What we know, what we don't, who holds the data, and what the water sector can build itself.
Olivier Mills · Baobab Tech
UNC WATER & HEALTH 2026 TOOLS, TRADE-OFFS & TAKEAWAYS
The wish, 2024
“I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.”
Joanna Maciejewska · 29 March 2024
Source: Maciejewska 2024, via WIST [34]
The promise, 2026
AI will cure disease.
Universal high income.
AI will solve the climate crisis.
"I'd rather bet on AI solving the problem." Eric Schmidt, October 2024
Trump, Nvidia's Jensen Huang and tech executives at a White House AI meeting, 29 September. Photo: Kent Nishimura/AFP via Getty Images
“Mostly murder, surveillance, and surge pricing for the McRib.”
“We can apparently capitalism our way to socialism.”
Jon Stewart · The Daily Show · 6 October 2026
Sources: The Register 2024 [35]; The Daily Show 2026 [36]
0%
of LLM use runs on models that disclose nothing on energy or water.
Model cardEnvironmental
Energy per query
Not disclosed
Water per query
Not disclosed
Training emissions
Not disclosed
Hardware
Not disclosed
Data-centre region
Not disclosed
Share of tokens on OpenRouter's top 20 models, May 2025
Source: Luccioni, Gamazaychikov, Alves da Costa & Strubell 2025, Misinformation by Omission [1]
A few facts.
Found by researchers and reporters. Rarely disclosed.
0.26 → 0.73mL
Water per Gemini prompt. Google counts cooling only; with power-station water it nearly triples.
Elsworth et al. 2025; Sharma et al. 2026 [2, 54]
0.26 vs 2.2mL
Cooling water per prompt: Gemini (Google, 2025) vs GPT-3 (academic estimate, 2023). 8× apart.
Elsworth et al. 2025; Li et al. 2023 [2, 3]
1.34×
One Meta site’s peak-day water demand against its town utility’s capacity.
Akinade et al. 2026 [4]
14×
Meta’s permitted water against its stated use, Richland Parish. Contracts often confidential.
CRS 2026 [5]
1,262
Projected US deaths from data-centre air pollution in 2028 under high growth ($20.9bn).
Han et al. 2026 [6]
1,450×
Energy for image generation against text classification.
Luccioni et al. 2024 [7]
17×
Energy to post-train a reasoning model against its plain instruction twin.
Morrison et al. 2026 [8]
6.7Mt
CO₂ from one Microsoft–XTO oil deal in 2025: 43% of Microsoft’s footprint, outside its Scope 3.
Roussilhe et al. 2026, preprint [9]
415→945TWh
Data-centre electricity, 2024 to 2030 (IEA base case).
IEA 2025 [10]
1/4
of data centres are in water-gap areas
34%
of AI capacity sits in high water stress
Rosa et al. 2026; IEA 2026 [11, 12]
96.6%
Air cooling cut one site’s water 44%, then moved 96.6% of the rest to power plants.
Alston et al. 2026 [13]
19×
Satellite NOx at xAI’s Southaven gas plant against what its permit implies.
Gauld et al. 2026 [14]
Full references in references.md, by number.
Enabled emissions
Data centres IEA estimate, 2025
AI in oil and gas added emissions
up to 13×
01×3×5×10×13×
0.47 to 1.8 Gt CO₂ a year. To break even, AI would have to help renewables 4 to 5 times more than it helps oil and gas.
Sources: Alpine, Alpine et al. 2026, npj Climate Action [15]; Inside Climate News 2026 [16]; Axios 2026 [17]. Scale: multiples of the IEA's data-centre estimate.
Do you know what you agreed to?
Terms of Service · Privacy Policy
These Terms govern your use of our websites, apps and services. By creating an account or using the Services, you agree to these Terms and our Privacy Policy.
By using the Services you grant us a worldwide licence to host, store and process your Content. To deliver and develop the Services, your conversations and files may be reviewed and used to improve our models unless you change your settings. You can delete individual conversations at any time. Content that has been used for training is retained for up to five years. Feedback you submit may be kept for longer. Features such as memory store details about you across sessions so that responses can draw on earlier conversations. Connected apps may share data with us under their own terms. We may update these terms; continued use means you accept the changes.
