Julius Olavarria

AI

1) AI Water Consumption up to 2050

Base44 vibecoding 

Attempt at accurate water consumption for the AI industry as it expands unto 2050. There are currently lots of estimates so I attempted to synthesize different benchmarks for each decade. There is not a single estimate out there that is typically used, so this work may be a novel synthesis as of July 18, 2026.

Findings:

AI water consumption is not noticeable unless there is a continued explosion in AI demand up until 2050. So if I were to input a custom growth rate starting at a standard high growth path (700 Billion Litres in 2025, 50% compounding growth) and then taper to 40% and 30% for the next two decades, AI would place as the second highest water usage on the Earth by 2050, behind agriculture. I recommend using the custom growth rate to see whether or not AI would destroy our environment — but even for official low and high estimates by published researchers, AI water use is high but negligible compared to other uses on Earth. 

Below I have included, after a few hours of edits, the prompts that put the pieces together: 

I need to change most of the data. In 2025 there was 4500 trillion litres used of water total (or 4.5 trillion m cubed) in the world. I want a disclaimer that says this is given by UNESCO / FAO and represents direct water usage and not total water footprint (e.g. rainwater absorbed by crops), or in other words withdrawal-only (blue water) accounting. From this we can do rough estimates, about 2520 Trillion liters is agriculture (56%); livestock factory farming is 630 Trillion liters (14%); other 1349 Trillion liters (30%), and then the AI figures below. We can expect this to grow with population levels at about 0.8% up until 2035, 0.6% from 2035 to 2045, and then 0.4% from 2045 - 2050.

For AI, "indirect water consumption is incredibly difficult to measure, and tech companies almost never disclose it;" The Rare Exception: Industry researchers note that Meta is one of the only companies to explicitly break down both. In its disclosures, Meta reported consuming "3 billion liters directly versus a staggering 72 billion liters indirectly." Continue to document this and add disclaimers: Since the scales are so low for AI compared to other water usage, we will include both direct and indirect effects of what it takes to power AI, for example, powering datacenters, whereas the if we were to calculate the indirect effects of water usage for the other groups we would end with an astronomically high number. Additionally, calculating this number would be extremely difficult and is probably impossible without much disagreement, added on to the idea that "withdrawal statistics are built to measure water directly extracted from a source because that's the number that matters for local water stress, aquifer depletion, and allocation conflicts." AI is electricity heavy as well, so calculating indirect costs of electricity, in which the industry depends on so much, makes sense as the industry scales. *This is a developing thought but it is the core rationale behind the numbers.

I want sources now for each and you can fit them in attractive places that would give this model more not only more credibility but better optics. I want you to put somewhere that I chose 300 ish billion and 700 ish billion from the best available high / low estimate data (see below), and the 9.1 T litre increase (high estimate - to 2030, 65% compounding growth) because that was the most aggressive estimate for AI expansion. I also want them to know that this is an estimate broadly that considers two poles, to which you can configure custom growth rate by decade.

I also want a disclaimer here: "The United Nations University (UNU-INWEH): Their global report details long-term impacts, estimating that AI's expanding footprint will match the basic domestic water needs of 1.3 billion people by 2030. This contributes to the high estimate. As well as here: Professor Shaolei Ren (UC Riverside): A leading pioneer in data center hydrology, his foundational papers mapped out the exact water costs per individual AI query (e.g., how a 50-question prompt session can evaporate 500ml of water). And one more: (Nuance) The Nuance: UN researchers warn of a "rebound effect." Even if software gets more efficient, the resulting drop in computing costs could cause an explosion in demand, ultimately driving total water consumption up at its highest possible year-over-year rate." Fit this info in convenient locations. 

312.5 - 765 billion litres AI water usage (note billions, not trillions, as given by the science journal Patterns, by researcher Alex de Vries-Gao - also add De Vries-Gao's methodology explicitly adds direct (on-site cooling) + indirect (water used to generate the electricity AI consumes) into one number. Indirect is the larger piece — his data puts electricity-generation water use at roughly 3–4x the direct cooling water, per kWh). Again, I want a disclaimer added here above the low, high, and custom growth rate buttons so the user knows where we are coming from. Also, this is in 2025 - so this is the start of the period in which we are measuring. 

So from there let's do high growth. Let's add another source - By 2030: The United Nations University (UNU-INWEH) calculates that if AI spikes to consume 40% of all data center electricity without strict cooling overrides, "AI-specific water consumption will hit 9.3 trillion liters annually." This is by 2030 - if it's high (starting at 765 billion litres). You can write this, disclaimer again: The percentage increase here would be, doing the math, roughly 65% COMPOUNDING growth up until 2030, and then to avoid absurd results (in which consumption would be at more water than Earth provides) we taper off to 15% from 2030 - 2040, and then 6% from 2040 - 2050. 

I want you to write the assumptions / rationalization: UNU-INWEH report — "Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints" (released June 3, 2026), as covered by the United Nations University news release, which states "global data-centre electricity use, estimated at 448 TWh in 2025, could reach 945 TWh by 2030." This is corroborated independently by PBS and KSL coverage of the same report. Checking the CAGR: (945/448)^(1/5) − 1 ≈ 16.1%/yr ; so match this with water and we get a safe assumption. However, this assumes 1:1 comparison with water to electricity; it's most likely close.

