Last week, I gave a talk on “AI and the environment” to a few hundred young students in Europe. They were well versed on the risks of AI, but much less on its benefits. Here is what I learned.
First, talking about energy is tricky because most people don’t have good intuition about power. So here are a few interesting comparisons that can help anchor the debate: A single complex query to an LLM uses about as much energy as driving a car for six feet or two meters. A single full-size rack of high-end GPUs for training AI, uses about as much power as three gasoline-powered cars driving down the highway.
I love cars, but AI is probably doing more to help the environment than cars, watt-per-watt. For example, AI can accelerate the search for more energy-efficient chips, more effective solar cells, higher density batteries, and it can even accelerate the design of fusion reactors that will eventually deliver free energy.
Then there are other environmental benefits introduced by AI that are more indirect: For example, the use of pesticide has reduced by 70% due to the use of AI software that can recognize when and when not to spray, with positive downstream implication to chemical runoff and food quality. AI is being applied to develop new processes for carbon sequestration, water desalination, wildlife protection, fire prevention, PFAS detection, and to numerous other environmental challenges that have baffled human engineers for decades. The list is long.
Many were concerned about water usage of datacenters. However, modern datacenters don't need to consume any water at all.
When it comes to cooling, AI can be just as efficient as a car. Car engines are cooled using a closed water circuit: Water is pumped through the engine to cool it, then the hot water flows through a radiator where the heat is blown away with a simple fan, and the cycle repeats.
Similarly, a GPU rack can be cooled by flowing water to cool the processors, then the hot water is pumped through a radiator on the datacenter’s roof where the heat is blown away with a simple fan, and the cycle repeats.
This form of closed-loop “convection cooling” is efficient – it uses just a fraction (~3%) of the energy of the engine or rack, consumes no water, and requires no AC.
Many were concerned about energy supply. But notice this trend: Because of AI, suddenly, everyone is investing in nuclear, fusion, even space solar. There are thousands of new startups in this space. Unlike past booms in green energy research, this time, the interests are not driven by government subsidies, but are motivated by sound commercial viability.
So what do I recommend? Let’s move past emotional arguments against AI and the environment, to a rational discussion of specific environmental costs and benefits. Risks should not be ignored but every risk and cost, like energy, cooling, and materials, is an opportunity for new innovations.
Let’s not repeat the mistake our ancestors made in the 1970’s, when clean nuclear energy was shunned, only to leave us paying the price today.
First, talking about energy is tricky because most people don’t have good intuition about power. So here are a few interesting comparisons that can help anchor the debate: A single complex query to an LLM uses about as much energy as driving a car for six feet or two meters. A single full-size rack of high-end GPUs for training AI, uses about as much power as three gasoline-powered cars driving down the highway.
I love cars, but AI is probably doing more to help the environment than cars, watt-per-watt. For example, AI can accelerate the search for more energy-efficient chips, more effective solar cells, higher density batteries, and it can even accelerate the design of fusion reactors that will eventually deliver free energy.
Then there are other environmental benefits introduced by AI that are more indirect: For example, the use of pesticide has reduced by 70% due to the use of AI software that can recognize when and when not to spray, with positive downstream implication to chemical runoff and food quality. AI is being applied to develop new processes for carbon sequestration, water desalination, wildlife protection, fire prevention, PFAS detection, and to numerous other environmental challenges that have baffled human engineers for decades. The list is long.
Many were concerned about water usage of datacenters. However, modern datacenters don't need to consume any water at all.
When it comes to cooling, AI can be just as efficient as a car. Car engines are cooled using a closed water circuit: Water is pumped through the engine to cool it, then the hot water flows through a radiator where the heat is blown away with a simple fan, and the cycle repeats.
Similarly, a GPU rack can be cooled by flowing water to cool the processors, then the hot water is pumped through a radiator on the datacenter’s roof where the heat is blown away with a simple fan, and the cycle repeats.
This form of closed-loop “convection cooling” is efficient – it uses just a fraction (~3%) of the energy of the engine or rack, consumes no water, and requires no AC.
Many were concerned about energy supply. But notice this trend: Because of AI, suddenly, everyone is investing in nuclear, fusion, even space solar. There are thousands of new startups in this space. Unlike past booms in green energy research, this time, the interests are not driven by government subsidies, but are motivated by sound commercial viability.
So what do I recommend? Let’s move past emotional arguments against AI and the environment, to a rational discussion of specific environmental costs and benefits. Risks should not be ignored but every risk and cost, like energy, cooling, and materials, is an opportunity for new innovations.
Let’s not repeat the mistake our ancestors made in the 1970’s, when clean nuclear energy was shunned, only to leave us paying the price today.
