The Energy Wall: What AI Data Centers Can Teach Crypto About Infrastructure Bottlenecks

The numbers didn't lie, but my trust did. I learned this lesson in 2017, staring at a drained treasury contract while wondering how a line of code I had personally audited could harbor a reentrancy vulnerability I missed. Now, standing on the other side of the crypto winter, I find myself watching a different kind of audit unfold — not in Solidity, but in the steel and silicon of America's electrical grid. The warning signs have been flashing for two years. Nobody in the AI industry seems to be listening.

Rich McCormick, former Goldman Sachs executive and now an outspoken critic of AI infrastructure economics, has been sounding alarms about what he calls the unsustainable trajectory of US data center expansion. His thesis is stark: the energy required to feed artificial intelligence models is approaching the physical limits of what the American grid can deliver. The story he tells is not about model capability or algorithmic breakthroughs. It is about transformers, power lines, cooling water, and the slow, grinding reality of physical infrastructure. In crypto, we understand this language intimately. We built our own infrastructure on rented compute, on energy we never saw, on assumptions that collapsed in 2022.

Over the past seven days, the pattern has become unmistakable. Microsoft and Google have each announced multi-billion dollar power purchase agreements with nuclear energy providers. AWS has acknowledged that grid connection delays are now the primary bottleneck in its data center construction pipeline. Meanwhile, the average transformer delivery time in the United States has stretched from a matter of weeks to over twelve months. The bottleneck has moved upstream — from silicon to electrons, from chips to megawatts. This is not a prediction. It is a ledger of commitments already made, colliding with a physical world that does not move at the speed of capital.

The Architecture of Hunger

I first understood infrastructure hunger in mid-2020, when I deployed fifty thousand dollars of my own capital into an arbitrage bot on Curve Finance. The bot worked. It printed small, steady profits from the spread between stablecoin pools. But what I had built was not a trading strategy — it was a liquidity architecture, and I was renting both the compute and the capital efficiency from a protocol whose incentive structure I had not fully reverse-engineered. When the competing protocol attempted to manipulate yields through coordinated LP deposits, my strategy survived because I understood the game theory beneath the code. The lesson was not technical. It was structural: in any system, the constraint that appears most distant from the value creation is the one that will ultimately throttle it.

The AI industry is now confronting its own version of this lesson. The scaling law — the observation that model capability improves predictably with compute investment — has driven an insatiable appetite for training and inference capacity. Between GPT-3 and GPT-4, estimated single-training energy consumption grew from approximately 1.3 gigawatt-hours to roughly 50 gigawatt-hours. That is a thirty-eight-fold increase in a single generational leap. Meanwhile, power density in AI-optimized data centers has surged from the traditional five to ten kilowatts per rack to thirty, sometimes one hundred kilowatts per rack. The physical architecture of these facilities is no longer compatible with the electrical infrastructure that was designed for a different era of computing.

The International Energy Agency projects that global data center electricity consumption will more than double from 460 terawatt-hours in 2022 to over 1,000 terawatt-hours by 2026. American data centers alone are forecast to grow from approximately three percent of national electricity consumption in 2022 to between eight and ten percent by 2030. Put differently: the entire computational layer of the American economy is about to consume as much power as some of its largest industrial sectors, and the infrastructure required to deliver that power is not being built fast enough.

Order Flow: Where the Capital Actually Goes

Here is the analysis that matters. I spend my days reading order flow — watching where capital moves before price moves, where commitments are made before they are publicly acknowledged. The same analytical framework applies to infrastructure markets. The question is not whether AI data centers will expand. The question is where the energy will come from, at what cost, and who will bear the friction when supply falls short of demand.

The commercial picture is unambiguous. Energy costs now represent thirty to fifty percent of the total cost of ownership for an AI-optimized data center, compared to fifteen to twenty percent for traditional facilities. This shift transforms energy from a peripheral operational expense into the single largest variable cost in the business model. The hyperscalers — Microsoft, Google, Amazon, Meta — collectively project capital expenditures exceeding two hundred billion dollars in 2024 alone, the vast majority of which is directed at AI infrastructure. But this capital is being deployed into a market where the upstream constraint is not chip availability, which can be addressed through supply chain investment, but electrical capacity, which requires physical construction of generation, transmission, and distribution assets that cannot be accelerated through engineering alone.

The grid connection queue has become the new semiconductor bottleneck. In 2020, a data center project could expect to receive interconnection approval within approximately one year. By 2024, that timeline had stretched to two to four years, with some projects facing indefinite deferral. This is not a cyclical fluctuation. It is a structural mismatch between the pace at which capital is deploying compute infrastructure and the pace at which the grid can absorb new load. The United States Department of Energy has acknowledged that transformer lead times now exceed one year — a component that was once available on demand is now a gating factor in billion-dollar projects.

