Do we need AI?
Talen Energy announced its sale of a 960-megawatt data center campus to cloud service provider Amazon Web Services (AWS), a subsidiary of Amazon, for $650 million.
The data center, Cumulus Data Assets, sits on a 1,200-acre campus in Pennsylvania and is directly powered by the adjacent Susquehanna Steam Electric Station, which generates 2.5 gigawatts of power.
“We believe this is a transformative transaction with long term benefits,” said Mark “Mac” McFarland, Talen president and chief executive officer of Talen, on a Monday call with investors and media. As power demand continues to rise worldwide, “data centers are at the heart of that growth,” he added.
“Several years ago, Amazon set an ambitious goal to reach net-zero carbon by 2040—ten years ahead of the Paris Agreement. As part of that goal, we’re on a path to power our operations with 100 percent renewable energy by 2025—five years ahead of our original 2030 target,” an Amazon spokesperson said. “To supplement our wind and solar energy projects, which depend on weather conditions to generate energy, we’re also exploring new innovations and technologies and investing in other sources of clean, carbon-free energy. This agreement with Talen Energy for carbon-free energy is one project in that effort.”
And the advent of artificial intelligence (AI), which requires far more processing power than standard computing, has put rocket boosters under the data centre world.
His company has just built a huge new data centre in Portland, Oregon dedicated to AI. Just how different is this from an ordinary data centre?
“A normal data centre needs 32 megawatts of power flowing into the building. For an AI data centre it’s 80 megawatts,” says Mr Sharp.
AI systems are using all this extra electricity simply because they are doing so much more processing than standard computing. They are chewing through far more data.
Mr Sharp also points out that the entire web of technical support demanded by AI is greater. “You have five times more cabling, for instance.”
All of this points to a problem. How can AI grow when it requires so much more power to function?
Demanding ever more juice from the existing grid means competing with homes and other industries, and is not going to win the data centre sector any friends if blackouts result.
“Our industry has to find another source of power,” Mr Sharp declares. He reckons that is nuclear.
More pressingly, he predicts that data centres in the not too distant future will come with their own dedicated, built-in nuclear reactors.
The technology in question is the much-touted Small Modular Reactor (SMR). These are designs for advanced reactors with about a third of the power generation of a traditional, large nuclear plant.
While there are currently no SMRs in commercial operation around the world, China is building the world’s first, and similar technology is already used by nuclear-powered submarines.
Meanwhile, universities, such as the UK’s Imperial College London, have for years operated small nuclear reactors for teaching and training purposes. Imperial’s own reactor, located just outside London, was operational from 1965 to 2010.
Today most companies developing SMRs for commercial use are focusing on helping towns and cities to keep their lights on. However, a clutch of specialist firms have decided that data centres are the best candidates for their SMR designs.
Dr Michael Bluck runs the Centre for Nuclear Engineering at Imperial College London. “Data centres are power hungry things, but with AI we’re moving into a new level of power requirements,” he says.
“There are about 50 SMR designs out there. The challenge is to build them in repeatable units, factory style, standardising production lines.
“There’s no reason why a small fast reactor can’t power a data centre, except that you have to get it past the regulator.”
======
A few years ago, I interviewed Arjun Makhijani PhD, a skeptic of nuclear energy, who raised compelling issues about one of the main purported advantages of “small modular reactors”.
Thank you for the clip from Arjun, who is always brilliant. There are a few additional points to consider in that Boeing-style supply chain analogy. The first is fuel–where does the Uranium come from when we are already well past peak-U and having to consider extraction from seawater and reprocessing of spent fuel? The second is enrichment. Governments have traditionally paid for that and it is extremely energy intensive. Fuel fabrication can scale, SMR construction can scale, operations can scale–all following typical startup industry models. But there are two add-ons that Boeing or Tesla don’t have–security and waste disposal. Security includes the proliferation risk which will also scale. Waste includes emissions, both accidental and unavoidable, at all stages of the fuel cycle. Those are significant but largely (and intentionally, historically) hidden costs (to public health, safety, welfare and democracy) that should be the real deal breakers.
It’s the training of AI that is critical to whether using exaflops of processing power comes up with an answer just as stupid as a human would.
Just waiting for AI to implement the Matrix power source.