Commodity Watch: AI is Thirsty

<div class="grid grid--33-66-col"><div class="col"><img loading="lazy" data-fr-image-pasted="true" src="/getContentAsset/e4db1c4c-2687-44cd-adbd-db1eb849e5d2/cb87803a-320c-480f-ab75-7b9029eaaf79/Ahmad-Al-Sati-new.jpg?language=en" alt="Ahmad Al Sati" title="Ahmad Al Sati" class="fr-fic fr-dii" style="width: 180px"></div><span style="font-size: 12px"><div class="col"><strong>AHMAD AL-SATI<br></strong>PORTFOLIO MANAGER<br><br><p>Ahmad is the President and portfolio manager for Gemcorp Capital Advisors LLC, based in New York.&nbsp;<br><br>Ahmad has spent most of his career in the global credit markets. Prior to Gemcorp, Ahmad was President of Pandion Mine Finance and RiverMet Resource Capital, LP - a fund focused on investing in precious metals, where he was responsible for managing the investments and the day-to-day operations of the registered investment adviser.&nbsp;</p></div></span></div><hr><p>For each one-hundred-word email, ChatGPT consumes 500ml of water – slightly more than a 16-ounce bottle. AI inferences results in similar water consumption patterns (see map below for number of inferences per 500ml). For reference, the average adult should consume 2-3 litres per day.</p><p><br></p><p>Water is required across the AI value chain. Purified water is required for chip manufacturing. It is also used in significant amounts to generate thermoelectric power from gas and coal (the U.S. national average water withdrawal, for example, is estimated to be 43.8 L/kWh). Data centres, the front of the chain, require water to manage humidity and heat. Data centres demand a lot of electricity (see point above re electricity) and generate significant heat. The servers require specific atmospheric conditions – temperatures between 64-80°F and humidity between 40-60% rH. Closed and open water loops keep the internal atmosphere constant. New water is needed, however, because the water evaporates or is discharged to eject salt and minerals which accumulate over time.</p><p><br></p><p>Meta used 22 million litres of water to train its open-source language models. Google used almost 20 billion litres of potable water for onsite cooling in 2022. In July 2022, one month before OpenAI completed training GPT-4, Microsoft used ~11.5 million gallons of water in its Des Moines data centres where GPT-4 was being trained. That water accounted for 6% of all the water used in the district. Demand for water will only increase as data centre water use increases. Globally, AI demand for water is expected to be about 6.6 billion cubic meters of water in 2027, which is equivalent to 6x Denmark’s municipal, industrial, and agricultural water use.</p><p><br></p><p>Yet, freshwater is finite and scarce. Water covers 70% of the earth but only 3% of the world’s water is freshwater and only 0.5% sits outside the polar caps. Water is not uniformly distributed and is increasingly impacted by droughts. Today, ~2.7 billion people suffer severe water scarcity at least one month of the year. AI will compete for water with humans.</p><p><br></p><p>Managing water resources could minimize water competition. Locating data centres in cooler, less humid climes will help minimize water use (as outside air is used to manage the heat). Using data centres more at night also helps manage heat and humidity.</p><p><br></p><p>These constraints mean managing resources in the physical world is critical. Unlike previous tech innovations that decoupled the physical from the digital (to a large extent), AI is deeply grounded in material resources. It requires silicon, water, electricity and batteries (among other materials and commodities) in substantial amounts. This reliance on physical goods could result in chokepoints and cost overruns that may affect the ultimate speed and scale of AI. Thus, securing the entire value chain of the requisite materials will be key to pushing the AI revolution forward.</p>