Making AI actually work day to day is becoming its own job.

On July 16, hear from three people doing it: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. They'll share what the role really looks like, how it came to exist, the skills worth hiring for, and the challenges they're tackling right now. Bring your questions, since the best moments happen live.

Register to save your spot.

Reserve my spot

Hi {{first_name}}

Every time I ask an AI a question, somewhere a data center gets a little warmer and a little water disappears. How much water is one of the most contested numbers in tech right now: the companies say a few drops, the researchers say it depends what you count, and almost nobody publishes their books.

Craig

How much water does one AI query use? The honest answer is that it depends who is counting and what they count. Google, in the most detailed first-party disclosure so far, puts the median Gemini text prompt at about 0.26 milliliters of on-site water, roughly five drops. Sam Altman has offered a similar figure for ChatGPT, about 0.3 milliliters. If those numbers were the whole story, this would be a short article.

They are not the whole story, because of scope. Those figures count only the water evaporated cooling the company's own servers, not the water consumed generating the electricity that runs them. Mistral, the only other lab publishing first-party numbers, reports about 45 milliliters for a 400-token reply once indirect water is included, more than 150 times Google's on-site figure. A recent expert assessment of AI's environmental risks collects independent estimates showing a ten-page report can consume anywhere from 0.7 liters to 60 liters depending on the model and the infrastructure behind it. Same task, hundred-fold spread, all in how you draw the boundary.

The study that put this on the map is "Making AI Less Thirsty" from Shaolei Ren's team at UC Riverside, now published in Communications of the ACM. Its headline findings: training GPT-3 directly evaporated an estimated 700,000 liters of clean freshwater (about 5.4 million liters counting electricity), and global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal by 2027, more than four times what Denmark uses in a year. Training numbers for today's frontier models have never been published by anyone.

Here is the part that makes this a genuinely interesting story rather than a scare story: the science has been correcting itself in public. Ren now says the viral "bottle of water per prompt" figure overshot, and that a GPT-4 prompt likely used closer to 15 milliliters all-in, while models have gotten more efficient since. The "10 gallons per AI image" claim circulating online has no study behind it at all; working from measured energy data, an image runs closer to 15 to 60 milliliters. Getting the numbers right cuts both ways.

Why it still matters even at the lower numbers: totals and location. Google's US data centers alone consumed an estimated 12.7 billion liters in 2021, and the build-out has accelerated enormously since. Per-query efficiency does not help a town whose aquifer hosts the data center, which is why the local backlash stories we have covered in past issues, from Virginia to Visakhapatnam, keep coming. Water is the most local resource there is; averages hide the places that pay.

Now the other side of the ledger, because it is real. AI systems are doing some of the best water work in the world right now. Google's flood forecasting provides warnings up to five days ahead across more than 80 countries, in a sector where flooding causes an estimated $50 billion in damages a year, and machine learning is improving irrigation, leak detection in aging pipes, and the grid balancing that reduces water-hungry thermal generation, all documented in a good npj Climate Action review. Ren's own research points to fixes that cost almost nothing: schedule training runs for cooler hours and cooler places. As he puts it, we don't water our lawns at noon; we shouldn't train AI models at the hottest hour either.

What we still don't know is the biggest finding of all. Two companies publish first-party numbers. There is no shared standard for what counts, on-site versus lifecycle, withdrawal versus consumption, annual average versus drought-season peak. The EU now requires data centers to report water use, and the UN has proposed an AI environmental transparency initiative, but for most of the industry the honest answer to "how much water does your model use" remains: they haven't said.

My read: the drops-versus-bottles fight is mostly a distraction. The real question is why, three years into the AI boom, disclosure is still voluntary, unstandardized, and rare, and the practical move for anyone in our field is to keep asking for the numbers, in procurement, in policy, and in public.

Research Engineer, Rule of Law
Anthropic · San Francisco or New York
Build the technical foundations for keeping AI systems accountable to law, at the research lab treating the rule of law as an engineering problem as well as a legal one.

Research funding: Scaling AI Safety for a Multi-Agent World
Cooperative AI Foundation, with Schmidt Sciences, Google DeepMind, and ARIA · grants up to $1,000,000 · closes August 8
A $10 million fund for research on what happens when AI agents interact with each other, one of the least understood safety questions of the next few years. Open to researchers worldwide.

