The Power Wall

Why AI’s Next Crisis Isn’t Talent or Capital — It’s Electricity

“The only way of discovering the limits of the possible is to venture a little way past them into the impossible.” — Arthur C. Clarke

The Problem No One Wants to Name

In 2026, what matters more to AI: capital or power?

Capital is abundant. Venture funds have raised hundreds of billions specifically for AI investments. Microsoft, Google, Meta, Amazon — each could write checks for tens of billions without breaking stride. When OpenAI or Anthropic need funding, investors line up. Money flows toward proven scaling approaches like water finding the path of least resistance.

Power is scarce.

Not electrical generation capacity in the abstract — the world generates plenty of electricity. But power where it’s needed, when it’s needed, at scales AI training demands. And increasingly, that’s the constraint that matters.

OpenAI is reportedly seeking electricity contracts for 2030 approaching the output of China’s Three Gorges Dam — roughly 22,500 megawatts of continuous generation capacity. That’s enough to power tens of millions of homes. For training models.

Microsoft negotiated to restart Three Mile Island’s dormant nuclear reactor. Not to power a city. To power AI data centers.

Google, Amazon, and Meta are all in similar competitions for power capacity. They’re not competing for algorithms or talent or even data anymore. They’re competing for gigawatts.

Capital can’t solve this. You can’t throw money at physics. Building new nuclear plants takes ten to fifteen years. Solar and wind require massive storage solutions that don’t yet exist at necessary scales. Natural gas plants face regulatory and environmental opposition. The power grid wasn’t designed for AI-scale compute loads, and upgrading it is measured in decades, not quarters.

This is the power wall. And it changes everything about the AI race.

The Economics: What Does a Gigawatt Actually Cost?

“The future is already here — it’s just not evenly distributed.” — William Gibson

The case for understanding the power wall begins with a simple question: what does $50 billion buy you in AI infrastructure?

Based on current estimates from Bernstein Research ($35B per GW for AI-optimized data centers) and Nvidia ($50–60B per GW), that investment buys roughly 1 to 1.5 gigawatts of compute capacity on Earth. That’s 4.3 million GPU-equivalents. Annual operating costs run about $2.5 billion per year — electricity, cooling, maintenance, real estate.

The five-year total cost of ownership: $63.5 billion. And you cannot expand beyond initial capacity, because the grid won’t let you.

At $3 per GPU-hour and 50% utilization, that facility generates $140 billion in revenue over five years, netting $76.5 billion in profit. Impressive. But the ceiling is hard and permanent.

The question for CXOs and investors isn’t whether AI compute is profitable. It is. The question is whether it’s scalable. And on Earth, with current grid constraints, the answer is: barely.

The Thorium Variable

China recognized the energy bottleneck early. While Western AI labs focused on algorithmic improvements and model scaling, China invested heavily in thorium reactor technology. Thorium — more abundant than uranium, inherently safer, producing less long-lived nuclear waste — could be the energy breakthrough that reshapes geopolitical competition.

China’s experimental thorium molten salt reactor went operational in 2021. Copenhagen Atomics in Denmark has designed a container-sized thorium reactor delivering 100 megawatts thermal. ThorCon is building 500-megawatt reactors inside ship hulls. These are not theoretical designs — they are engineered products approaching commercial deployment.

Here’s the beautiful thing about thorium molten salt reactors: they run at atmospheric pressure. That means no giant pressure vessel. No risk of steam explosions. If something goes wrong, the fuel drains into a catch basin and solidifies. They’re compact, modular, and remarkably safe compared to conventional nuclear.

If China achieves commercial-scale thorium power while the West remains dependent on conventional nuclear and fossil fuels, the economic advantage is profound. Training AI models at a fraction of Western electricity costs doesn’t just make Chinese AI cheaper. It makes it potentially unlimited in scale.

The assumption underlying Western AI strategy — that compute is available if you can pay for it — might only be true for whoever controls future energy infrastructure.

Jevons Paradox: Why Efficiency Won’t Save Us

“The first principle is that you must not fool yourself — and you are the easiest person to fool.” — Richard Feynman

The question everyone asks: what if AI gets so efficient we don’t need this much power anymore?

The answer: efficiency gains unlock demand. It’s called Jevons Paradox, observed in 1865 when steam engines got more efficient and coal consumption exploded instead of declining.

GPU efficiency has improved 11x since 2016. The P100 delivered 21 FLOPS per watt; the B200 delivers 240. Yet AI power consumption has increased 30x in the same period — from roughly 1 GW globally to 30 GW. Demand is outpacing efficiency 3:1.

Why? Because efficiency unlocks new use cases. In 2016, AI couldn’t generate video, fold proteins at atomic resolution, drive cars, or write code. Each new capability requires massive compute. And each new capability creates demand for the next one.