The Contrarian Read: Energy Is the New Liquidity

I built a liquidity pool, but lost my liquidity. This phrase captures the paradox of the AI data center expansion. The industry is pouring capital into compute infrastructure — the equivalent of building pools — while the underlying resource required to sustain that infrastructure — energy — is becoming increasingly scarce and expensive. The parallel to DeFi liquidity mining is not superficial. In 2021, I watched protocols issue tokens at unsustainable emission rates to attract TVL, creating the illusion of demand while subsidizing the very metric they were trying to grow. The AI industry is now doing something structurally identical: it is committing to compute capacity that assumes energy will remain available at current margins, when the physical constraints suggest the opposite.

The contrarian insight is this: the AI industry's bottleneck is shifting from compute to energy, and this shift will fundamentally reshape the geography, economics, and competitive dynamics of the global AI race. Just as I warned that post-Dencun blob data would saturate within two years and force a reevaluation of L2 economics, the energy constraint will force a reevaluation of AI infrastructure economics. The question is whether the market will price this risk in advance — as it should — or whether the narrative momentum of scaling will override the physical reality until a sharper correction forces reckoning.

The geographic dimension is already visible. Data center construction is migrating toward regions with energy surplus — Texas, Iowa, Ohio — and away from energy-constrained markets like California and the Northeast corridor. This redistribution mirrors what happened in crypto when gas fees on Ethereum mainnet forced computation toward L2s and alternative chains. The constraint is not absolute; it is relative. But the direction is clear: energy endowment is becoming a competitive moat, and nations and regions with abundant power generation capacity will command a structural advantage in the AI infrastructure race.

The Game Theory of Infrastructure Competition

In the copy trading community I founded during the bear market, I learned that trust is the scarcest resource — more valuable than any strategy, any signal, any piece of proprietary code. The same principle applies at the geopolitical level. The United States holds approximately forty percent of the world's hyperscale data center capacity, a position of structural dominance. But this advantage is being tested by a variable that no amount of semiconductor investment can neutralize: the physical grid.

China's advantage in this dimension is not incidental. Its ultra-high-voltage transmission network, constructed over the past two decades, was designed for a different era of electricity consumption — one without AI. But that same infrastructure provides a head start in absorbing the new load. Meanwhile, the Middle East — Saudi Arabia, the United Arab Emirates — is positioning itself as an energy-abundant alternative for AI data center deployment, leveraging petrodollar reserves to attract hyperscaler investment. This is not a hypothetical future. Microsoft and Google have already announced significant data center investments in the Gulf region. The energy-to-compute conversion is becoming a geopolitical instrument, and the countries that control the upstream resource will increasingly shape the downstream architecture of global AI.

Silence is the loudest audit. The silence from the AI industry about its energy footprint is itself a signal. Major hyperscalers have signed renewable energy purchase agreements worth billions of dollars, but these agreements have not prevented the construction of fossil-fuel-dependent backup capacity. The carbon-neutral narrative and the actual energy procurement strategy operate on different timelines and different accounting frameworks. This is the same pattern I observed in DeFi: the incentive structure and the stated mission diverge, and the divergence reveals the true priority. In the AI industry's case, the priority is capacity at any cost — and the cost is being socialized onto the grid, onto ratepayers, and onto the environment in ways that the industry's public narrative does not acknowledge.

Takeaway: What the Grid Will Tell Us

The signal to watch is not the next model launch or the next benchmark score. It is the interconnection queue at PJM Interconnection and ERCOT — the independent system operators managing electricity flow in the eastern and western United States. When a data center project is denied interconnection, or when the approval timeline extends beyond four years, the constraint is no longer theoretical. It is priced into the infrastructure buildout timeline, and it will eventually be priced into the compute supply that AI companies are counting on.

Art burns hot; patience burns colder. The AI industry is burning hot — deploying capital at unprecedented rates, committing to compute capacity that assumes infinite energy availability. But the physical world has its own rhythm, its own constraints, its own patience. The grid will not bend to the timeline of a product launch. The transformers will not arrive faster because the stock price demands it. And the energy required to train the next generation of models will either be available at a sustainable cost or it will not.

The question I am left with, after auditing contracts and watching liquidity pools drain and building communities from nothing, is this: when the energy wall arrives — and it will arrive — will the AI industry have built the infrastructure to absorb the shock, or will it discover, like so many before it, that the constraint it treated as a given was the one variable it never properly priced? The answer is not in the code. It is in the grid.