Sam Altman, the chief executive of OpenAI, took to social media this week to push back forcefully against claims that a single ChatGPT query consumes an entire bottle of water — a statistic that has circulated widely in environmental debates about artificial intelligence. But even as Altman disputed the water figures, he acknowledged what many in the industry have quietly accepted: AI’s energy appetite is enormous, growing, and demands serious attention.
The exchange, which played out on X (formerly Twitter), has reignited a broader conversation about the environmental footprint of generative AI systems at a moment when the technology is being deployed at unprecedented scale across industries. With data center construction booming and power grids straining under new demand, the question of what AI actually costs the planet — in water, electricity, and carbon — has moved from academic circles to boardrooms and regulatory bodies.
The Water Claim That Sparked a CEO’s Rebuttal
The controversy centers on a widely cited figure suggesting that each ChatGPT conversation uses roughly 500 milliliters of water — about one standard bottle. This claim traces back to research from the University of California, Riverside, published in 2023, which estimated the water footprint of large language model training and inference. The figure gained traction in media coverage and environmental advocacy, becoming a shorthand for AI’s hidden resource consumption.
As reported by TechRadar, Altman called these claims “completely untrue” in a post on X. He did not provide specific counter-figures or point to OpenAI’s own water consumption data, but he was emphatic in his denial. The statement came in response to a user who repeated the water-per-query statistic, and Altman’s reply was blunt: the numbers being shared publicly do not reflect reality.
What the Research Actually Says — and What It Doesn’t
The UC Riverside study, authored by Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren, attempted to quantify the water consumption associated with AI model training and inference across different data center locations and cooling systems. The researchers estimated that training GPT-3 alone consumed approximately 700,000 liters of fresh water. For inference — the process of generating individual responses — they estimated water usage that, when averaged across queries, approached the 500-milliliter figure that has since gone viral.
However, the researchers themselves noted significant caveats. Water consumption varies dramatically depending on data center location, the type of cooling technology employed, the time of year, and ambient temperature. A data center in a cool northern climate using air-based cooling will consume far less water than one in a hot, arid region relying on evaporative cooling systems. The 500-milliliter estimate was an average that encompassed wide variation, and it included both direct water use (for cooling) and indirect water use (associated with electricity generation at power plants). Critics of the viral statistic argue that conflating these two categories overstates the direct environmental impact of any single query.
Altman Concedes the Energy Problem Is Real
What made Altman’s response notable was not just his denial of the water claims but his candid admission about energy. According to TechRadar, the OpenAI CEO acknowledged that AI’s energy consumption is a legitimate concern and one the company takes seriously. This admission aligns with a growing consensus among technology leaders that the power demands of training and running large AI models represent a genuine infrastructure challenge.
OpenAI’s own operations have expanded rapidly. ChatGPT now serves hundreds of millions of users, and the computational resources required to handle that traffic are staggering. Each query to a large language model requires significantly more processing power — and therefore more electricity — than a traditional web search. Goldman Sachs estimated in a 2024 research note that a single ChatGPT query uses roughly ten times the electricity of a Google search. The International Energy Agency has projected that data center electricity consumption could double by 2026, driven in large part by AI workloads.
The Data Center Boom and Its Discontents
The environmental debate around AI cannot be separated from the massive data center buildout currently underway across the United States and globally. Microsoft, Google, Amazon, and Meta have all announced multi-billion-dollar expansions of their data center infrastructure, much of it explicitly tied to AI capabilities. Microsoft alone has committed to spending more than $80 billion on AI-capable data centers in its current fiscal year.
These facilities require enormous quantities of electricity and, in many cases, water for cooling. In communities where data centers are being built or expanded, local residents and officials have raised concerns about strain on power grids and municipal water supplies. In parts of Oregon, Virginia, and Arizona, data center water consumption has become a political issue. Google’s 2024 environmental report revealed that the company’s water consumption rose 17% year-over-year, a trend the company attributed partly to AI-related computing demands.
OpenAI’s Transparency Gap
One of the persistent criticisms leveled at OpenAI — and one that Altman’s social media rebuttal did little to address — is the company’s lack of detailed public reporting on its environmental footprint. Unlike Google, Microsoft, and Meta, which publish annual sustainability reports with specific data on energy consumption, water use, and carbon emissions, OpenAI has not released comparable disclosures.
This opacity makes it difficult to independently verify Altman’s claim that the water statistics are untrue. Without OpenAI providing its own figures — broken down by data center location, cooling method, and workload type — outside researchers and journalists are left to rely on estimates and extrapolations. The UC Riverside researchers used publicly available information about data center operations and Microsoft’s infrastructure (OpenAI runs primarily on Microsoft Azure) to build their models. If those models are flawed, as Altman suggests, the most effective rebuttal would be data, not denial.
Industry Peers Are Investing Heavily in Mitigation
Other major AI companies have taken a more proactive approach to addressing environmental concerns. Microsoft has pledged to be carbon negative by 2030 and water positive by the same year, meaning it intends to replenish more water than it consumes. Google has committed to operating on 24/7 carbon-free energy across all its data centers by 2030. Both companies have invested in advanced cooling technologies, including liquid cooling systems that reduce or eliminate the need for water-based evaporative cooling.
Amazon Web Services, which operates one of the world’s largest cloud infrastructure networks, has similarly invested in renewable energy procurement and water recycling systems. These commitments reflect an industry-wide recognition that the environmental costs of AI infrastructure are not merely a public relations problem but a material business risk, particularly as regulators in the European Union and elsewhere begin to consider mandatory environmental reporting for data center operators.
The Broader Question: Can AI Growth and Environmental Responsibility Coexist?
Altman has previously expressed optimism that AI itself could help solve energy and environmental challenges — for instance, by accelerating the development of nuclear fusion or improving the efficiency of power grids. OpenAI has invested in Helion Energy, a fusion startup, and Altman has spoken publicly about the need for abundant, cheap, clean energy to power the AI future he envisions.
But critics argue that these long-term technological bets should not distract from the near-term reality: AI systems are consuming vast and growing quantities of electricity and water today, and the infrastructure being built to support them will operate for decades. The Uptime Institute, a data center research organization, has noted that the average data center has a lifespan of 20 to 25 years, meaning facilities built today will shape energy and water consumption patterns well into the 2040s.
The tension between AI’s promise and its resource demands is unlikely to resolve itself quickly. What Altman’s exchange this week made clear is that even the leaders of the most prominent AI companies recognize the stakes. Whether they are willing to match that recognition with transparency and concrete action remains an open question — one that investors, regulators, and the public are increasingly unwilling to leave unanswered.
For now, the debate over a bottle of water has become a proxy for something much larger: the terms on which society will permit the continued, rapid expansion of artificial intelligence, and who will bear the environmental costs of that expansion.