How to Get Ahead of AI’s Coming Power Crunch

Dr. Ryan Ries here. Lately, I’ve been thinking a lot about the power problem in the AI industry, and I don’t see enough people talking about it.

While this may feel a bit doom and gloom at first, I promise there are solutions your business can (and should) be thinking about today as you’re planning your AI projects. Just keep reading!

Quick sidebar. I am co-hosting a re:Invent recap this Thursday. Register here if you’re interested in hearing about which re:Invent announcements are most important for your business!

Power Moves

Everyone's building AI strategies. Roadmaps. Proof of concepts. Business cases projecting ROI from machine learning models that'll revolutionize operations.

But there is a BIG problem: there might not be enough power to run them.

We talked about this in a previous Matrix when the deal was inked, but Microsoft just reopened Three Mile Island. Google signed with Kairos Power for 500 megawatts of molten salt reactors, the first one online in 2030. Amazon bought a data center campus next to a nuclear plant for $650 million, locking down 960 megawatts. Oracle's planning a one-gigawatt facility powered by three small modular reactors.

Why This Matters

An NVIDIA Blackwell GPU pulls 1,200 watts. Modern AI racks hit 240 kilowatts—enough to power 200 homes from one rack. A proper training cluster needs 500 megawatts running continuously. That's a mid-sized city's worth of power.

By 2030, US data centers will consume somewhere between 8-12% of national electricity, up from 4% today. The deployed H100 chips alone will use as much electricity as an entire country! That’s just mindblowing to me.

The part that is a bit scary is that the grid can't deliver it. There's power generation capacity out there, but the distribution infrastructure is saturated. Upgrading transmission lines typically takes a minimum of five to ten years.

AI companies need power now.

What This Means for Your Projects

When you're scoping your next machine learning project, you're probably thinking about data quality, model architecture, and engineering resources. You're not thinking about whether AWS or Azure will have enough power allocation to run your workload.

Cloud providers are going to start rationing compute. They already are in some regions. That beautiful architecture you designed for real-time inference might not matter if you can't get the GPU instances to run it.

This creates 3 immediate problems:

1. Cost Volatility

Power scarcity drives compute costs up. When hyperscalers build their own nuclear reactors to guarantee supply, they will pass those infrastructure costs through to their customers. Ensuring your business is as cost-optimized as possible is crucial. 

2. Availability Constraints

Getting access to high-end compute is becoming harder. Reserved instances are selling out faster. Spot pricing is more volatile. If your AI workload requires consistent GPU access and you haven't locked in capacity, you might find yourself unable to train or serve models when you need to.

3. Strategic Risk

Your competitors who secured compute capacity early have a structural advantage. They can iterate faster, train bigger models, serve more requests. If you're waiting to "see how AI develops" before committing resources, you might find there's nothing left to commit to.

The Nuclear Option

The tech giants aren't building nuclear reactors because they think it's fun. They're doing it because they've run the numbers and realized the grid can't support what they're planning to build.

NuScale got regulatory approval for their 77-megawatt small modular reactor design in May 2025. They're the only SMR vendor with full NRC approval, on track for deployment by 2030. 

But five years is a long time when you need compute today.

This Is The Key: What You Should Be Thinking About

When you're planning AI projects, add power availability to your risk register. 

Ask your cloud provider about their capacity roadmap. Understand whether the regions you're targeting have sufficient infrastructure. 

AWS, for example, is addressing power constraints through efficiency gains. Their data centers run 4.1 times more efficiently than typical on-premises setups. They’re developing custom processors like Graviton that use 60% less energy than comparable instances - stretching their power capacity further.

Consider whether your workloads can be optimized to run on less power-intensive hardware. Maybe that massive foundation model isn't necessary. Maybe a smaller, fine-tuned model running on cheaper compute gets you 80% of the value at 20% of the power cost. This is one of the reasons AWS was talking about training your own small model using Nova and doing model distillation at re:Invent.

In terms of processors, at Mission, we recommend AMD’s processors for our customers as they tend to be most efficient. 

Think about workload scheduling. Can your training jobs run during off-peak hours when power is cheaper and more available? Can you architect for intermittent compute access rather than assuming always-on availability?

Most of all, recognize that AI infrastructure is becoming a strategic asset, not just a utility you rent by the hour. The companies that secure their compute supply chain now will have a profound advantage over those who assume infinite cloud elasticity.

My Take

I spent years working on advanced imaging systems for the Department of Defense, building hardware that pushed the boundaries of what sensors could detect. We always designed with power constraints in mind because when you're deploying in the field, you can't just plug into unlimited electricity.

The AI industry is learning this lesson the hard way. All the brilliant algorithms in the world don't matter if you can't power the chips to run them.

The nuclear reactor deals aren't about being environmentally friendly or forward-thinking. They're about survival. These companies have looked at the math and realized that without dedicated power infrastructure, their AI ambitions aren’t going to work.

If you're building AI capabilities for your organization, you need to be thinking about this too.

If this raised questions or ideas for your team, subscribe to the newsletter. It’s the easiest way to get in touch and keep the conversation going.

Until next time,
Ryan

Now, time for this week’s AI-generated image and the prompt I used to create it:

Create a professional, high-resolution profile photo, maintaining the exact facial structure, identity, and key features of the person in the input image. The subject is framed from the chest up, with ample headroom and negative space above their head, ensuring the top of their head is not cropped. The person looks directly at the camera, and the subject's body is also directly facing the camera. They are styled for a professional photo studio shoot, wearing a smart casual blazer. The background is a solid '#141414' neutral studio. Shot from a high angle with bright and airy soft, diffused studio lighting, gently illuminating the face and creating a subtle catchlight in the eyes, conveying a sense of clarity. Captured on an 85mm f/1.8 lens with a shallow depth of field, exquisite focus on the eyes, and beautiful, soft bokeh. Observe crisp detail on the fabric texture of the blazer, individual strands of hair, and natural, realistic skin texture. The atmosphere exudes confidence, professionalism, and approachability. Clean and bright cinematic color grading with subtle warmth and balanced tones, ensuring a polished and contemporary feel.