Innovation Professorships (Associate Professor / Professor)
Aarhus University · Denmark · closes September 10
Multiple senior posts for researchers who want to commercialize science for the public good, with themes including water technology and catchment-scale water challenges. Directly on today's theme.

Team Lead, Socioeconomics and Just Energy Transitions
IRENA · Abu Dhabi · P4, roughly $86,000 to $97,000 net · closes August 5
Lead the research team analyzing how the energy transition lands on jobs, welfare, and equity, at the intergovernmental agency for renewable energy.

Director of Product, Gaming
Mozilla · Remote (US or Canada) · $205,000 to $351,000 depending on level and location
Build new products at the company owned by a nonprofit foundation, where the mandate is an internet that serves people first.

Research & Methods Lead, Partnership for Battery Action
Global Development Incubator · Remote · closes July 17
Shape the evidence base for cleaning up battery supply chains, from cobalt mines to recycling, at a partnership working across the Global South.

Senior Digital Safety Research Associate
Family Online Safety Institute · Washington, DC (hybrid) · $75,000
Research how families actually experience online risk, from AI companions to age verification, and turn it into guidance parents can use.

Corporate Sustainability & Climate Change Consultant
ERM · Lima, Peru
Advise companies on climate strategy and disclosure at the world's largest pure-play sustainability consultancy, from its Lima office.

PhD in Sustainability Science, Finance and the Energy Transition
Lund University · Lund, Sweden · closes August 28
A funded doctorate on how financial systems speed up or slow down the energy transition, at one of Europe's leading sustainability research centers.

Three roles at ABM: Fundraising Specialist, Governance Fellow, Operations Fellow
ABM · Remote, worldwide
Help build the backbone of a Global South-led movement using art, data, and technology for justice. Three openings across fundraising, governance, and operations.

Manager, Cybersecurity and Infrastructure
Oxfam America · Boston, US
Protect the systems behind one of the world's best-known anti-poverty organizations, where security failures have humanitarian consequences.

Governance, Risk, and Compliance Analyst
You.com · San Francisco (hybrid)
Build the compliance backbone at an AI search company, exactly the kind of unglamorous governance work today's feature argues the industry needs more of.

Consultancy: Digital Dashboard for Nature-Based Solutions
IISD · Johannesburg-focused, remote-friendly
Build the minimum viable product that shows a city what its nature-based water and climate solutions are actually delivering. Data visualization in service of infrastructure decisions.

Digital Systems Capacity Building Consultant
Resolve to Save Lives · Addis Ababa, Ethiopia
Strengthen the digital systems Ethiopia's public health institutions rely on for disease surveillance and response.

On July 16th at 1PM ET, beehiiv is going live with a look at the future of publishing, audience growth, and digital business.

What started as a newsletter platform has evolved into something much bigger: a place where creators and brands can grow, monetize, and own their audiences without stitching together half the internet to make it work.

The next chapter starts live at the Summer Release Event. 

Join us to see what’s coming next.

RSVP now.

Impact joke of the day. We asked our AI how much water it drinks. It said "that depends on your methodology," which is also what my uncle says about his diet.

Three reads that go deeper on today's theme. Links go to the original reporting.

"The global water cycle is out of balance" (Science Arena). A substantive interview with Shaolei Ren on blue water, why AI's cooling draw competes directly with human use, and why local communities need a seat at the table when data centers come to town.

How the "bottle of water per prompt" number was born (Andy Masley). A careful detective story about a viral statistic, with the original researcher's own correction. A great case study in how environmental numbers get made, spread, and fixed.

Green and intelligent: AI's role in the climate transition (npj Climate Action). The peer-reviewed overview of where AI genuinely helps, from flood early warning in 80+ countries to smarter grids, and what it takes to keep the benefits ahead of the footprint.

Jobs, jobs, jobs. The best place to start is our own AI for Impact Career Platform, tracking roles across the AI-and-impact world, and improving every week with your feedback. The PCDN job board carries 1,100+ live social-impact roles beyond it, refreshed daily.

Person to follow: Shaolei Ren, the UC Riverside professor whose research put AI's water footprint on the world's agenda, who has advised a UN panel on AI's environmental impacts, and who is refreshingly willing to correct his own numbers in public. The person behind most of what you just read.

A role we missed, feedback on the career platform, or your own take on the water question: reply to this email, or use the form below. A human reads every message.

Know someone who thinks about AI's real-world costs? Forward this along. They can subscribe at impactai.beehiiv.com.