OpenAI CEO Sam Altman is pushing back on one of the recurring criticisms of AI infrastructure: that ChatGPT queries and AI data centers consume large amounts of water. According to Tom’s Hardware, Altman made the comments on the Sources Podcast with Alex Heath, where he argued that the water-use narrative around ChatGPT has been exaggerated and has not kept pace with changes in data-center cooling. The headline claim was a striking comparison: Altman said, from memory and with a caveat that the number might not be exact, that producing one almond in California uses about as much water as 38,000 ChatGPT queries. He framed that as “total true water accounting,” not simply water running through one facility, Tom’s Hardware reports. Altman also argued that the common image of thirsty AI data centers is outdated. He said older data centers used evaporative cooling, but claimed modern very large data centers no longer rely on that approach in the same way and use an amount of water comparable to an office building for ordinary plumbing needs such as sinks and toilets. He characterized the broader concern as a persistent internet narrative that, in his view, does not survive scrutiny. Tom’s Hardware’s article complicates that comparison. The outlet notes that researchers estimate it takes about 1.1 gallons of water to grow a single almond. It also says Altman’s own prior comments have put a ChatGPT response at roughly 1 to 50 milliliters of water, a range Tom’s Hardware says would equate to about 85 to 5,500 ChatGPT queries per almond depending on the cooling method — still far from the 38,000-query figure Altman cited in the podcast. The outlet also points out that facility age and cooling design matter. It cites estimates that training GPT-3 on older systems used nearly 185,000 gallons of water, and it notes that AI agents can consume more tokens than simple chatbot interactions. That does not directly refute Altman’s point about modern facilities, but it underscores that “water use per AI interaction” is not a single fixed number. Tom’s Hardware further reports that data-center projects in the United States have faced local water scrutiny. One site in Fayette County, Georgia, used 29 million gallons over 15 months, according to the article, while another development in Morgan County, Georgia, allegedly caused neighbors’ water to turn muddy. Those examples keep the issue grounded in local infrastructure, not only in abstract per-query estimates. The state of the evidence here is narrow: this cluster contains one reputable report, and Altman’s most eye-catching number was explicitly delivered from memory. What is clear is that OpenAI’s CEO is trying to reframe AI water use as a cooling-technology and facility-design issue rather than a simple per-query indictment. What remains unresolved from the provided material is which accounting method should be used across old facilities, new facilities, training runs, inference and agent-heavy workloads. Who benefits: AI infrastructure developers benefit if newer cooling systems can credibly reduce local water concerns. OpenAI also benefits if the public debate shifts from per-query comparisons to facility-specific engineering details. Who's exposed: Data-center operators remain exposed where local water use is measurable, contested or tied to community complaints. Companies making broad environmental claims are also exposed if their headline comparisons rely on assumptions that are not transparent.