No, AI isn’t going to use all the world’s energy

The opinions stated here are my own, not those of my company.

We’re still in the AI age, which I thought would be full of awesome AI demos. There are a lot of them, and I’ve built my own. However, there seem to be even more bad takes on the future of AI based on not any real understanding of computers but on unreasonable speculation.

Take this op-ed in the Times from a few months ago, which I came across last week. In it, the author talks to economists about the future of work and jobs. The editorial started out okay but soon devolved into a strange point from Noah Smith:

He wrote that the future of human work hinges on whether A.I. is or isn’t allowed to consume all the energy that’s available. If it isn’t, “then humans will have some energy to consume, and then the logic of comparative advantage is in full effect.” He added: “From this line of reasoning we can see that if we want government to protect human jobs, we don’t need a thicket of job-specific regulations. All we need is ONE regulation — a limit on the fraction of energy that can go to data centers.”

People who believe that AI might consume all of the world’s energy do not have a rational understanding of AI, and it’s definitely important to set expectations correctly or we’ll end up fighting our own shadows. It doesn’t actually allow us to develop good regulations and it will cause us to miss out on the actual flaws of the technology.

Energy Distribution

In 2022, the world consumed over 170,000 TW of energy. Keep in mind that’s all energy and not just electricity.

The US generated 4,500 TWh of electricity last year.

Total energy consumed was nearly 26,000 TWh. Because total energy is a lot more than just the electricity coming from an outlet. Oil is a common energy input for vehicle travel and natural gas is used for heating in many places.

So even if AI data centers took up 100% of all electrical demand, it still wouldn’t use anywhere near all the energy.

Even within electricity itself, 38% of demand comes from the residential sector. But let’s look at the commercial sector. Even there, data centers make up a fraction of commercial electricity usage. In fact, even computers make up a fraction of data center usage. Much more electricity is used for HVAC, heating and cooling spaces, than the actual electricity needed to run computations.

Data centers currently use about 20 GW of electricity and that’s expected to grow to 30 GW by the end of the decade. But even that’s a small, small percentage of all the electricity in the US and the world.

Keep in mind that even the field of computing has a lot of extraneous energy usage beyond just running the computers. Servers require maintainers, and maintainers require food and comfortable temperatures. (And healthcare and food and homes and many other things.) While you can isolate data centers on a graph, you can’t truly disentangle them from the world.

Data centers would need to scale not by 50% but by 800,000x to get close to total energy. That’s simply unrealistic. It’s orders of magnitude away from realism. Ilya Stutskever says we’ll need to cover the planet with solar panels and data centers, but we’er far from that being necessary. Anyone who would claim AI is close to using up all energy is not a serious person.

Energy Transmission

If AI was going to use all energy, it would need to use every last watt in the world. It’s very easy for a data center in Ohio to pull electricity from the nearby Ohio power grid, but there’s no way to connect it to remote microgrids in Tanzania.

Let’s say we do that. The two areas are about 8,000 miles apart. You’d need to build transmission lines that go across nations, across continents, and even thousands of miles across the ocean. Even if you completed such an unnecessary mega-project, you’d waste a significant portion of that energy along the way.

It just doesn’t make sense to bother tracking down every bit of energy and trying to wire it all together to a singular geolocation.

Let’s say we could draw petawatts across the Pacific through reconducturing and that we could convert all the oil and natural gas into electricity and had enough transmission lines to draw that to one place.

There’d be no way to handle that much compute in a single place. The amount of generated heat would be immense. The electrical grid would be a gigantic knot with a single point of failure. There are physical limits that we can mask, but never eliminate. There’s no way to achieve infinite compute.

You might say obviously AGI doesn’t need to be in a single place. As it’s a digital entity, not a physical one, it can be distributed in data centers around the world.

We’d still have limitations to infinite compute. Even if we built a data center in rural Tanzania, how would we be able to employ it in a single planetary brain? We’d need a lot of fiber.

Here’s a cool map of all the subway cables around the world. We’ve laid out enough fiber for 20GW of global data centers and trying to build more but it’s a slow process that requires a physical build-out of the world.

Fiber is fast, literally the speed of light, but still it can't transmit data halfway around the world in a timely fashion. State-of-the-art is transmitting 301 Tbps, achieved in experiments just in the last few months.

But a highly advanced AGI is going to need to shuttle that much data around a lot faster. One website says the human brain has 2500 terabytes of memory. I have no idea if that’s true. But taking ten seconds to compute and send a human brain’s worth of data is not AGI. Superhuman intelligence would need more memory and be slower.