These Terms govern your use of our websites, apps and services. By creating an account or using the Services, you agree to these Terms and our Privacy Policy.
By using the Services you grant us a worldwide licence to host, store and process your Content. To deliver and develop the Services, your conversations and files may be reviewed and used to improve our models unless you change your settings. You can delete individual conversations at any time. Content that has been used for training is retained for up to five years. Feedback you submit may be kept for longer. Features such as memory store details about you across sessions so that responses can draw on earlier conversations. Connected apps may share data with us under their own terms. We may update these terms; continued use means you accept the changes.
Anthropic · August 2025
Consumer chats train models if you click Accept
Kept up to 5 yearswhen you agree; 30 days if you don't
Memory keeps what you said last month. ZDR, zero data retention, exists on commercial terms. Do your tech providers even ask for it?
Sources: Anthropic 2025 [18]; Tom's Guide 2025 [19]; OpenAI Memory FAQ [20]; Anthropic API data retention [21]. Terms text is a generic illustration.
Open data
Trained on, then sold back.
OpenStreetMap · non-profit sector & volunteers
Frontier model training
Harvested at scale
ChatGPT · Claude · Gemini
Subscriptions & enterprise lock-ins
$20+ / month
100,000 IP addresses scraping OpenStreetMap at once, January 2026.
Sources: heise 2026 [22]; Boettcher, State of the Map 2026 [23]
Open data in the age of AI
Two last miles, pulling against each other.
Last mile 1 · the opportunity
Getting data and analysis into the hands of the people who need it.
Last mile 2 · the governance gap
Keeping licences and attribution intact inside someone else's workflow.
"Simply providing unbounded anonymous access may turn out to be unsustainable."Earthmover, LinkedIn, October 2026
Source: Earthmover 2026 and reply thread (A. Logan, Cecil) [24]
Where AI is going
Use of AI
Horizontal
General purpose: broad and shallow. Chat, drafting, summaries.
VerticalDomain specific: narrow and deep. Built for one use case. KII coding · support triage · signal anomaly detection
DiagonalAI products now go wide and deep. connectors · skills · agents
Framework: Baobab Tech
Who is moving
Corporate vs Non-profit and Funders
Companies
0%
are scaling AI across the business, up from 38%. Nearly 90% use it somewhere.
Nonprofits and funders
0%
of nonprofits have a one- to two-year AI roadmap.
0%
of foundations offer grantees no support to adopt AI.
Sources: McKinsey, State of AI 2026 [25]; Bridgespan & NTEN 2026 [26]; Center for Effective Philanthropy 2025 [27]
Black October · 2026
Vulnerability found by AI. Exploited within the hour.
0%of websites with a known server language run PHP.
00:00
21 OCTTwo critical flaws disclosed, found with AI models
+45 MINTens of thousands of exploit attempts
EVENINGMillions of attempts; sites taken offline
CVE-2026-109618 · CVE-2026-109619 · undetected for over a decade
Sources: Fortune, 21 Oct 2026 [28]; Axios, 22 Oct 2026 [29]
When anyone can build anything.
Jev · 15 September 2026
Closed decision model, $40M seed round
What's a decision model?
Text and questions in
{
"text":"Floods cut the clinic road.",
"sector": ["health", "water", "food"],
"urgent":"yes or no",
"severity":"1 to 5"
}
→
Answers out
{Confidence
"sector":"water",91%
"urgent":"yes",97%
"severity":484%
}
250+
open copies within 2 weeks
LayaKevOpenJevLiquid d1GLiNER2.5-Decide+ 250 more
“Shovelware”
AI slop started with text. Now it's slop for data and software, and we're seeing it everywhere.
So we need trusted tech partners, developers and data layers.
Sources: Wikipedia, Jev [30]; Baobab Tech, decision-models-experiments, from Han Xiao's All about Jev [52]; Wikipedia, Shovelware [33]
Academic and Research sector
0%
of biomedical abstracts showed signs of LLM-assisted writing by December 2025, up from 31% in 2024. In full papers: 89%.
