represent several liters to roughly 12 liters of water use or water footprint depending on the accounting method, while Sam Altman has estimated an average ChatGPT query at approximately 0.32 milliliters. Those figures imply that producing one almond can correspond to the water associated with roughly 11,000 to 38,000 average ChatGPT queries.
That is less catchy.
It is also much harder to misinterpret.
Is Altman Right That Modern Data Centers Use Water Like Office Buildings?
There is genuine evidence supporting this part of his argument.
Virginia's Joint Legislative Audit and Review Commission conducted one of the most extensive state-level examinations of data centers in 2024.
Virginia is particularly important because Northern Virginia is one of the largest data-center markets on Earth.
The commission found that most data centers used approximately the same amount of water as, or less than, an average large office building.
The average large-office benchmark used in the analysis was approximately 6.7 million gallons annually.
A presentation accompanying the state study found that about 83% of the data centers examined fell at or below that level.
So when Altman compares a modern data center's direct water use with an office building, he is not simply inventing the analogy.
A major government study in America's largest data-center region produced a remarkably similar finding.
But “Most” Is Doing Important Work
The Virginia report also shows why broad statements about data centers can be misleading.
Some facilities used dramatically more water.
The study found that 11 individual data-center buildings consumed more than 50 million gallons in 2023, while one building used approximately 243 million gallons—roughly 10% of the sector's measured state total.
Overall, Virginia data centers used an estimated 2.1 billion gallons in 2023.
Just over one-third came from reclaimed water rather than new freshwater withdrawals, and the industry's total represented less than 0.5% of statewide water withdrawals.
That creates a nuanced conclusion.
At the statewide level, the industry's direct water footprint was relatively small.
At most individual facilities, usage resembled a large office.
But some data centers were enormous water users.
And local effects can matter even when a statewide percentage looks small.
Water Is a Local Resource
This is why comparing national or statewide totals can hide the real environmental issue.
Imagine a data center requiring a certain amount of freshwater in a region with abundant rivers and reservoirs.
Now place the same facility in a drought-stressed basin with declining groundwater.
Identical water consumption can have very different consequences.
The Virginia commission itself noted that while Virginia overall is relatively water-rich, some individual localities face greater limitations, particularly where large surface-water resources are unavailable or groundwater is under management pressure.
That means a facility does not become environmentally harmless simply because its annual consumption resembles an office building.
Location matters.
Timing matters.
Water source matters.
There Is Also a Water Footprint Outside the Data Center
This is the biggest challenge to using only on-site water consumption when evaluating AI infrastructure.
Data centers consume extraordinary amounts of electricity.
Generating that electricity can itself require water.
Thermal power stations may need cooling.
Hydroelectric generation depends directly on water flows.
Therefore, a facility can use relatively little water within its own property boundaries while still being indirectly connected to far greater water withdrawals elsewhere.
A major Ceres report released in August 2026 focused specifically on this issue.
Ceres examined seven states hosting roughly half of U.S. data-center capacity and estimated that data-center electricity demand depended on approximately 3.4 trillion gallons of freshwater annually across the power-generation system.
The seven states were Virginia, Texas, California, Illinois, Georgia, Ohio and Arizona.
Ceres found that 78% of electricity generated in the states studied came from power plants that use water to operate, while many of those plants were in regions already facing significant water stress.
This does not contradict the Virginia office-building comparison.
It measures something different.
A Data Center Can Have Low Direct Water Use and High Indirect Water Dependency
Imagine an AI facility using an advanced closed-loop cooling system.
Its direct water demand may be modest.
That sounds excellent.
But if it draws enormous amounts of electricity from water-intensive power generation, another water footprint exists upstream.
Conversely, a facility with greater on-site water consumption might run primarily on wind and solar electricity, which generally require much less operational water than conventional thermal generation.
Which data center is “better”?
You cannot answer without defining what water use you are measuring.
That is why environmental accounting gets complicated so quickly.