Why 6% for 2040–2050? It's reasonable because it's roughly half the prior decade's rate, consistent with a maturing industry where the "AI share of electricity" driver is already saturated (it can't double again — you're already near its ceiling by 2030–2040), leaving growth driven mainly by new compute/electricity demand rather than a AI / electricity share-shift.

So from here we can go low estimate, low growth rate (starts at 312.5 B litres) - add the source - By 2030: "The International Energy Agency (IEA) maps out a highly conservative baseline where optimized data centers restrict global AI and cloud water consumption to 1.2 trillion liters—nearly 8 times less than the UN's high-growth 2030 spike."

This would suggest a compounding level (300 to 1.2 ) Growth rate, 300B → 1.2T (2025–2030):

Ratio = 1,200/300 = 4x. CAGR = 4^(1/5) − 1 ≈ 32%/yr — roughly half the 65%/yr from the high scenario, which makes sense as a low bookend. Write out this math as an assumption of course. 

ScienceDirect/EarthArXiv - the paper "Sustainable AI infrastructure: A scenario-based forecast of water footprint under uncertainty" rationale (write this in source): this is a neutral paper that gives us actual predicted benchmarks. So combined with the low estimate (300B to 1.2 T), this makes sense to use given its prestige. From 2030 - 2040 we will see 7.5% (which sits at the middle of their two predicted paths: moderate intervention at 5% and business-as-usual at 10%) and 2040 - 2050 at 3% per year, given things taper off. It's helpful to note that this paper predicts three paths. Here I made the heuristic that two of the paths are most true for the future: Business as usual and moderate intervention. Sustainable intervention would probably require utopian imagining and resources, and connectedness that the world does not have. Write this assumption.

2) ESG Rankings: What's are the best startups?

Claude ranked top startups assuming AI goes well. These could be good ESG-esque investments for businesses as of August 18, 2026. Some of these are applicable to most/all businesses, and all of these are relevant for VC firms or businesses with VC arms. It will also be good to see how this plays out in the future. 

"Rank them via moral weights — how much they can provide to people in the economy (economic value) for people, sustainability (their environmental footprint, water usage, etc.), ability for misuse (e.g., surveillance technology), and ability for benefit (e.g., how well it delivers on important, tractable, and neglected social problems)." 

How do we give an objective metric? We have a metric for the economy but that does not tell us how good the company is or will be. Startups in the AI business right now do not generate returns and many are heavily in debt. So we must pivot to funding stages in which anything below series-C makes sense (let's say under $50M, although the ones I have included here are well below this) because that is where these companies are most likely to (1) have a following/reporting with references that are citable and (2) compose of diverse stakeholders in which many business investors (and perhaps individuals) can join.

We should go off of companies that are currently integrating with AI. So startups in the tech industry. Why? AI is probably the future. And theoretical applications can be tied more closely to yes/no metrics: "Does the customer base include defense/intelligence contracts, is facial recognition or biometric tracking a core product line, is the company on the Commerce Department Entity List, does it have a published Responsible Scaling Policy with any teeth?"

What does assuming AI goes well mean? Assuming AI is not misaligned and kills all of humanity? Sure. But more importantly, we should generally orient this ranking toward a world where AI starts generating sufficient returns by 2030. Where are we now and is this feasible? Let's see:

"As of mid-2026: worldwide AI spending is projected to reach $2.59 trillion in 2026, yet only 6% of organizations qualify as AI 'high performers' with measurable bottom-line impact, per McKinsey's survey of nearly 2,000 companies."

Gwkinvest frames 2026–2030 as "the crucial test for AI commercialization," where at that time investments must start generating returns or "face potential write-downs."

It seems AI companies themselves have no plan on how they manage to accomplish this, unfortunately. Let's put 2030 down as an optimistic threshold, but one that AI companies must hit to save the industry[?] and beat these estimates. 

I am going to rank 8 of the best funding-stage startups based on (1) how much they can provide to people in the economy (economic value) for people, (2) sustainability (their environmental footprint, water usage, etc.), (3) ability for misuse (e.g., surveillance technology), and (4) ability for benefit (e.g., how well it delivers on important, tractable, and neglected social problems).

How do we narrow down the field so we do not get AI slop on our search results for general startups that would take years to individually evaluate? For that I should make assumptions that you should largely agree with to narrow down the field, and for some contestable ones I have clarified in parenthesis. I should also note that some metrics may have repeating justifications and/or a "low" ranking in cases in which there is no connection to the field. For example, "low" economic value for biosecurity (the very first ranking) does not just mean economic value will be indeterminate, but that there is no obvious connection between improving biosecurity and simultaneously generating economic value over and above potentially saving humanity. This would also be double counting the "human benefit" metric. Finally, in this context is important to note we are talking about small companies in an early funding stage and (likely) small ESG business investments. 

What is the safest field here based on these assumptions? It seems, surprisingly, agricultural AI and climate disaster modeling investments would be, based on these weights, the best fields for companies to make ESG investments. 

So from there we can compile our list (given by Claude and edited/fact-checked by me): 

Agricultural AI 

Climate/disaster modeling AI 

https://score-impact-ai.base44.app [note that the app includes the justifications explained above, and also included numerical rankings for the companies on top of the work that I did —  I think that all 8 recommended startups deliver smart ESG investments for businesses, so take these rankings as unofficial.]

This is a heuristic assuming businesses will check relevant details before making decisions. All sources are public.