Off-Prem

Fission impossible? Meta wants up to 4GW of American atomic power for AI

Facebook titan targets early 2030s for reactor deployment


Meta believes it will need one to four gigawatts of nuclear power, in additional to the energy it already consumes, to fuel its AI ambitions. As such, it will put out a request for proposals (RFP) to find developers capable of supplying that level of electricity in the United States by early 2030.

"Advancing the technologies that will build the future of human connection — including the next wave of AI innovation — requires electric grids to expand and embrace new sources of reliable, clean and renewable energy," the Facebook parent company wrote in a blog post announcing the RFP on Tuesday.

But while Meta plans to continue investing in solar and wind, hyperscalers seem convinced that harnessing the atom is the only practical means of meeting AI's thirst for power while making good on its sustainability commitments.

This wouldn't be the first time Meta has pursued nuclear fission power. As we previously reported, Meta had planned to build an atomic datacenter complex, but was foiled after a rare species of bees were discovered on a prospective site, resulting in its cancellation.

Meta has become a leading developer of generative AI models with Llama 3.1 405B being among its most sophisticated. To support the development of these and future models, Zuckerberg has committed to deploying some 600,000 GPUs, which require a prodigious amount of power to run.

As we understand it, additional details regarding the nature of the RFP will be provided to qualified companies. However, we do know that Meta is looking for someone to deploy between one and four gigawatts of nuclear power, suggesting they're still a little uncertain as to the extent of power that'll be required to achieve their goals and that these plans are destined for the US.

The blog post also mentions the prospect of deploying multiple units to cut costs. Given the timeline, this suggests that Meta is very likely looking at small modular reactors (SMRs).

As their name suggests, SMRs are really just miniaturized reactors not unlike those found in nuclear submarines and aircraft carriers, which can be manufactured and co-located alongside datacenters and other industrial buildings.

Many hyperscalers and cloud providers faced with AI's energy demands have turned to SMRs for salvation, and there's certainly no shortage of options to choose from. Oklo, X-energy, Terapower, Kairos Power, and NuScale Power are just a handful of the companies actively developing reactor designs. However, it's worth noting that despite all the hype around these itty bitty reactors, nobody has actually managed to prove their commercial viability.

But with few alternatives that don't involve abandoning their lofty sustainability pledges, many datacenter operators are pushing ahead with power purchase agreements with SMR vendors. Most recently, Sam Altman-backed startup Oklo revealed it had obtained letters of intent from two major datacenter providers to deliver 750 megawatts of power.

Amazon has also committed to investing in nuclear power. Back in October, the e-commerce and cloud giant announced it was working with X-energy to construct several SMRs. Google, meanwhile, has teamed up with Kairos on a similar plan, and Oracle says it's obtained building permits for a trio of SMRs to power a one gigawatt datacenter campus.

However, it remains to be seen whether these plans will ever pan out. In addition to strict regulatory controls on nuclear power, the technology is seen by many as unsafe despite evidence to the contrary. Perhaps more pressing is the fact that SMRs, at least in the early days, won't be cheap.

Earlier this year, the Institute for Energy Economics and Financial Analysis argued that SMRs are "too expensive, too slow to build, and too risky to play a significant role in transitioning away from fossil fuels."

Of course, SMRs aren't Meta's only option here. Microsoft is working with Constellation Energy to bring the decommissioned Unit 1 reactor at Three Mile Island back online. Meanwhile, Amazon earlier this year purchased Talen Energy's Cumulus atomic datacenter co-located alongside the 2.5 gigawatt Susquehanna nuclear power plant in Pennsylvania for $650 million.

However, even existing nuclear infrastructure isn't a sure bet. This November, Amazon hit a roadblock after federal regulators rejected a deal that would have let it increase its power draw at the site from 300 to 480 megawatts.

Bringing these plans online is by no means trivial. As we previously reported, the Palisades nuclear power plant in Michigan, which received a $1.5 billion loan from Uncle Sam, will require substantial and costly repairs to its steam generator tubes. ®

Send us news
26 Comments

Day after nuclear power vow, Meta announces largest-ever datacenter powered by fossil fuels

Louisiana facility's three natural gas turbine plants to churn out 2,262 MW

AI's rising tide lifts all chips as AMD Instinct, cloudy silicon vie for a slice of Nvidia's pie

Analyst estimates show growing apetite for alternative infrastructure

Million GPU clusters, gigawatts of power – the scale of AI defies logic

It's not just one hyperbolic billionaire – the entire industry is chasing the AI dragon

US bipartisan group publishes laundry list of AI policy requests

Chair Jay Obernolte urges Congress to act – whether it will is another matter

Take a closer look at Nvidia's buy of Run.ai, European Commission told

Campaign groups, non-profit orgs urge action to prevent GPU maker tightening grip on AI industry

Infosec experts divided on AI's potential to assist red teams

Yes, LLMs can do the heavy lifting. But good luck getting one to give evidence

Apple and Meta trade barbs over interoperability requests

Both are only thinking about the best interests of users, of course

Are you better value for money than AI?

Tech vendors start saying the quiet part out loud – do enterprises really need all that headcount?

Apple called on to ditch AI headline summaries after BBC debacle

'Facts can't be decided by a roll of the dice'

Ireland fines Meta for 2018 'View As' breach that exposed 30M accounts

€251 million? Zuck can find that in his couch cushions, but Meta still vows to appeal

Cheat codes for LLM performance: An introduction to speculative decoding

Sometimes two models really are faster than one

American cops are using AI to draft police reports, and the ACLU isn't happy

Do we really need to explain why this is a problem?