GPT-2 in 2019 used 1.5 billion parameters and cost $50,000 to train. GPT-4 in 2023 used an estimated 1.8 trillion parameters and cost $100 million. The next generation will likely use 10 trillion parameters and cost $500 million or more. Even with 11x more efficient GPUs, models got 100x bigger.

Then there’s the inference explosion. Today, training and inference consume roughly equal power globally — about 15 GW each. By 2030, inference is projected to hit 150 GW as billions of people use AI daily, autonomous vehicles scale, and IoT devices run edge AI. Training may be the glamorous problem, but inference is the quiet killer.

Efficiency doesn’t kill demand. Efficiency feeds demand.

The Economic Cascade: Who Actually Benefits

The conventional wisdom about large infrastructure projects: they’re vanity projects for the already-wealthy that create a handful of high-paid engineering jobs and nothing for ordinary people. The data tells a different story.

Aerospace and energy manufacturing creates 3,500 jobs per billion dollars invested, compared to 800 for software. And 60% of those jobs are blue-collar — welders, machinists, assembly workers, technicians — paying $35,000 to $60,000 per year. No college degree required.

These aren’t jobs in Silicon Valley. They’re jobs in places like Brownsville, Texas, where SpaceX’s presence has already increased high school graduation rates by 8 percentage points and college attendance by 17 points. Kids see a pathway: six months of welding certification leads to a $48,000-per-year job at age 19.

The supply chain multiplier matters too. When a major infrastructure project buys $2 billion in stainless steel, that money flows to mills in Gary, Indiana and Birmingham, Alabama — Rust Belt cities desperate for manufacturing jobs. Reactor components built in South Korean shipyards, electronic assemblies in Mexican factories, precision parts from small machine shops across rural America.

For every 10,000 manufacturing jobs created, another 1,500 service sector jobs follow — restaurants, retail, childcare, healthcare. Workers spend 70% of their income locally. Communities get built.

The irony: building infrastructure to support AI creates more working-class jobs than building the AI itself.

Infrastructure is how working-class communities get built. The power wall isn’t just an AI problem. Solving it is an economic development opportunity.

The Geopolitical Dimension

The power wall doesn’t exist in a political vacuum. Whoever solves it first doesn’t just gain advantage — they potentially define the dominant paradigm for AI development.

China has the world’s only operating commercial molten salt reactor, proven lunar landing capability, and a $15-billion-per-year space program. They are not standing still.

The United States has SpaceX, the world’s most advanced launch system, and the largest AI industry on Earth. But it lacks commercial thorium reactor technology and faces a regulatory environment that can delay nuclear projects by a decade or more.

Europe has Copenhagen Atomics and strong engineering traditions, but moves slowly on both energy and AI policy.

India has the world’s largest thorium reserves and a growing AI sector, but limited capital for infrastructure at this scale.

The Middle East has sovereign wealth funds with hundreds of billions available, but no indigenous technology base in either nuclear or AI.

No single nation can solve the power wall alone. The capital requirement ($50 billion or more), the technical breadth (reactors, compute, launch systems, robotics), and the regulatory complexity span borders.

The organizations that move first — whether nations, corporations, or consortia — will define the next phase of AI development. Everyone else will pay their prices.

What Choice Do We Have?

The AI industry will spend over $200 billion between now and 2030 trying to secure 50 to 100 gigawatts of compute capacity on Earth. Grid capacity, environmental reviews, regulatory approval, political opposition — every constraint is real, and every constraint is fatal to the timeline.

The alternative isn’t finding power somewhere else on Earth. Every location faces the same fundamental limits: regulations, competing demands, grid infrastructure that was never designed for this. Whether you build in the Arctic, the desert, or the deep ocean, you’re still on Earth. Still subject to the same political and physical constraints.

The alternative to solving the power wall is stopping. And stopping means someone else doesn’t.

By 2030, AI companies will have exhausted the easy options. Every nuclear restart contract will be signed. Every available grid connection will be claimed. Every regulatory shortcut will have been tried. And demand will still be growing.

The power wall isn’t a future problem. It’s today’s problem with a five-year fuse.

Whoever solves it — through new energy technology, new infrastructure models, new international partnerships, or some combination we haven’t imagined yet — will own the AI industry for the next generation.

The question isn’t whether this will be solved. It’s who solves it first, and whether the rest of us are inside the tent or outside watching.

“There are decades where nothing happens; and there are weeks where decades happen.”

Philip Cheevers is a systems architect and writer with four decades of experience building complex infrastructure across telecommunications, healthcare, energy, and AI governance. He is the inventor of the Michelangelo AI governance framework and writes about AI strategy, chronolinguistics, and the intersection of technology and human systems.