Last week I shared an impressive medical AI outcome – an OpenAI model performed better than ER doctors in making diagnostic and treatment recommendations. When every doctor has an AI partner to provide a second opinion and deep analysis, it's pretty much a sure thing that lives will be saved and patients will have a better experience.
Here at Innovating with AI, though, we work very hard not to be myopic AI boosters. We're optimists but realists too. We are excited for the future but cautious about the transition. We believe that bringing more voices into the AI conversation is a good thing – and that everyday people, not just "tech people," need a seat at the table as we plan out the future.
So even when we someday roll out an AI partner to every doctor and every teacher, the practical reality is that most AI usage is going to be for more mundane tasks. It is these mundane tasks that bring up the serious discussion of environmental trade-offs, and that's what I'll be covering in depth today.
Every AI prompt uses electricity and other resources, and those things have to come from somewhere. If you're talking about saving lives with AI in the emergency room, pretty much any expense is justifiable. But what about for white-collar office work? What about generating images and videos? The ethical waters are much muddier when the tasks can be perceived as non-essential.
That said, I don't think it would be better for the average human being if we "went back in time" and used less electricity. Life is measurably better today than it was 100 years ago, in large part because of advances like electricity, refrigeration and fertilizer (which helps us grow enough food to feed more people). So the question becomes not how to stop progress, but:
How do we fuel impressive new technologies without messing up the planet?
Clean energy and AI go hand in hand
If you solve the question of cheap electricity generation, the vast majority of problems go away.
Right now, natural gas is usually the quickest-turnaround way to generate new electricity in the United States. This is better than using coal, but still a net-negative for the environment and public health. Data centers themselves aren't inherently "dirty." They're basically just big buildings. They are only a problem if you burn fossil fuels to create electricity to run the computers inside.
At a certain threshold of electricity-cheapness, water problems also disappear. This is because desalination – taking the salt out of ocean water – is possible but electricity-intensive. But when you get electricity costs low enough, it makes sense and gives us something close to an unlimited water supply. (In Saudi Arabia, half of drinking water already comes from desalinated ocean water.)
In the long run, cheaper electricity solves both the fossil fuel pollution problem and the water scarcity problem. And it's accomplishable with power-generation technology that already exists.
Solar – the easy and obvious step
You will occasionally see claims that solar is bad because creating the panels is mineral-intensive, but the math clearly shows that solar is a better way to generate electricity than fossil fuels. We already see a ton of data center projects using solar to support their electricity needs. For example, Apple runs a data center in Denmark entirely on solar, wind and hydroelectric power.
Apple's data center in Viborg, Denmark runs on 100% renewable solar, wind and hydro power.
The drawback of solar and wind is that they only generate power when it's sunny or windy. This can be offset by battery storage, but that gets expensive and slows the buildout. Solar is extremely cheap in the daytime, but figuring out the battery cost and capacity question for nighttime is the next financial frontier.
So, if you can do an all-solar data center, to me it's a no-brainer. The challenge is that companies want to move faster and not wait for the battery-storage question to be figured out in a cost-effective way. So we end up with proposals like the Meta data center in Louisiana, which couples solar with natural gas. The proposed Meta data center would use something like 100x the power of the Apple one pictured above (50 megawatts for Apple, 5,000 to 7,000 megawatts for Meta), and it would mix one-third renewable energy with two-thirds natural gas power. That means building up to seven new gas power plants near the data center – so you can understand why folks in Louisiana are not happy about the idea.
This is a problem with any type of large facility, not just data centers. The cleanliness and pollution is mostly about how they generate power. With on-site gas, they spew carcinogens. If they could switch to clean energy they'd be something close to neutral for the environment.
But if you want to move fast, natural gas is one of the few viable ways to generate electricity during the off-hours for solar and wind. So the next step is to unlock clean, 24/7 energy from other sources.
Nuclear is safer than coal and as clean as solar
Obviously nuclear is scary. But just as plane travel is safer than car travel (but people are scared of flying and no one is scared of driving), there is a big reality vs. perception problem with actual nuclear power generation.
If you look at coal power in terms of the air pollution it generates, you get seemingly insane statistics like:
One coal power plant in Pennsylvania was responsible for 600 deaths per year until recently, when it installed better scrubbers and got that down to "just" 80 deaths per year
Natural gas is 10% as dangerous as coal, so that coal plant killing 80 people per year would kill ~8 annually if it were natural gas.
But look at nuclear, solar and wind. The danger goes down to 1% of gas and 0.1% of coal.
Mortality rate from pollution and accidents for each energy source. (Source)
When you look at the real numbers, nuclear is so safe that it is almost laughable to consider replacing it with natural gas, oil or coal. It clearly belongs in the "clean, safe energy" bucket where we happily categorize wind and solar.
