Transformer lead times stretched past 12 months. Grid connection queues hit 2-4 years. And the four largest cloud operators are planning to spend north of $200 billion in 2024 on infrastructure that, in many cases, cannot get a grid interconnection date before 2027.
This is not a supply chain story. It is a physics story. The AI data center buildout has hit a wall that no amount of chip allocation can fix. The bottleneck shifted from TSMC's fab capacity to the local substation transformer. And the market is only beginning to price this.
Rich McCormick's warning about AI data center expansion risks deserves more attention than it's getting. Not because the warning is new, but because the data behind it is being systematically ignored by investors who still treat AI as an infinite growth story. The numbers tell a different story. IEA data shows global data center electricity consumption moving from 460 TWh in 2022 to over 1,000 TWh by 2026. In the US, data centers will go from roughly 3% of national electricity demand to 8-10% by 2030. That is not a linear extrapolation. That is a hockey stick colliding with a physical ceiling.
I've spent the last three years building trading infrastructure around on-chain data. I've traced the exact block where the Terra peg broke. I've monitored GBTC premium spreads for arbitrage. What I've learned is simple: infrastructure outlasts innovation. And right now, the infrastructure constraint is no longer silicon. It's electrons.
The Power Density Problem
Let me break down the technical reality. Traditional data centers run at 5-10 kW per rack. AI data centers need 30-100 kW per rack. That's not an incremental change. That's an order of magnitude shift in power density requirements, which fundamentally changes the cooling, electrical distribution, and grid interface design.
Uptime Institute's 2024 data confirms this. The power density jump is real and it's happening faster than the electrical grid can adapt. The US grid has an average infrastructure age exceeding 30 years. Transformer manufacturing capacity hasn't kept pace with demand. The result? A transformer that used to ship in weeks now takes over a year. For a data center developer, that's not an inconvenience. That's a project-killing delay.
I saw this pattern before. In 2022, when Terra collapsed, I traced the exact block where the algorithmic peg broke due to a flash loan exploit. The sequence was: liquidity pool depletion → arbitrage failure → death spiral. The market didn't see it coming because everyone was focused on the yield narrative, not the mechanics. The same thing is happening with AI infrastructure. Everyone is focused on the compute narrative, not the energy mechanics.
The Cost Structure Shift
Here's the number that matters most. Energy costs as a percentage of total cost of ownership (TCO) for data centers have shifted from 15-20% in traditional facilities to 30-50% in AI data centers. That's the single largest variable cost, and it's rising.
Let me put this in trading terms. When your variable costs go from 20% to 50% of your cost structure, your margin sensitivity to energy prices quadruples. A 10% increase in energy costs that used to hit your bottom line by 2% now hits it by 5%. That's not a margin squeeze. That's a margin restructuring.
The cloud providers haven't fully passed these costs to customers yet. API pricing per token remains relatively stable. But that's a temporary state. The economics are simple: if energy costs rise and prices don't, margins compress. If margins compress, either prices rise or capex slows. Either way, the growth narrative takes a hit.
I don't predict, I react. And what I'm reacting to is a cost structure that is fundamentally deteriorating at the margin.
The Training vs. Inference Blind Spot
Here's something the mainstream coverage misses. The energy narrative around AI typically focuses on training runs. GPT-3's training consumed approximately 1.3 GWh. GPT-4's training is estimated at around 50 GWh. That's a 38x increase. Impressive numbers. But they miss the bigger story.
Inference energy is about to overtake training energy. By 2026, inference will account for more than half of AI data center energy consumption. Why does this matter? Because training is a one-time cost. You train a model, you incur the energy cost once, and you're done. Inference is ongoing. Every user query, every API call, every token generated consumes energy. This is a recurring, growing energy obligation that scales with adoption.
This changes the investment calculus. Training energy costs are front-loaded capex. Inference energy costs are opex that grows with revenue. If you're modeling AI company economics, the inference energy curve is the one that matters. And it's steeper than most analysts realize.
The Geographic Arbitrage
Energy costs are not uniform. They vary dramatically by region. This creates a geographic arbitrage opportunity that sophisticated players are already exploiting. Texas, Ohio, and Iowa have cheap energy and are attracting data center investment. California and New York, with their higher energy costs and regulatory friction, are being bypassed.
This isn't just about electricity prices. It's about grid interconnection timelines. In regions with congested grids, interconnection queues stretch to 4+ years. In regions with available capacity, projects can come online in 18-24 months. That difference is now the critical competitive variable in AI infrastructure buildout.
I've been tracking this from a market perspective. The data shows that energy-rich regions are becoming the new compute hubs. This is a structural shift that will reshape the geography of AI infrastructure over the next five years.
The Nuclear Bet
Microsoft signed a nuclear power agreement with Constellation Energy in 2024. Google invested in SMR (small modular reactor) startups. Amazon has made similar moves. These aren't PR stunts. They're recognition that renewable energy alone can't provide the stable, 24/7 baseload power that AI data centers need.
But here's the reality check. SMRs are not commercially viable at scale yet. The regulatory approval process for new nuclear in the US takes 10-15 years. Even with accelerated pathways, you're looking at 5-8 years before SMRs meaningfully contribute to data center power supply. That's a long time when your grid connection queue is already 3 years.
The nuclear bet is real, but it's a long-duration play. In the meantime, the immediate constraint is grid infrastructure. And that's not getting fixed quickly.
