By 2030, AI needs India's power and America's drinking water

1. Resource gap, stalled projects — ▶ 1:18

   By 2030 AI needs the incremental power of the whole of India and the incremental drinking water of the whole United States, and Beck confirms that is per year and growing exponentially. A water gap already existed before AI. Today 60% of data center projects are delayed, mostly because natural resources are missing or communities push back. Those communities are "you and I": neighbors, families and employees who fear their water will be taken and their power bills will rise.

2. Closed loop, car-wash water — ▶ 3:59

   Ecolab's approach is to reuse water the way nature does, in a closed circuit. Today's data centers are big air-conditioned rooms cooled by water-hungry cooling towers. Direct-to-chip cooling brings a water-based product to each chip, cools it and returns it, so compute grows with no incremental water. Beck pitches politicians that a new data center will use no more water than a car wash, which "changes the discussion". Ecolab claims 103 years in water, 3 million locations and 40 industries.

3. Water used three times — ▶ 6:08

   Chip power and heat are almost directly linked: more compute performance needs more power and creates more heat, and unremoved heat melts the chip. Water enters at three points. Power generation is the second-largest water user in the world behind agriculture. Cooling towers use it to cool the chips. Chips are also manufactured in ultra pure water. Solving cooling alone therefore addresses only part of AI's water footprint.

4. 97% legacy cooling base — ▶ 8:49

   Roughly 10,000 data centers exist and one is being built about every week. Some 95-97% of them are air cooled, using lots of water and power and not always sited sensibly. Newer builds are liquid cooled, and the very newest use closed-loop circuits that consume no water. Beck frames it as good news and bad news: the latest generation fixes the problem, but almost everything built so far does not.

5. Recycled water goes bad — ▶ 12:58

   Circulating water is the easy part. Warm water standing for long periods grows biological contamination, scales like a coffee pot, corrodes and leaks, and any of these can stop a whole data center. It must be reconditioned on every return, and chip-cooling water must be purer than drinking water. Ecolab learned this from chip fabs, whose water is 1,000 to 1 million times purer than that in a drug. One sugar cube dropped into Lake Michigan would make the lake no longer ultra pure.

6. Heat equals nuclear plants — ▶ 16:16

   One gigawatt of compute produces roughly one gigawatt of heat, about one nuclear plant, so an 8-gigawatt buildout means eight. Options are venting it to air, district heating for homes and industrial processes, or generating electricity from it, which Beck calls the future. Data centers spend roughly 60% of power on compute and 40% on cooling. Cutting cooling to 10% would put 90% of power into compute.

7. One a week, chips every 10 months — ▶ 20:07

   Before Covid the world built roughly one data center a year. Now it is one a week: 50% in the US, 30% in China and 20% elsewhere. Chip generations arrive every 10 months, each with power multiplied by 10, and the interval is shrinking toward 6 months. Those chips get installed in older data centers, which must be retrofitted. Otherwise power and water needs "explode once again".

8. The moratorium worry is fair — ▶ 22:16

   New York is heading toward a moratorium on data center building. Beck says the worry is justified because recent builds ignored water supply and nearby natural reserves. He also notes that $900 billion goes into data center infrastructure in 2026 alone, and no city or state wants to miss that money. He calls today's debate "a little bit extreme" but expects communities and builders to find ways to make it work.

9. Warm versus cool chips — ▶ 25:09

   Nvidia and others are discussing running chips on warmer water, 30-50C, so less cooling is needed. Beck doubts it will stick, because chip performance rises as cooling improves. After 30 years in the business he says "humans have a tendency to go for performance first and to figure out the problems later". He bets cooler chips prevail, with Ecolab delivering that cooling using zero water through a closed circuit.

10. Power: 50 plants to 100 — ▶ 27:20

    World power demand was flat for 20 years and will now rise 5% a year. AI uses the equivalent of roughly 50 nuclear plants and will need 100 by 2030, but nuclear takes 5 to 20 years to build. The interim fix is on-site diesel or gas turbines, which Beck concedes is not the cleanest route. Minnesota's 80% carbon-free by 2030 target points to a cleaner model. China makes 90% of solar panels and wind turbines, and the West is starting from ground zero.

11. A 35-mile bubble — ▶ 33:13

    All the world's fresh water in one sphere would be 35 miles across, a bit bigger than Manhattan. Supply fell 56% short of needs before AI. Beck argues "there will be no AI if we can't reuse the water", and no data center without community consent. The failure path is the current one: forced communities, drained water, and rivals who care less about those questions getting the capacity first. He expects a messy quarter-by-quarter path but rising demand.

AI will not be limited by capital. It will be limited by water, power and permission. Christophe Beck's estimate is that by 2030 AI needs the incremental power of all of India and the incremental drinking water of the whole United States, per year, on a curve that is exponential rather than linear. The money is there: $900 billion is going into data center infrastructure in 2026 alone. The projects are not. Some 60% are delayed, mostly because the natural resources are missing or because communities push back.

