ANALYSIS

A job ad from Microsoft shows it’s looking to harness the power of the atom to fuel its growing datacenter footprint, but it’d hardly be the first.

Redmond is looking for someone who “will be responsible for maturing and implementing a global Small Modular Reactor (SMR) and microreactor energy strategy.”

Microsoft’s interest in SMRs isn’t even the company’s first foray into nuclear power. In May, Redmond signed up to buy power from fusion energy startup Helion beginning in 2028, despite the fact the company has yet to prove its tech works.

And over the past few months, we’ve seen numerous datacenter operators embrace nuclear power as a means to end their reliance on fossil fuels.

Last month, Green Energy Partners and IP3 detailed a plan to build a massive datacenter campus in Virginia powered entirely on small modular reactors (SMRs) — more on those later. Meanwhile, Swedish datacenter operator Bahnhof is also investigating the use of the miniaturized reactors to not only power its operations, but as many as 30,000 households in the surrounding area.

Meanwhile, Cumulus Data opened the doors on a nuclear-powered datacenter in Pennsylvania this January. The 48 megawatt facility is located alongside the 2.45 gigawatt Susquehanna power plant.

DCs thirst for power

What’s behind this recent trend? It’s simple, datacenters are consuming more power than ever before thanks – in part – to the explosive growth of generative AI.

To put things in perspective, today a reasonably large cloud or hyperscale campus might be rated for 50 megawatts of capacity. However, anyone who’s been paying attention to the GPUs and other accelerators that AI workloads run on knows that while they’ve been getting progressively more effective, they’re also using a lot more power.

AMD CEO Lisa Su highlighted this inconvenient truth during a speech at the International Solid-State Circuits Conference earlier this year. The problem, she explained, is that while the performance of CPUs and GPUs has roughly doubled every 2.4 years, power efficiency hasn’t kept up.

Su estimated that a zettaflop-class supercomputer — one 1,000x faster than the US’s chart-topping Frontier system — was only about a decade away, but would require somewhere in the neighborhood of 500 megawatts of power. “That’s on the scale of what a nuclear power plant would be,” she said at the time.

You might be thinking something along the lines of “yeah, but that’s a supercomputer,” but the fact is the AI clusters used to train large language models like OpenAI’s GPT-4 are constructed from many of the same building blocks.

In fact, many modern AI clusters deployed by the likes of Microsoft, Google, and Meta would theoretically place among the 10 most powerful systems on the Top500 ranking of supercomputers.

The rise of the microgrid

The amount of power those facilities draw means datacenter operators know they can’t always tap the juice they need from the grid.

In locations like Northern Virginia or Dublin, Ireland, where many datacenters congregate, power is often cheap but also in incredibly high demand. During peak operating hours, it’s not uncommon for datacenters in these markets to supplement grid power using backup generators.

Some datacenter operators have gone so far as to build their own power plants — sometimes called “microgrids” — to deal with this instability in electrical supply. For example, back in July, Microsoft got the green light to build a 170 megawatt natural gas generator to keep the lights on at its €900 million datacenter development outside Dublin. The plant, which consists of 22 individual generators and reportedly cost €100 million to build, is designed to operate during periods in which the national grid can’t keep up.

Microsoft isn’t alone. Amazon plans to employ natural gas fuel cells at some of its Oregon datacenters, after reportedly running into trouble securing adequate grid capacity to feed its needs.

While both examples address the cloud provider’s power supply challenges, they still rely on fossil fuels. It’s hard to say you’re serious about cutting your carbon footprint while continuing to run your datacenters on decomposed dinosaurs.

And this is where SMRs start to make sense. They emit little or no CO2 during normal operations and are capable of generating a considerable amount of power for a relatively small footprint.

• Microsoft hiring a nuclear power program manager, because AI needs lots of ‘leccy
• Virginia industrial park wants to power DCs with mini nuclear reactors, clean hydrogen
• First US nuclear power plant built this century goes online
• Why do cloud titans keep building datacenters in America’s hottest city?

Not some newfangled fantasy

SMR’s are essentially miniaturized nuclear reactors that can – in theory – be mass produced and scaled out as additional power is required.

But while commercial examples are only now gathering steam, the underlying technology dates back to the 1950s in reactors built for nuclear submarines like the USS Nautilus.

Depending on the SMR in question — there are actually quite a few designs from the likes of NuScale, TerraPower, and Westinghouse and others — the International Atomic Energy Agency (IAEA), says these reactors can generate anywhere from tens to hundreds of megawatts of electrical output.

NuScale’s SMR designs are particularly notable, as they recently received certification by the Nuclear Regulatory Commission necessary for commercial deployment. According to NuScale’s website, its reactor design is rated to produce about 77 megawatts of power.

Roadblocks remain

In a report published last year, Omdia analysts concluded that SMRs could prove to be a viable alternative to generators powered by fossil fuels, especially if regulatory and financial headwinds can be overcome. But despite recent developments and interest around SMRs, the technology is likely years off, Omdia’s Alan Howard told The Register.

“The most optimistic deployment of an SMR here in the United States is by 2030,” he said in an interview. “The notion of it being used on a datacenter campus, that’s going to be – and I’m only speculating here –between 10 and 15 years away.”

As you can imagine, building a small nuclear reactor isn’t the hard part; we’ve been doing that for decades. The real challenge is making the tech price competitive with other sources of power.

NuScale claims to be getting close with a levelized cost estimate (LCOE) of between $40/MWh to $65/MWh, when they reach commercial availability during the latter half of the decade. For reference, the LCOE for something like natural gas is around $37/MWh.

However, this target likely remains a ways off, according to a Utah City council meeting agenda which put the cost of development for NuScale’s modular reactor at closer to $89/MWh.

With that said, there are several SMR startups out there working to address this very problem. At the same time, the cost of fossil fuel is likely to continue rising over the next few years, and the datacenters are only getting hungrier. ®