A quick note on scale – the chart above is in terawatts, which means 1,000 gigawatts. In turn a gigawatt is 1,000 megawatts. If you assume the Meta facility uses 5 gigawatts from gas power 24 hours a day, that means it would use 43,800 gigawatt-hours, or 43 terawatt-hours, in a year. All electricity generation in the United States adds up to about 4,400 terawatt-hours per year.
So a maxed-out Meta data center using gas would use 43 terawatt-hours in a year, which would lead to ~121 deaths from pollution and accidents. If you switched that to nuclear, you're down to ~1 death from accidents and pollution, saving ~120 lives every year without changing anything else about the data center itself.
France is proof that nuclear makes sense. The country generates 66% of its electricity with nuclear. Fossil fuel electricity sources are so minuscule that you can barely see them at the left end of this chart.
France generates 66% of its electricity with nuclear and uses nearly zero fossil fuel for electricity. (Source)
Call me crazy, but I don't think of France as a dangerous nuclear wasteland. Instead, it's an admirable leader in clean electricity generation, and if the U.S. followed France's lead we could almost entirely solve the clean energy problem.
Most of the barriers to this come from the outdated assumption that nuclear is dangerous. In practice, a coal plant operating normally is roughly as dangerous as a nuclear power plant malfunctioning. Nuclear accidents are obviously more scary than natural gas accidents, but reality simply does not support shutting down existing nuclear plants or discouraging new ones. Sadly, both New York and Germany have shut down nuclear plants and ended up with a dirtier electricity mix as a result.
And then there's the newest nuclear tech, small modular reactors. These systems can produce 300 megawatts (enough for six of those Apple data centers pictured above) and are small enough to fit into nice-looking structures, like the concept below. (The vertical ones are about the height of two school buses stacked on top of each other, and about the width of two school buses parked next to each other.)
Small modular reactors can fit into structures like these and produce huge amounts of clean electricity. (Image Source)
Geothermal can complement solar in much of the western United States
Lastly, I'll touch on geothermal energy. While this is not feasible everywhere, on the west coast and in the mountain states, it is possible to generate clean electricity simply by drilling down into the ground and collecting heat from deep in the earth. Like nuclear, geothermal plants can operate 24/7 regardless of weather conditions, and have no nuclear-like risk of major accidents.
Coincidentally, the parts of the U.S. that have geothermal potential also are very sunny, so it is the perfect complement to solar. Here's a map from the Department of Energy showing favorable sites for geothermal plants:
Geothermal is basically the best of all worlds in large swathes of the country. (Map Source)
Again, we can see indisputable evidence of geothermal success in other countries. Iceland generates 27% of its electricity with geothermal and 73% from hydropower, which means 99.9% of its electricity for its population of 400,000 comes from clean renewables. In Kenya (population 58 million), 43% of electricity comes from geothermal and 90% of all electricity comes from renewables.
Kenya's electricity mix, with geothermal in the lead. (Source)
The data center buildout is an opportunity to push for clean, safe, cheap electricity
We've looked at the electricity generation mix for several other countries today. Now it's time to look at the United States:
Forty-two percent gas, 15 percent coal. All in all, more than 58 percent of American electricity comes from fossil fuels. That number is nearly zero in Kenya, Iceland and France.
So when we talk about data center pollution, we aren't really talking about something specific to AI or computing. Instead, we're acknowledging that anything we do in the United States that uses electricity is inherently a pollution-generating activity.
Stop and think about this for a second.
"Electricity = pollution" is such a base assumption for Americans that we rarely consider it could be any other way. But right now, at this very moment, you could go to Kenya, France or Iceland and use electricity without generating any pollution! Together, Kenya, France and Iceland have roughly the population of the five most populous U.S. states (CA, TX, FL, NY, PA) combined.
There's no physical barrier to the U.S. generating power the way other countries do – we're choosing to do otherwise for a variety of reasons, ranging from cost-optimization to nostalgia.
My view is that the desire for a rapid AI data center buildout is an opportunity to push companies to invest in clean energy in the very near future. Perhaps it is unreasonable for Louisiana to require Meta to build advanced nuclear and solar today, but clearly state and local governments could force Meta to commit to a 10-year clean-energy plan in exchange for using gas in the meantime.
I think this is where an optimistic but skeptical public approach really shines. If you simply say "no data centers allowed" because the current electricity-generation mix sucks, you basically remove yourself from the conversation, and the most likely result is that companies and political leaders move forward without your input.
If you instead reframe it as "How can we make the most of AI's potential while also doing it in a way that doesn't create pollution," there are many viable options. Changing the American electricity mix will not be frictionless, but it is clearly possible.