Top tech companies turn to hydrogen and nuclear energy for AI data centers

Yuval Bachar knows data centers. He's worked on them for Meta, Microsoft and Cisco, but now, his startup is looking to help Silicon Valley run data centers with lower carbon dioxide emissions.

ECL, Bachar's startup, builds hydrogen-powered data centers.

Hydrogen is a novel energy source for data centers that is more eco-friendly, and more importantly for tech companies that need to quickly expand their infrastructure, they can be placed into service in half the time that it takes to construct data centers that connect to the grid, Bachar said.

There's one of these hydrogen-powered data centers, with a measly 1-megawatt capacity, next to ECL's headquarters in Mountain View, California. Twice a month, a diesel truck hauls in hydrogen in a tank from Southern California or northern Nevada. The hydrogen mainly derives from natural gas, which is the top energy source for electricity in the U.S.

Bachar and others developing technologies that can fuel data centers with minimal emissions discuss their work in a new CNBC documentary, which you can watch above.

Since OpenAI released ChatGPT in 2022, Amazon, Google, Microsoft and other companies have been racing to open data centers that can handle generative artificial intelligence. These buildings are typically filled with power-hungry Nvidia graphics processing units. GPUs are the standard for training and running large language models that produce impressive chunks of text with a few words of human input. Executives across industries have seen what ChatGPT can do, and now they want to infuse generative AI into their products and internal operations, sometimes with hopes of boosting productivity.

If your data center doesn't have enough power for GPUs today, then executives will look elsewhere. Bachar knows that. It's a big part of his pitch.

He likes to say that utilities in some places, such as California and Virginia, can't help you right now if you want a lot of power for a data center. OpenAI's Sam Altman has invested hundreds of millions in nuclear startups, but they won't be ready to deliver energy for years, Bachar said.

After establishing ECL in 2021, Bachar has signed up two paying customers, with several other organizations that have placed orders for future delivery.

"It's the Microsofts, Facebooks, Amazons and Googles of the world ... which require all of this technology to be placed somewhere, and right now, somewhere is nowhere," said Bachar, explaining that traditional data centers in the U.S. can't be easily repurposed to work with AI.

ECL has plans to operate its sites efficiently, but as of now, it's tiny, with 10 employees and 18 contractors. That's much smaller than Altman's nuclear fusion investment, Helion, and the fission startup he backed, Oklo. Together the two employ nearly 600 people, representatives said.

Microsoft has committed to working with Helion, and the software company also signed a power purchase agreement in September to restart a nuclear reactor at Pennsylvania's Three Mile Island that shut down in 2019.

Nuclear installations inherently prompt questions about safety and the handling of waste, but their carbon-free status makes them attractive. Amazon, Google and Oracle have all explored small modular reactors with lower capacity than the ones at Three Mile Island.

The big tech companies are carefully watching their emissions in the AI age.

By 2030, Google wants to have net-zero emissions while Microsoft's goal is to be carbon negative by that year. Amazon has pledged to reach net-zero carbon by 2040.

"We're working with major tech companies, as well as various industrial players, to help them integrate our plug and play solution for on-site power generation into data centers," said Bret Kugelmass, founder and CEO of Last Energy, a Washington startup working on small modular reactors.

Bachar is fascinated with nuclear energy, but he said getting more of those facilities online will take time.

"We have a problem that we have to solve right now," he said.

In addition to his nuclear investments, OpenAI's Altman has bet on solar startup Exowatt. It has partners developing data centers that are consuming more than half of the energy available in their states in some locations, co-founder and CEO Hannan Happi said.

Geothermal energy has also garnered fresh interest in the modern AI era, with Google collaborating with startup Fervo Energy in Nevada. Tim Latimer, the startup's CEO, said Fervo has found a way to generate gigawatts of electricity in a single place by drilling horizontal holes underground, rather than the traditional vertical way.

Gigawatts are a serious quantity, but drilling holes for geothermal plants can be expensive, said Adrian Cockcroft, a former Amazon sustainability executive.

ECL intends to build a large-scale, 1-gigawatt data center in Texas over the next four years, with the help of hydrogen pipelines. It will probably take that long to move to zero-carbon green hydrogen using electrolyzers that convert water into hydrogen and oxygen, Bachar said.

But generating green hydrogen through electrolysis isn't cheap, said Kittu Kolluri, managing director of Neotribe Ventures.

The price of green hydrogen is to be determined, especially now that Donald Trump is U.S. president again, Bachar said.

Still, every gigawatt matters.

In 2028, U.S. data center demand could come in between 74 gigawatts and 132 gigawatts, according to a December report from the Lawrence Berkeley National Laboratory. Data centers might account for 6.7% to 12% of total U.S. energy consumption in 2028, up from 4.4% in 2023, the report said.

"The concern we have is can we grow fast enough to address the unprecedented demand for AI data centers," Bachar said.