Why We Don't Need More Data Centers To Build Better AI

Sending data centers into space. Building them at the bottom of the ocean. Converting California fairgrounds - and San Francisco's Cow Palace - into massive buildings filled with servers..

The explosive growth of AI technology is fueling massive demand for data centers, spurring some seemingly outlandish ideas about where to put these new facilities. At the same time, growing concern about the environmental costs of data centers - which gobble up large amounts of energy and water - is prompting some to swear off AI altogether.

But taking the time to build AI smarter, rather than just bigger, could allow us to reap the scientific and social benefits of this technology with a smaller environmental footprint, argue two UC Berkeley experts in a commentary published online today (Tuesday, Aug. 11) in the journal Nature. The authors include Berkeley Professors Carl Boettiger and Fernando Pérez, both leaders at the Eric and Wendy Schmidt Center for Data Science & Environment, along with Cassie Buhler, a postdoctoral fellow at the Cooperative Institute for Research in Environmental Sciences (CIRES) at CU Boulder.

In an interview with UC Berkeley News, Boettiger and Pérez describe how, for many applications, smaller and more efficient open-source large language models are rapidly catching up to the capabilities of the large enterprise models like ChatGPT, Claude and Gemini. They believe that, by embracing these open source models, scientists and the public have the opportunity to create AI technologies that are better suited to their needs while providing alternatives to large, privately-owned platforms.

The news is filled with stories about the environmental impacts of data centers, from the noise they generate to the large amount of energy and water they consume. Which of their environmental costs do you find the most concerning?

Carl Boettiger: Data centers don't comprise a huge amount of the energy footprint of the planet compared to things like air conditioning or other big electrical uses, but they are the most concentrated energy demands we've ever created. And because the energy grid is very localized, they spike the power rates in the communities where they are built. In addition, they are too often located in communities that can least afford it, so these impacts are multiplied.

The story we are being told is that great intelligence requires great power, and that this is like a law of physics - it can't be done any other way. The truth is that AI can be much more energy efficient, but for economics and business reasons we are building them in this very energy intensive way.

Why are companies taking this power-intensive approach when there are more energy-efficient approaches available?

Boettiger: Computing has always followed a pattern where new algorithms are relatively inefficient, and as they get better and better they run on lower and lower power. Similarly, when you create a new chip or a new piece of hardware, initially you make the data center version that needs lots of power, and eventually it gets miniaturized. We've seen this a million times: Computers have gone from filling up a whole room to fitting inside the phone I'm holding in my hand.

Right now, AI development has become an economic race. If you're a company trying to build market share, you don't want to be selling a model that's a year behind. That's immensely far in AI terms. And so they are racing to the frontier by building bigger models that can be run only on that energy-demanding architecture rather than waiting until it gets more energy efficient.

They also want the models to think quickly. Our intuition is that bigger means smarter, but today, bigger often means faster. A user might not want to wait for the model to take 30 minutes to write code - it has to be done in 30 seconds or they'll switch providers.

Fernando Pérez: Building very large models that run in centralized locations also allows them to collect everyone's data. If you build a high efficiency, high throughput and low latency model, and get all the users all under one roof, then not only do you own all their data, but you can use that data to continue improving your models.

Could you walk me through some of the alternatives to these energy-intensive frontier AI models? How do they compare in terms of performance?

Boettiger: Open models have been around since day one, but until recently only academics and researchers knew about them. Initially, they required large servers to run - servers that might be the size of a fridge, sit in a rack, and have a bunch of GPUs. The performance of these models was also considerably behind, so they were mostly considered research curiosities.

In 2023, Facebook's open source model Llama accidentally leaked, and for a long time, that was the lead model. It was two weeks before someone had it running on a laptop. The open source community just chewed through challenges that companies like Google thought were insurmountable.

There was another big wake-up moment in January 2025 when the Chinese model DeepSeek came out, which was open source and trained on much less energy. It caused a momentary panic on Wall Street. There was a huge multi-billion dollar blip in the NVIDIA stock price.

Depending on the exact benchmark and the exact task, open models have maintained a pretty steady pace roughly six months behind the frontier models, and just in lockstep. And at the same time, we've seen the resources you need to use those open models get smaller and smaller.

When do you think open source models will be able to do useful work?

Boettiger: I think we're basically at the transition point where the technologies that are six months behind the frontier can do useful work, and in another six months they will be able to run on common hardware. That's why major laptop providers are starting to ship the new NVIDIA RTX Spark chips - they know that it will soon be possible to run tasks on these devices without calling out to a data center. And this is coming just in time, too, as data centers are going to get harder and harder to construct - the political will is starting to change and there's so much opposition.

Pérez: The massive frontier models are also reaching the point of being complete overkill for many, many real-world tasks. Most people don't need a Formula One car that costs millions of dollars and that requires an army of people to operate when they could buy a perfectly reliable Toyota Corolla. We need to remind people that there's a lot they can do with these alternative technologies that could be in many ways more beneficial to them.

Berkeley has long been a leader in the advancement of open-source technology. What are the other benefits of embracing open-source platforms, particularly in scientific research?

Pérez: I believe it is important that scientists actually have control of their scientific instruments. Scientists need tools that they can take apart, reassemble, reinvent and reimagine to suit their expertise and their needs. We can't do that if a company is providing the toolkit and you can't touch the parts because they're not yours. I'm not particularly interested in a future where the future of science is paying a subscription to any one company, no matter how good their tools are.

It's ironic because what these AI companies are building would be impossible were it not for an entire ecosystem of openly available tools, infrastructure, data and software that has been built up for the last 35 years. The current AI revolution would not have happened if everybody was clicking around in Microsoft Excel using Windows 95.

There often seems to be a tension between the desire to use AI to help solve issues related to biodiversity and climate change and the fear that AI is driving climate change by burning energy and natural resources. How do you reconcile this tension in your work?

Boettiger: AI has always been a bit of a floating signifier that refers to the cutting edge of the technology, and increasingly refers to massive models running in data centers. I think in some way, that's become a disservice. It's as if "transportation" was our only word, and we couldn't distinguish between a bike and an airplane when talking about carbon footprints.

We're seeing environmental scientists saying that they won't touch AI because of its environmental impacts. But our call to action is to learn to use the lower-footprint tools. It's not the physics that's blocking you, it's the business models. This is a chance for us to re-embrace open science and open source technologies to create a greener future.

This interview has been edited for length and clarity.