Shuttling all that data around is runs into physical limits. Can we compress the data? Sort of, but we still run into Shannon’s limit on how much data can be compressed over the wire. There are fundamental limitations to compute capacity, as Stephen Wolfram notes.

Barriers to Growth

When you think of all the potential ways things can go wrong, it’s more likely we won’t get AGI anytime soon. Even if you wanted AI to use all the energy in the US, you’d need the technology sector to grow by 800,000x.

That’s a kind of growth that is physically impossible and also politically impossible. We’re lucky to get 3% growth, and trying to get one more percent is actually quite difficult.

There would need to be a massive program of enormous scale to hire millions of people to construct new data centers, new transmission lines, new fiber, new CPUs, new GPUs, new storage, and every other little component that all needs to be created and shipped and coordinated in perfect synchronization. Even then, we may be able to get 6% growth.

In the data center power usage article above, estimates are that we’ll only get 50% growth in energy over the next decade. That’s a lot of growth and we’re already running into power constraints.

We need to create more power, a lot more. This is a serious constraint. Data centers are now seeking to become power producers. Amazon recently acquired a data center powered by nuclear energy.

But our ability to build more power also faces a lot of barriers. Nuclear energy has a lot of regulatory barriers that slow down development and deployment. Attempts to build clean energy transmission get blocked in a lawsuit by nominally environmental groups.

Paper cuts like that slow down everything. But it happens everywhere. Every town is going to have lawsuits against more power, more compute, more of whatever else is required to build a society devoted singularly to AGI. Partisans will reflexively be against it. They’d rather go to jail than turn off their light to give the AGI an extra 5 watts.

Beyond that, there are many other barriers. Remember the Internet boom of the 90s? The dot-com bubble happened and resulted in a graveyard of failed companies.

In the 2010s, the Internet of Things was poised to change the world. It happened. It’s still happening. I wouldn’t say it failed, but it definitely didn’t grow exponentially and change the world.

The future can be hard to predict. Is AI in a bubble today? Possibly. I guess we’ll only know in hindsight. But my point is that it’s a lot easier for things to go wrong than they are to keep moving exponentially forever.

Chip fabs could easily be disrupted by another pandemic, or an earthquake, or a war. In those cases, the chip shortage and lack of GPUs would make it difficult to make technological leaps. Any supply chain disruption could lead to another AI winter.

Energy Abundance

We need energy abundance. We need to figure out what nuanced barriers exist to building more solar, more housing, and more productivity. This is not easy. It’s not just about plugging data into a computer and getting back a perfect plan on paper. It requires political work to modify laws and get votes in your favor.

Electricity demand in the US is rising, and that’s a good thing. We’ve spent years talking about ending fossil fuels and switching to clean energy, but now it’s finally happening. Electrification is going to lead to a massive rise in electricity demand. Heat pumps, EVs, electric stoves, and more are all critical pieces of this journey.

This is already putting pressure on our grid and we are having serious issues in building more. The interconnect queue to add more power sources to the energy grid is full of clean energy projects waiting for their turn, waiting for the regulatory process to be completed. Permit reform is a critical step here.

Data centers are not going to grow unless we grow power in tandem. And that power needs to be clean.

More generally, we need to speed up processes and that means putting an end to the endless red tape put up by NIMBYs who use the law selfishly.

That also includes immigration reform. There are a lot of people who want to come to this country and get jobs, and we put up a lot of barriers to force them into the shadows. We should make it easier to come here, get a job, pay taxes, and innovate.

If we do all of that, do we get AGI? No. We get 6% growth instead of 3%. We get affordable housing in nice cities. We get a clean, livable environment. We get more tech companies with immigrant founders. AI will still be incremental, still centuries away from gobbling up all the energy in the world.

Conclusion

I wanted to write to push back against all the grifters trying to tell you that AI is something that is existentially dangerous and you need them as the whisperer to protect humanity.

Normies are right. You don’t need to worry about AI in schools. We won’t all be dead in fifteen years. It is not a digital species.

This is all very silly. Look, AI can be dangerous in the same way that a lawnmower is dangerous. That is to say, it genuinely is a dangerous tool. A lawnmower has a dead man’s switch so that this dangerous tool always needs a human in the loop to run. That’s a regulation that is specific and targeted and is reasonable. We don’t need to pretend a lawnmower is something that it’s not.