Water Withdrawal Is Not the Same as Water Consumption
Another distinction regularly disappears from viral discussions.
A power plant may withdraw water from a river, use it for cooling and return much of it.
Only part may be consumed, primarily through evaporation or other losses.
Hydroelectric plants can move enormous volumes of water through turbines while returning the water downstream.
Therefore, trillion-gallon withdrawal figures should never automatically be interpreted as trillions of gallons permanently disappearing.
The same principle applies to agriculture.
Water applied to fields, evapotranspiration, rainfall and broader water-footprint accounting measure different things.
When people compare one number from AI with one number from farming without checking the definitions, dramatic but misleading conclusions become easy.
Altman Is Right About One Major Public-Perception Problem
People are not naturally good at evaluating environmental impacts across categories.
An almond is small.
A data center is enormous.
Therefore, intuitively, the data center feels environmentally worse.
But environmental accounting does not operate through visual scale alone.
Agriculture is one of humanity's largest users of freshwater.
Growing food requires photosynthesis, transpiration and huge areas of land.
A single nut may embody months of irrigation and plant growth.
A digital query, meanwhile, represents a microscopic fraction of a server's operating resources.
Comparing the two can therefore produce numbers that feel absurd even when the broad order of magnitude is plausible.
That is precisely why Altman's almond analogy is rhetorically effective.
But Agriculture and AI Serve Different Functions
The comparison can also become misleading in another direction.
Human beings require food.
They do not biologically require chatbot prompts.
That does not mean AI has no social value.
It means environmental comparisons should consider the service being provided.
A better question than:
“Does one almond use more water than ChatGPT?”
might be:
“How much value do we receive per unit of water, electricity, land and carbon from each activity?”
That answer depends heavily on the use case.
A ChatGPT query helping someone understand homework differs from one generating disposable spam.
An almond providing nutrition differs from food discarded uneaten.
Resource efficiency cannot be understood only through unit comparisons.
Scale Changes Everything
Even an extremely small per-query footprint becomes meaningful when multiplied by billions of interactions.
Suppose, purely as an illustration, Altman's 0.000085-gallon estimate were applied to one billion queries.
That would equal about:
85,000 gallons of water.
At one billion queries every day for a year, the total would exceed:
31 million gallons.
Again, this is only an illustration because actual query volumes, models and resource intensity vary.
The point is mathematical.
Tiny multiplied by enormous can become substantial.
This is exactly why environmental researchers focus not only on efficiency per query but on the growth of total AI demand.
Efficiency Can Improve While Total Consumption Still Rises
This phenomenon has appeared repeatedly throughout technological history.
A machine becomes more efficient.
Its cost falls.
More people use it.
Total consumption rises anyway.
AI may follow the same pattern.
If the water and energy required for one query fall by 90%, but global AI usage grows by 1,000%, total resource demand can still increase significantly.
Therefore, both statements can be true simultaneously:
AI queries are becoming extraordinarily efficient.
And:
AI infrastructure's total environmental footprint is growing.
These are not contradictory claims.
The AI Water Debate Often Confuses Four Completely Different Questions
When people ask, “How much water does AI use?” they may actually mean one of several things.
They might be asking about water consumed directly by cooling equipment.
They might mean water withdrawn but returned.
They might mean water associated with producing electricity.
Or they may be asking for a full lifecycle footprint including chip fabrication, construction and supply chains.
Those questions can produce radically different answers.
This is why one article can say a ChatGPT query consumes less than a milliliter while another analysis produces much larger numbers.
Different boundaries can create different results without either calculation necessarily being fraudulent.
OpenAI’s 0.32-Milliliter Figure Needs More Transparency
Altman's disclosure was useful because very few major AI companies have historically published clear per-query environmental numbers.
But the next step should be methodology.
A scientifically robust disclosure would explain what counts as an “average query,” which model families are represented, whether reasoning workloads are included, what hardware is assumed, what cooling systems are represented and whether water used for electricity generation is included.
Without that information, independent researchers cannot reproduce the figure.