Permission is the scarcer input, and Beck treats the reasons for its scarcity as legitimate. The people objecting are neighbors, families and employees. Their fear is simple: someone will take their water and their power bill will go up. The water gap is real before a single chip is plugged in. Beck imagines all the world's fresh water in one sphere. It would be 35 miles across, a bit bigger than Manhattan, and demand already exceeded supply by 56% before AI. New York is heading toward a moratorium. Beck says the worry is justified, though he finds the debate a little extreme. He concedes the industry earned some of it. Sites were bought first and checked later, with no real water nearby or a natural reserve next door.

His answer is to stop spending water on cooling. The typical hall today is a big air-conditioned room chilled by a cooling tower that consumes water. Direct-to-chip cooling replaces that. Thin tubes bring liquid to a small radiator on each chip. The liquid returns to a cooling distribution unit, is cleaned and brought back to temperature, and goes round again. The idea is 20 years old, first built for gaming GPUs that ran too hot. The result, Beck says, is no incremental water at all. New sites would use less than a car wash. He uses that comparison deliberately with politicians, because it changes what the argument is about.

That describes the newest buildings, not the stock. Roughly 10,000 data centers exist, and 95-97% of them are air cooled. Construction has gone from about one a year before Covid to about one a week, with 50% in the United States, 30% in China and 20% elsewhere. Chip generations now arrive every 10 months, each multiplying power by 10, and the gap is shrinking toward 6 months. Those chips get dropped into old buildings. Retrofit is therefore not optional. Without it, power and water needs explode again.

Closing the loop is harder than it sounds. Warm water that stands still behaves like a pond or a coffee pot. It turns green, scales, corrodes and leaks, and any of those can stop an entire data center. So the water must be reconditioned each time it returns, and it must be purer than drinking water. This is where Ecolab claims an edge. Chip fabs use the purest water on the planet, 1,000 to 1 million times cleaner than water in a drug, and Ecolab learned to manage it there. Beck's image is a sugar cube dropped into Lake Michigan: the lake is no longer ultra pure. The product is less the pipe than the chemistry and data that keep the loop clean.

Cooling is also only one of three places water enters the system. Power generation is the second-largest water user in the world after agriculture, and chips are made in ultra pure water. That is why power matters as much as water. A data center spends roughly 60% of its power on compute and 40% on cooling, and cooling does nothing for an AI application. Moving it to 10% sends 90% of the power to compute. The heat is a resource too. One gigawatt of compute throws off about one gigawatt of heat, roughly one nuclear plant, so an 8-gigawatt plan means eight. It can be vented to air, used for district heating, or turned back into electricity. Beck thinks the last is the future.

Supply is the harder problem. World power demand was flat for 20 years. It now grows 5% a year. AI uses the equivalent of roughly 50 nuclear plants and will need 100 by 2030. Nuclear takes 5 to 20 years to build, so it cannot close that gap. In the meantime, companies put diesel engines or gas turbines beside the data center and skip the grid. Beck admits that is not the cleanest route. It is performance first, in practice. The cleaner route is a state like Minnesota, targeting 80% carbon-free power by 2030, which would make data centers both water neutral and carbon neutral. But China makes 90% of solar panels and wind turbines, and the West is starting from zero. Conventional power fills the gap for a while.

Here is the objection to the whole argument. The efficiency story depends on a choice the industry has not made. Nvidia and others are discussing running chips on warmer water, 30-50C, so that less cooling is needed. Beck doubts it will stick. A cooler chip performs better, and after 30 years in the business he says "humans have a tendency to go for performance first and to figure out the problems later". He bets cooler chips win and that Ecolab can deliver them with zero water. That bet admits something. The same instinct that favors cooler chips also favors on-site gas turbines. The sustainable path is a forecast, not a trajectory.

It also matters who is forecasting. Beck runs a company that sells water treatment to data centers, and the conversation is sponsored by it. The India and US comparisons are his estimates. The car-wash line describes the newest builds, which are a small share of the stock.

His reply is that constraint will force the outcome. "There will be no AI if we can't reuse the water." There will be no power without more renewables, and no data center without communities in democratic countries saying yes. He does not expect a smooth line. Wall Street works quarter by quarter, so the next three months will be messier than the next 10 years. But he argues demand cannot fall. People will not stop using their phones, and companies will not give up the benefits. He points to Mayo Clinic expertise reaching anyone with a phone. He thinks that is worth solving the problems for.

The bad path, in his telling, is the current one: communities forced, water drained, and countries that care less about these questions getting the capacity first. That last point cuts both ways. For Beck, competition is the reason to do it right: build where the infrastructure fits, reuse water forever and lead. It is just as easily the reason to push ahead and settle with communities later. Which reading wins decides whether the 60% of delayed projects are a pause for better engineering or a sign that the buildout will be fought over place by place.