The Efficiency Paradox
Now let me offer the contrarian angle. The energy crisis narrative is real, but it's also incomplete. There are efficiency gains happening that partially offset the demand growth.
NVIDIA's transition from H100 to B200 represents a significant efficiency improvement per unit of compute. Algorithmic innovations like FlashAttention and Mixture-of-Experts (MoE) architectures reduce the compute required for equivalent model performance. Quantization, distillation, and sparsification techniques are reducing the energy intensity of inference.
PUE (Power Usage Effectiveness) optimization is another lever. Moving from 1.5 to 1.2 PUE reduces total energy consumption by about 20%. Liquid cooling, which is transitioning from 10% penetration in 2023 to an expected 40%+ by 2028, dramatically improves cooling efficiency.
These gains matter. But they don't negate the core problem. They slow the curve, they don't flatten it. The IEA projections already account for some efficiency improvements, and they still show doubling of data center electricity consumption by 2026.
The efficiency paradox is this: efficiency improvements make AI cheaper, which drives more adoption, which drives more energy demand. Jevons Paradox in action. The more efficient we make AI compute, the more AI compute we use. The energy curve keeps going up.
What the Market Is Mispricing
From my trading seat, the market is making a classic error. It's pricing AI infrastructure as if energy constraints don't exist. The capex numbers from Microsoft, Google, Amazon, and Meta are being treated as pure growth signals. But a significant portion of that capex is now going to energy infrastructure — power purchase agreements, grid interconnection fees, on-site generation, cooling systems. That's not growth capex. That's defensive capex.
When you dig into the numbers, the picture becomes clearer. Energy-related infrastructure investment is consuming an increasing share of AI data center capex. Dell'Oro Group projects global AI data center investment exceeding $300 billion in 2025, with energy-related components growing fastest. That's capital being deployed to maintain the status quo, not to expand compute capacity.
Code doesn't lie, but markets do. The market is telling you these are growth investments. The data tells you a significant chunk is survival spending.
The Political Dimension
The energy constraint isn't just technical. It's political. Data center energy consumption is becoming a local political issue. Virginia, which hosts the largest concentration of data centers in the US, has seen residents complain about rising electricity bills. Washington state is discussing additional energy taxes on data centers.
This creates regulatory risk that isn't priced into AI infrastructure investments. If data centers face energy surcharges or stricter permitting requirements, the economics shift further. And this isn't just a US issue. Europe is facing similar pressures, with Ireland and the Netherlands imposing moratoriums on new data center connections due to grid constraints.
The geopolitical dimension adds another layer. The US-China competition in AI is now extending to energy infrastructure. China has advantages in grid construction and renewable energy deployment. The US has advantages in chip technology. The question is which constraint binds first — chip access or energy access.
The Water Issue Nobody's Talking About
Here's a detail most analysis misses. Data centers don't just consume electricity. They consume water for cooling. A typical 100 MW data center can use 1-3 million gallons of water per day, depending on cooling technology. In water-stressed regions, this creates an additional constraint that's harder to solve than electricity.
Liquid cooling reduces water consumption compared to traditional evaporative cooling, but it increases the complexity and cost of the facility. The trade-off between water and electricity is another dimension of the infrastructure constraint that's not being adequately modeled.
The Investment Implication
What does this mean for positioning? The energy bottleneck creates a clear investment thesis. Companies that provide energy infrastructure for data centers — grid equipment manufacturers, transformer producers, cooling system providers, energy storage companies — are positioned to benefit from the massive capital flows into AI energy infrastructure.
The companies that are most exposed to energy cost risk are the ones with the least pricing power. The AI service providers who can't pass through energy costs will see margin compression. The ones who can pass through costs will see their growth narrative challenged as prices rise.
Volatility is just unpriced risk. The risk here is that energy constraints slow the AI buildout, extend payback periods, and compress returns on invested capital. That risk is not fully priced into current valuations.
The Liquidity Angle
Liquidity is the only truth in markets. And the liquidity picture here is interesting. Capital is flowing into AI infrastructure at unprecedented rates. But capital is also flowing into energy infrastructure. The two are becoming coupled in ways that create both opportunities and risks.
If energy costs continue to rise, the marginal return on AI infrastructure investment declines. At some point, the expected return on a new data center doesn't justify the capital commitment. That's when the capex cycle turns. And when the capex cycle turns, the AI infrastructure narrative changes.
I don't know when that inflection point comes. But I know the direction of travel. Energy costs are rising, grid constraints are tightening, and the cost of building AI infrastructure is going up. The question is whether AI revenue growth can outpace the cost growth.
The Bottom Line
Here's what I'm watching. Transformer lead times. Grid interconnection queues. Energy cost trends in key data center markets. Cloud provider capex guidance. These are the leading indicators of AI infrastructure health. They matter more than model benchmark scores or AI adoption metrics.
The energy constraint is not a short-term issue. It's a structural shift that will shape the AI industry for the next decade. The companies and regions that solve the energy problem will be the winners. The ones that don't will see their growth constrained by physics.
AI data center expansion is not going to stop. The demand for compute is too strong. But the pace of expansion will be constrained by energy availability. And that constraint will reshape the competitive landscape in ways the market hasn't priced yet.
The next time you hear about a massive AI data center investment, ask one question: where's the power coming from? The answer to that question tells you more about the project's viability than any technology roadmap or revenue projection. Energy is the new compute. And the grid is the new GPU.