The viral claim that ChatGPT consumes 500 milliliters of water per query is wrong by orders of magnitude, but the real infrastructure crisis threatening AI's expansion is far more serious and barely discussed. A single ChatGPT response uses roughly 5 milliliters of water, not 500 milliliters, according to careful analysis of the original research that spawned the myth. Yet while individual users shouldn't feel guilty about their AI usage, the upstream story about what data centers do to power grids and mineral supplies reveals a genuine geopolitical vulnerability that could reshape the AI industry for decades.

Where Did the "Bottle of Water Per Query" Myth Come From?

The 500-milliliter-per-query statistic traces back to a 2023 research paper by Shaolei Ren at UC Riverside titled "Making AI Less Thirsty," which was later published in Communications of the ACM. However, Ren never claimed that a single query consumed 500 milliliters. Instead, he calculated that 500 milliliters was needed for 10 to 50 medium-length GPT-3 responses. The "per query" framing was a citation-chain mutation that escaped onto Twitter and never came back, eventually becoming filed as a discrete misinformation incident by the AIAAIC database.

Software engineer Sean Goedecke retraced Ren's math and found that the original estimate had compounded a per-page power figure misread as per-request, then applied it to GPT-3, a model roughly 10 times less efficient than current systems. Goedecke's reconstruction puts a current ChatGPT response closer to 5 milliliters of water. Google reports a median Gemini text prompt at 0.26 milliliters of water and 0.24 watt-hours of energy, making it the only first-party disclosure available.

"The bottle-per-query thing is completely untrue, totally insane, no connection to reality," said Sam Altman, OpenAI's CEO, when pressed on the statistic at the India AI Impact Summit in February 2026.

Sam Altman, CEO at OpenAI

What Does 0.3 Watt-Hours Actually Mean for Your Energy Use?

To understand whether AI queries are genuinely wasteful, it helps to translate the energy consumption into something tangible. A single ChatGPT text query costs roughly 0.3 watt-hours, a figure that Sam Altman published on his personal blog in June 2025 and that independent researchers at Epoch AI confirmed in February 2025 using GPT-4o on H100 GPUs. Hannah Ritchie at Our World in Data lined up Google's published Gemini number and confirmed the order of magnitude.

In practical terms, 0.3 watt-hours is equivalent to roughly two minutes of an LED bulb, four feet of driving a sedan, or about five seconds of streaming Netflix in high definition. If you sent a thousand ChatGPT prompts every single day for an entire year, you would raise your personal energy footprint by less than 1 percent. A single prompt costs roughly 1/150,000th of an average American's daily emissions.

However, not all AI queries are created equal. Reasoning models like OpenAI's o1 and Anthropic's Opus 4.7 High burn around 30 times more energy per query than chat-style queries, and long-context queries can hit 40 watt-hours. Video generation exists in another universe entirely, with a five-second AI-generated clip clocking in around 3.4 million joules according to MIT Tech Review investigation.

The Real Crisis: Where AI Gets Its Minerals

While the water myth has dominated headlines, the actual infrastructure bottleneck threatening AI's expansion is mineral supply. The $700 billion AI infrastructure buildout requires enormous quantities of copper, cobalt, nickel, lithium, and rare earth elements. Microsoft's 80-megawatt Chicago data center site alone required 2,100 tons of copper. Nickel, cobalt, and lithium are essential for the batteries that power data centers, while rare earths are critical to powering the magnets in server fans and hard drives that keep AI systems running.

The problem is severe: China is the leading refiner for 19 out of 20 of the most important strategic minerals, with an average market share of 70 percent according to the International Energy Agency. Without a new source of minerals, the United States faces a precarious dependence on China for the raw materials underpinning its technological and economic future. If left unaddressed, that vulnerability could hand Beijing enormous leverage over American industry for decades to come.

How Deep-Sea Mining Could Reshape AI's Supply Chain

In response to this mineral crisis, the Trump administration signed an executive order in 2025 directing the Department of Commerce and other agencies to pursue the exploration and exploitation of deep-sea resources. Several U.S. companies are now planning to extract polymetallic nodules, mineral-rich rock formations that sit unattached on the abyssal plain of the ocean floor.

American Ocean Minerals, which is in the process of closing a $1 billion merger with Odyssey Marine Exploration, holds investment interests in two companies that have earned licenses to conduct research in an exclusive economic zone around the Cook Islands in the South Pacific. The Cook Islands Seabed Mineral Authority estimates the region holds 6.7 billion metric tons of polymetallic nodules, including an estimated 20 million metric tons of cobalt, about 100 times the annual amount coming from the Democratic Republic of Congo, the world's top producer where China controls the majority of output.

"This could be hundreds of years of endowment," said Tom Albanese, former Rio Tinto CEO and now leading American Ocean Minerals, describing the scale of polymetallic nodules in one of the Cook Island zones.

Tom Albanese, CEO at American Ocean Minerals

Polymetallic nodules are potato-sized balls that form over millions of years as layers of metal oxides slowly accumulate around an existing object on the ocean floor, such as a shark tooth or shell fragment. They are composed primarily of manganese and iron but also contain economically valuable metals including nickel, cobalt, copper, and rare earth elements. The Cook Islands exclusive economic zone is over 770,000 square miles wide, nearly five times the size of California.

Steps to Understanding the Deep-Sea Mining Debate

• Environmental Concerns: The ocean floor is highly biodiverse, with the Ocean Exploration Trust finding "so many creatures we know absolutely nothing about" during a recent expedition around the Cook Islands. A separate report from the American Museum of Natural History estimated that deep-sea mining would decrease the abundance of animals by 37 percent in the Clarion-Clipperton Zone, a Pacific region also rich with polymetallic nodules.
• Regulatory Hurdles: The International Seabed Authority, which regulates mineral-resource-related activities in the ocean, has not yet approved any company to extract polymetallic nodules commercially. A meeting in March 2026 ended in a stalemate, and a growing coalition of 40 countries is now backing a moratorium on deep-sea mining, raising environmental, governance, and scientific concerns.
• Comparative Harm Assessment: Proponents argue that deep-sea mining may be less damaging than land mining, which contributes to massive habitat destruction, biodiversity loss, and toxic contamination of water sources. Many land operations, like those in the Democratic Republic of Congo, rely on forced labor, making alternative sources potentially more ethical despite environmental trade-offs.

The mineral demand is only expected to increase. The International Energy Agency estimated that demand for nickel, cobalt, and rare earth elements could more than double by 2040. For industry leaders like Albanese, the lack of mineral source diversification will remain a key geopolitical tension if new, independent sources aren't explored.

"I see the merits of continued engagement with China, but also see the risks of being overly dependent on the Chinese industry for any part of critical supply chains," said Albanese.

Tom Albanese, CEO at American Ocean Minerals

Why the Individual Guilt Framing Has Been Misleading

The discourse around AI's environmental impact has been dominated by individual-user guilt, with people worrying whether they should feel bad about asking Claude to summarize a PDF. This framing has been the most wrong part of the climate-AI debate for two years running, resembling how big corporations insist individuals be held responsible for not recycling while they slug down oil by the barrel full. The reality is that a single ChatGPT prompt raises your personal energy footprint by an unmeasurable amount.

The real conversation should focus on systemic infrastructure challenges: whether data centers can be powered sustainably, whether mineral supplies can be secured without geopolitical vulnerability, and whether the productivity gains from AI justify the environmental trade-offs. These are questions about policy, supply chains, and energy systems, not about individual user behavior.