Altman's 0.32-milliliter estimate may be accurate.
But reproducibility would make it far more valuable than a CEO's assertion.
Imrul
The 38,000 Number Deserves the Same Treatment
The podcast comment itself contained an important caveat.
Altman said he was working from memory and that the number might not be exact.
That should shape how the claim is reported.
It was not presented as a peer-reviewed result with a table of assumptions.
It was a conversational comparison intended to communicate scale.
The remarkable discovery is that a published 12-liter almond water-footprint estimate does reproduce almost exactly the ratio he remembered.
That lends the claim plausibility.
But it still does not prove the accounting boundaries are equivalent.
Why the 3.56-Liter Almond Number Is Also Not “The” Final Answer
The USGS-affiliated study estimated 3.56 liters using satellite-derived evapotranspiration and agricultural production.
Another peer-reviewed paper estimated about 12 liters when using a wider water-footprint framework.
Neither necessarily makes the other wrong.
They are measuring related but different concepts.
Water science is full of these distinctions.
Consumptive use.
Applied irrigation.
Blue water.
Green water.
Grey water.
Withdrawals.
Evapotranspiration.
Lifecycle consumption.
A social-media graphic usually removes all of them.
Then people argue over one number as though there were a single universally correct definition.
Does Eating Almonds Therefore Harm the Environment More Than Using ChatGPT?
That conclusion would be much too simplistic.
California almond production has genuine water implications, particularly because much of it occurs in a dry agricultural region.
But almonds also provide food, nutrition, agricultural income and export revenue.
AI has its own benefits and environmental costs.
A responsible comparison would examine multiple impacts, including:
Water scarcity.
Energy use.
Carbon emissions.
Land use.
Materials.
Economic value.
Social value.
Geographic context.
Reducing the conversation to “almonds bad, ChatGPT good” would misunderstand the point as badly as claiming every AI query drains a bottle of water.
Is AI Water Criticism a “Meme”?
Some claims circulating online have clearly exaggerated AI water use.
Altman's frustration is understandable.
But describing the entire concern as a meme risks overcorrecting.
Virginia's own government analysis found direct water use was currently sustainable statewide but explicitly warned that use was growing and could be better managed.
Ceres argues that purchased electricity introduces a large indirect water dependency that is frequently missing from company disclosures.
And rapid data-center expansion can produce significant local impacts even when average national statistics appear modest.
So the strongest conclusion is neither:
“AI is draining the planet dry.”
Nor:
“AI water concerns are meaningless.”
The evidence supports something much more interesting.
Modern AI Is Becoming More Water-Efficient, but Infrastructure Scale Still Matters
Data-center engineers have powerful incentives to reduce water use.
Water costs money.
Cooling infrastructure costs money.
Permitting delays cost money.
Community opposition costs money.
Efficient cooling can therefore produce both environmental and financial benefits.
Closed-loop systems, dry cooling, reclaimed water and climate-aware facility design can reduce freshwater requirements dramatically.
But AI companies are simultaneously building infrastructure at unprecedented scale.
That means engineering improvements must race against demand growth.
Sam Altman’s Strongest Point May Not Be the Almond Comparison
Near the end of his discussion, Altman offered a more important observation.
He argued that the AI industry's response to public skepticism cannot consist only of explaining that critics are wrong.
Companies have to demonstrate value.
Business Insider quoted him saying that the industry's job is to make AI products genuinely useful enough that people understand why the infrastructure exists.
That may be the real long-term challenge.
Communities asked to host enormous power-hungry infrastructure will want to know what they receive in return.
Not just globally.
Locally.
The Environmental Debate Will Not Be Settled by One Viral Comparison
“One almond equals 38,000 ChatGPT queries” is likely to spread because it compresses an enormously complicated infrastructure problem into one image.
A nut.
A chatbot.
A number.
But the real story requires several numbers.
Altman's reported average query:
About 0.32 milliliters.
USGS-affiliated almond crop-water estimate:
About 3.56 liters per almond.
