Microsoft’s Unlikely Three Mile Island AI Power Play   

Last month, Microsoft shocked power pundits, announcing its plan to bring the infamous Three Mile Island nuclear facility back from the dead. It was a PR agent’s worst nightmare. And if we’re serious about decarbonizing AI, it was the exact right move.   

Not Just an AI Problem

In 1979, Three Mile Island experienced the worst accident ever at a US commercial power plant. For the next three decades, heightened regulatory restrictions and chronic construction project mismanagement drove the US nuclear energy business ever deeper into the red. 

The 2011 Fukushima disaster and subsequent nuclear bans by prominent political leaders appeared to be the final nail in the nuclear coffin. Then generative AI changed everything. 

Or so it seemed. Goldman Sachs recently estimated each ChatGPT query needs ten times more electricity than a Google search and that AI will drive +160% absolute growth in data energy demand from 2024 to 2030.  These estimates are valid. 

But non-AI applications like gaming and cryptocurrency had already been tripling data center energy demand from 2015 to 2019, well before the 2022 launch of Chat GPT. And as of mid-2024, AI represented less than 10% of total data center energy consumption.  

AI presents a greenhouse gas challenge. But Microsoft’s energy breakthrough came almost in spite of it. 

Vast Leverage

As such, the best way to tackle AI’s energy problem may well be to piggyback on the global energy transition writ large. Microsoft’s gamble on nuclear power creates exceptional decarbonization leverage for any customer of the Microsoft cloud, and other cloud and energy leaders are racing down precisely the same energy path: 

• Google has partnered with Houston-based Fervo Energy to power its Nevada data center with 100% renewable, always-on geothermal energy. Fervo’s horizontal drilling breakthroughs have increased geothermal-ready locations on Earth from less than 5% to greater than 30%.  Expect more geothermally powered data center operations very soon.
• Schneider Electric templatizes power purchase agreements to help its customers secure rate advantage and hedge against price volatility. 
• Siemens and GE offer a new class of specialized “power-to-x” energy management services specifically tailored to decarbonize industrial process flows. 

Per previous Zero100 research, supply chain leaders who move fast to tap the energy transition ecosystem could tap a decarbonization flywheel that’s already well in motion. But innovations in direct carbon reduction offer equal promise.  

The key is to keep it simple. 

Common Sense Playbooks

A landmark HBR study of C-suite energy management strategies found that leaders who drive faster to net zero treat decarbonization like any other business transformation.  

AI acceleration calls for the aggressive review of technology levers to assure literacy in the latest tactical and systems-level efficiency breakthroughs. But the keys to success are strikingly aligned with other transformation fundamentals:  

Sharpen C-level AI carbon understanding and ambition  

• In 2017, Microsoft formed a central energy team made up of 14 experts in electricity markets, renewable energy, battery storage, and local generation, charging it with developing and executing the firm’s energy strategy. Microsoft’s CFO and President are now actively involved in setting and reviewing the company’s renewable energy targets. 
• Unilever, Walmart, and others actively cultivate relationships with late-stage energy venture and private equity investors like Energy Impact Partners and DBL Ventures to help narrow the list of renewable energy innovation candidates and identify promising long-range contenders. These include battery specialists like Tesla and Siemens, smart grid innovations from GE, ABB, IBM, and Schneider Electric, and next-generation nuclear innovations, eg, MIT’s Commonwealth Fusion Systems, which help crystallize potential 10+ year power transition roadmaps. 

Continuously evaluate and rank AI decarbonization opportunities

• For example, investments in tactical efficiencies may be leapfrogged by fundamental infrastructure breakthroughs. As just one illustration: Data center startup Sustainable Metal Cloud (SMC) recently demonstrated the ability to reduce AI chip energy consumption by up to 50% through its immersion cooling technology, which is particularly effective for cooling high-density GPU setups in data centers.   

Measure and reward decarbonization excellence 

• Johnson & Johnson and General Motors (GM) set capital aside for efficiency and carbon reduction projects ($40 million and $20 million, respectively). Others use an internal carbon price to determine allocation. 
• GE runs efficiency “treasure hunts,” compensating employees who discover and deploy local efficiency improvements, driving proverbial carbon-cost win-wins. Estimated savings: $150 million and counting.   

Embed continuous efficiency improvement into industrialized AI processes

• This is exemplified by P&G’s AI Factory.
• Seamlessly layer on new, energy-centric KPIs on balanced model performance scorecards. 

Run AI energy scenarios and simulations to anticipate energy what-ifs

• Build and evolve simulation models to model a range of potential AI decarbonization outcomes. 
• Inform AI plans and roadmaps with new information as efficiency breakthroughs and barriers emerge. 

And AI itself may help us accelerate the process of prediction. So, we gave it a try. 

AI, Heal Thyself

Using OpenAI’s newest experimental version of ChatGPT (o1), we decided to see if genAI could help ideate simulations of our AI energy future. The results were striking.

With three queries, ChatGPT was able to draft the logic and build a workable, global AI energy supply-demand prediction-simulation model. In one click, ChatGPT generated the code. With the next, ChatGPT gave us our answer: 

By 2032, the supply of carbon-free power will exceed AI-generated demand, with a sensitivity range of +5 years for factors like delayed returns to scale on newer, unproven sources and uses (see the model’s work and output here).  

Of course, this forecast, like all others, is wrong. But the draft logic was remarkably sound, saving us days of coding on the path to a dynamic what-if capability that should be in every supply chain executive’s toolkit.    

AI won’t fix carbon on its own. But aggressive AI adoption mixed with the courage to embrace systems reinvention may be the new supply chain winner’s code.