The system assumes infinite energy. It is the foundational axiom of the current AI buildout, a silent prerequisite buried in every hyperscaler's earnings call and every optimistic roadmap for artificial general intelligence. Code does not lie, but it does hide. And right now, the code of the AI industry is hiding a dependency that no amount of software optimization can patch: the physical grid.
Rich McCormick's recent warning about the risks of US AI data center expansion is not a contrarian take. It is a forensic observation of a system approaching a hard limit. The narrative of AI has been dominated by model parameters, benchmark scores, and the latest GPU die size. But the binding constraint has shifted. It is no longer the silicon. It is the carbon. It is the electron. The bottleneck has moved from the fab to the substation.
Over the past 24 months, I have watched the calculus of AI infrastructure change. As a DeFi security auditor, I am trained to look for the single point of failure in a system—the unchecked external call, the flawed access control list. The AI industry has a single point of failure, and it is not a smart contract. It is the transformer on a pole outside a data center in Virginia. The entire multi-trillion-dollar AI edifice is built on an assumption of infinite, cheap, and immediate power. That assumption is false.
The Physics of the Scaling Law
The article's analysis correctly identifies the core driver: the Scaling Law's insatiable appetite for compute. But it understates the physical reality. The current technical route—massive centralized pre-training on Transformer architectures—is not just computationally intensive; it is thermodynamically violent. We are not just running code; we are converting megawatts of electricity into a statistical approximation of human language.
The data points are stark. The International Energy Agency (IEA) projects global data center electricity consumption to surge from 460 TWh in 2022 to over 1,000 TWh by 2026. In the US, data centers are expected to consume 8-10% of national electricity by 2030, up from roughly 3% today. These are not incremental changes. They are step-function shifts in demand that the existing infrastructure was never designed to handle.
My own experience with energy-intensive systems comes from a different domain, but the principle is identical. In 2022, I built a quantitative risk model for the Terra-Luna collapse. The model stress-tested the mint/burn logic under varying gas fee scenarios and withdrawal constraints. The flaw was circular dependency. The system was designed to create stability from instability, and it failed because it assumed infinite demand for its own output. The AI data center buildout has a similar circular dependency. It assumes infinite energy supply to create intelligence, but the energy supply is finite, and the demand is not just for intelligence—it is for a specific, latency-sensitive, high-density form of compute.
The Architectural Autopsy: From Silicon to Substation
Let me perform an architectural autopsy on the current AI infrastructure stack. The traditional data center was a relatively benign load. Power density was 5-10 kW per rack. Cooling was an afterthought. The grid could absorb it. The AI data center is a different beast entirely. Power density has jumped to 30-100 kW per rack, and the latest GPU clusters are pushing even higher. This is not a linear increase; it is a step change in the physical demands placed on the electrical and cooling systems.
The first casualty is the grid interconnection queue. The US Department of Energy data shows that the average wait time for a data center to connect to the grid has stretched from roughly one year in 2020 to 2-4 years today. This is the new latency. It is not network latency; it is bureaucratic and physical latency. A project can have the best chips, the best algorithms, and the best team, but if it cannot get a grid interconnection, it is dead on arrival. I have seen this pattern before in DeFi. A protocol can have a brilliant design, but if it cannot get a reliable oracle feed, it is vulnerable to manipulation. The grid is the oracle for the AI economy, and it is currently being manipulated by physics.
The second casualty is the cooling system. Air cooling is insufficient for 100 kW racks. The industry is transitioning to liquid cooling—direct-to-chip and immersion cooling. TrendForce data suggests liquid cooling penetration will rise from ~10% in 2023 to over 40% by 2028. This is a massive infrastructure shift that requires new data center designs, new supply chains, and new operational expertise. It is not a simple retrofit. It is a rebuild.
The third casualty is the energy cost structure. In a traditional data center, energy accounts for 15-20% of total cost of ownership (TCO). In an AI data center, that figure jumps to 30-50%. Energy is no longer a utility cost; it is the primary variable cost. This changes the unit economics of AI. The price of a token from an API is not just a function of compute; it is a function of the price of a megawatt-hour in Northern Virginia. The market has not fully priced this in. It is a latent risk, sitting in the balance sheets of every AI company.
The Contrarian Angle: The Hidden Efficiency Offsets
The mainstream narrative is one of impending doom—a power crisis that will strangle AI. The contrarian angle, the one that the article's analysis misses, is that the system is not static. The efficiency gains are real, and they are accelerating. The article correctly notes that hardware efficiency (NVIDIA H100 to B200) and algorithmic efficiency (FlashAttention, Mixture-of-Experts) are improving. But it dismisses these as mere offsets. I see them as the market's natural hedge.