Equivalent under those two figures:
About 11,000 ChatGPT queries.
Broader California almond water-footprint estimate:
About 12 liters per almond.
Equivalent using that number:
About 37,000 to 38,000 queries.
Virginia's finding:
Most individual data centers examined used roughly the same or less water than an average large office building—but a minority used dramatically more.
Ceres's warning:
The electricity powering data centers introduces a much larger and often overlooked water dependency beyond the facility itself.
All of those facts can coexist.
Frequently Asked Questions About ChatGPT, Almonds and AI Water Use
Did Sam Altman really say one almond uses as much water as 38,000 ChatGPT queries?
Yes.
During the September 1, 2026 Sources podcast, Altman estimated from memory that approximately 38,000 ChatGPT queries correspond to the water used to produce one California almond.
Did Altman say the number was exact?
No.
He explicitly indicated that he was recalling the calculation from memory and that it might not be exact.
How much water does one ChatGPT query use?
Sam Altman previously estimated approximately 0.000085 U.S. gallons, or around 0.32 milliliters, for an average query.
Is 0.32 milliliters independently verified?
Not fully.
Altman's public post did not provide enough methodological detail for independent researchers to completely reproduce the calculation.
It should therefore be treated as OpenAI leadership's reported average rather than a universally established measurement.
How much water does one California almond use?
It depends on methodology.
A USGS-affiliated 2019 study estimated approximately 3.56 liters per almond based on crop evapotranspiration.
A separate peer-reviewed study estimated a broader total water footprint of roughly 12 liters per almond.
How many ChatGPT queries equal 3.56 liters?
Using Altman's roughly 0.32-milliliter estimate, approximately 11,000 average queries correspond to 3.56 liters.
How many queries equal 12 liters?
Using the same ChatGPT estimate, approximately 37,000 to 38,000 queries correspond to 12 liters.
Does that explain Altman’s 38,000 number?
Possibly.
His comparison aligns remarkably closely with the published 12-liter-per-almond water-footprint estimate.
However, there is no publicly available detailed calculation from Altman identifying that study as his source.
Why did Business Insider mention 3.56 liters?
Business Insider referenced the 2019 USGS-affiliated Central Valley crop-water study and said it could not independently verify Altman's exact ratio, although the broader claim that an almond uses vastly more water than one chatbot prompt was consistent with available estimates.
Which almond estimate is correct?
Both can be correct within their methodologies.
The 3.56-liter estimate focuses on crop water use derived from evapotranspiration.
The approximately 12-liter estimate uses a broader water-footprint framework incorporating multiple categories of water.
What is grey water in a water-footprint study?
Grey water is generally a conceptual estimate of the freshwater volume needed to dilute pollutants to meet specified water-quality thresholds.
It should not be interpreted as an equivalent amount of literal irrigation water poured onto a field.
Are almonds unusually water-intensive?
California almonds have substantial irrigation requirements and have become prominent in debates over agricultural water use, particularly because they are cultivated extensively in the state's dry Central Valley.
Water requirements vary by location, year, orchard conditions and accounting method.
Did Altman say data centers use no water?
No.
His argument was that public perception exaggerates their direct water demand and that modern very large facilities can use amounts comparable with office buildings.
Is the office-building comparison supported by evidence?
Yes, at least in Virginia.
A 2024 Virginia government study found that most data centers examined used about the same amount of water as, or less than, an average large office building.
What percentage of Virginia data centers were at or below that benchmark?
The state presentation reported approximately 83%.
How much water does a large office building use?
The Virginia comparison used approximately 6.7 million gallons per year as the average large-office benchmark.
Did some data centers use much more?
Yes.
Eleven Virginia data-center buildings exceeded 50 million gallons annually in 2023, and one reached approximately 243 million gallons.
How much water did Virginia data centers use in total?
The 2024 state analysis estimated approximately 2.1 billion gallons during 2023, with more than one-third supplied through reclaimed water.