Let me be precise. The Scaling Law is not a law of physics; it is a law of economics. It holds when the cost of compute is the binding constraint. When the cost of energy becomes the binding constraint, the market will adapt. The incentive to develop more efficient models—quantization, distillation, sparsification—becomes overwhelming. The cost of a single GPT-4 class training run is estimated at $50-100 million, a significant portion of which is energy. A 20% improvement in model efficiency is worth billions in avoided energy costs. The market will find that efficiency.
Furthermore, the article's analysis ignores the potential for a technical route change. The assumption that we will continue to scale monolithic Transformer models is not guaranteed. The industry is exploring distributed training, edge computing, and even fundamentally different architectures like neuromorphic or photonic computing. These are not science fiction; they are active research areas. A breakthrough in any of these could fundamentally alter the energy curve. The probability is low in the next 3-5 years, but it is not zero. I would put the probability of a major efficiency breakthrough that significantly reduces energy demand per unit of intelligence at 15-20% over the next five years. It is a tail risk, but it is a positive tail risk.
The Geopolitics of Electrons
The article correctly frames this as a geopolitical competition. The logic is simple: compute is power, and energy is the fuel for that power. The US has a lead in AI compute, but its grid is aging. The average age of a US transformer is over 30 years. China, by contrast, has invested heavily in ultra-high-voltage transmission and renewable energy capacity. This is a structural advantage that could manifest over the next decade.
The article's analysis touches on this but misses a key point: the energy endowment is becoming a new axis of geopolitical power. The Middle East, specifically Saudi Arabia and the UAE, is leveraging its energy wealth to attract AI data center investment. They are not just selling oil; they are selling compute. This is a strategic pivot. The US is trying to maintain its lead through export controls on chips, but it cannot export control its own grid capacity. The bottleneck is domestic.
I have seen this dynamic play out in the crypto world. The Bitcoin mining industry migrated to regions with cheap, stranded energy—Texas, upstate New York, and increasingly, the Middle East. The same logic applies to AI data centers. They will go where the energy is. This will lead to a geographic redistribution of compute, with energy-rich regions becoming the new hubs. This is not a prediction; it is an inevitability.
The Market Signal: The Rise of the Energy-Adjacent Asset
For investors, the signal is clear. The pure-play AI infrastructure trade is becoming crowded and increasingly risky. The energy-adjacent trade is the asymmetric opportunity. The article's analysis identifies this, but it does not go far enough. The opportunity is not just in renewable energy or grid upgrades. It is in the entire ecosystem that supports high-density compute: liquid cooling technology, advanced power management, energy storage, and even nuclear (SMRs).
Microsoft's deal with Constellation Energy to restart a reactor at Three Mile Island is a landmark event. It signals that the hyperscalers are willing to pay a premium for reliable, carbon-free baseload power. This is not a niche play; it is a strategic necessity. The market for SMRs is nascent, but the demand signal is real. I would estimate that the total addressable market for energy infrastructure specifically driven by AI data centers will exceed $500 billion over the next five years. This is the new frontier.
The Security Blind Spot: The Grid as an Attack Surface
As a security auditor, I cannot ignore the security implications. The article's analysis mentions national security, but it does not address the grid as an attack surface. A data center is not just a consumer of energy; it is a node in a critical infrastructure network. A sophisticated adversary could target the grid, not to steal data, but to disrupt AI operations. The 2021 Colonial Pipeline attack was a ransomware event. A future attack on a grid substation serving a major AI hub could be a strategic act of war.
The concentration of AI compute in specific geographic regions—Northern Virginia, for example—creates a single point of failure. A physical attack, a cyber attack, or even a severe weather event could take down a significant portion of the world's AI compute capacity. This is a systemic risk that is not being priced. The market is focused on the upside of AI, but it is ignoring the tail risk of a grid-scale failure. Security is a process, not a product. The process of securing the AI supply chain must include the energy supply chain.
The Takeaway: The New Constraint is the New Opportunity
The AI industry is hitting a wall. It is not a wall of intelligence; it is a wall of electrons. The era of assuming infinite compute is over. The era of managing finite energy has begun. This is not a negative thesis; it is a clarifying one. It will force the industry to become more efficient, more innovative, and more geographically distributed. It will create a new class of winners in the energy sector and a new set of risks for the incumbents who fail to adapt.
The question is not whether AI will be constrained by energy. It will be. The question is who will profit from that constraint. The market is just beginning to understand this. The next bull market in tech will not be led by the companies that build the biggest models. It will be led by the companies that build the most efficient systems to power them. The grid is the new GPU. The substation is the new fab. And the energy trader is the new chip designer.
Infinite loops are the only honest voids. The AI industry is in an infinite loop of scaling, but the loop is not infinite. It is bounded by physics. The smart money is already positioning for the end of the loop. The rest are still waiting for the next GPU shipment. The signal is in the grid interconnection queue. The signal is in the PUE. The signal is in the price of a megawatt-hour. The code of the AI industry is being rewritten, and the new language is not Python. It is power. `,