Was that a large percentage of Virginia’s overall water use?
No.
The report estimated data centers accounted for less than 0.5% of statewide withdrawals in 2023.
Does that prove water concerns are irrelevant?
No.
Water impacts are highly local.
A relatively small statewide percentage can still matter in a community with constrained groundwater or limited surface-water availability.
Virginia's commission specifically warned that availability varies substantially among localities.
Do data centers still require cooling?
Yes.
Computing equipment produces heat, and industrial-scale data centers require substantial cooling systems.
Some systems use more water than others.
The Virginia report identified facility size, computing density and cooling design as major determinants of water demand.
Is direct cooling water the entire AI water footprint?
No.
Electricity generation can have its own water requirements.
This indirect footprint may occur far from the data center itself.
What did Ceres find in 2026?
Ceres estimated that data centers in seven major U.S. data-center states depended on approximately 3.4 trillion gallons of freshwater annually through electricity generation.
Does that mean data centers consumed 3.4 trillion gallons?
Not necessarily.
Power-sector water accounting often includes large withdrawals in which much of the water is subsequently returned.
Withdrawal and consumption should not be treated as identical concepts.
Why does electricity need water?
Many thermal power plants use water for cooling.
Hydroelectric generation obviously depends on water flows.
Wind and solar photovoltaic electricity generally require far less operational water.
Can a low-water data center still have a large indirect footprint?
Yes.
If its electricity comes from water-intensive power generation, the upstream water dependency can exceed its direct cooling demand.
Can renewable energy reduce the water footprint?
Often.
Wind and solar photovoltaic generation typically require little operational water compared with thermal power generation.
Hydropower is renewable but remains directly dependent on water availability.
Is one ChatGPT query environmentally insignificant?
At an individual level, Altman's reported water figure is extremely small.
But aggregate impacts depend on billions of interactions, model complexity, electricity use and infrastructure growth.
Could AI become more efficient but still use more total water?
Yes.
If usage grows faster than efficiency improves, total water and electricity demand can rise even while each individual query becomes more efficient.
Is it fair to compare AI with agriculture?
It can help illustrate scale, but only if the accounting methods are clearly explained.
An almond water footprint and a ChatGPT operational-water estimate may include different environmental boundaries.
Is eating almonds worse than using ChatGPT?
There is no scientifically useful universal answer.
Food production and AI provide different benefits and have different environmental impacts.
Meaningful comparisons require more than a single water metric.
What is the biggest problem with the viral 38,000-query claim?
The number can appear far more precise than the underlying evidence supports.
It matches one published almond water-footprint estimate remarkably well, but another respected study produces a ratio closer to 11,000.
And OpenAI has not publicly released enough methodology to determine whether identical water-accounting boundaries were used on both sides.
What is the most accurate conclusion?
Sam Altman's larger argument is supported better than the exact viral number.
One ordinary ChatGPT query appears to represent a very small amount of water when using his published estimate.
Producing a California almond can represent several liters or more of water depending on the accounting methodology.
That means one almond plausibly represents the water associated with many thousands of ordinary ChatGPT queries.
Whether the most defensible number is 11,000, 38,000 or something else depends on what exactly is being counted.
And that is the lesson hidden beneath the viral comparison.
Environmental accounting is not just about finding the biggest number.
It is about asking what the number means.
A data center can use little water directly while relying on water-intensive electricity.
An almond can be assigned 3.56 liters under one crop-water method and around 12 liters under a broader footprint method.
A ChatGPT query can average 0.32 milliliters under Altman's estimate while individual workloads differ substantially.
All of those statements can be true at once.
The useful debate is therefore not whether Sam Altman “destroyed” criticism of AI water use with an almond.
Nor is it whether every ChatGPT prompt is secretly draining reservoirs.
The useful question is much more practical:
Can AI companies make increasingly powerful computing systems while transparently measuring—and continuously reducing—the real resources required to run them?
Altman believes the answer is yes.
As AI infrastructure expands, the numbers will have to prove it.
