The latest warning from Rich McCormick, as filtered through Crypto Briefing, is a study in missing variables. It flags the energy pressure of expanding AI data centers. That warning is correct. The analysis around it, however, is incomplete. It treats the energy constraint as a headline risk, not as the fundamental operating parameter for the next decade of compute.
Observe the core problem: AI's growth curve is colliding with a physical ceiling. We are not running out of silicon. We are running out of grid. This is not a hypothetical stress test. It is the current operating reality.
The Context: The Bottleneck Has Shifted
For two years, the narrative was the GPU shortage. The market obsessed over wafer supply, memory bandwidth, and the export controls on specific accelerators. That was the 2023-2024 battlefront. The front has moved.
The real constraint for 2025 is the step-down transformer. The connection queue. The power purchase agreement. The multi-year timeline for a substation upgrade. We have shifted from a manufacturing bottleneck to a physical infrastructure bottleneck.
The IEA data is clear. Global data center power consumption is projected to jump from 460 TWh in 2022 to over 1,000 TWh by 2026. In the US, data centers are expected to consume roughly 8-10% of national electricity by 2030, up from about 3% in 2022. The power density per rack has moved from 5-10 kW in traditional facilities to 30-100 kW for AI. That is a shift in kind, not just in degree. Air cooling breaks down at those densities. You cannot simply bolt on more fans. This is a physics problem.
The Core: The Grid's Critical Latency
Let me be specific about the fault lines. The data is brutal, and it is the core of the issue.
First, the latency. The queue to connect a new data center to the US grid has grown from roughly one year in 2020 to over four years in 2024, according to the US Department of Energy. A four-year wait is an eternity in this market. It means the infrastructure you plan today is obsolete by the time it comes online. It forces firms to pre-purchase land and power capacity years in advance, which creates a secondary market in power rights that does not function efficiently.
Second, the energy cost structure. In a traditional data center, energy accounts for 15-20% of total ownership cost. In an AI data center, that number jumps to 30-50%. Energy becomes the largest variable cost, and it is only heading upward. The math here is unforgiving. If you assume a 100MW facility and a 50% energy cost share, a 30% increase in electricity prices does not just dent margins. It destroys the unit economics of the hosting service. It forces a re-pricing of the AI service.
Third, the cooling shift. The density is forcing a transition to liquid cooling. Immersion cooling, direct-to-chip. This is not a choice; it is the only way to pull heat away from a 100kW rack. TrendForce data suggests liquid cooling penetration will rise from roughly 10% in 2023 to over 40% by 2028. This adds a layer of operational complexity and water dependency that the market has not fully priced in.
In my experience auditing systems, this is where the narrative breaks. The complexity is often a veil for incompetence. The new efficiency metrics hide the real physical constraints. A firm can report an improved PUE. But if the grid cannot supply the power, that number is irrelevant. Trust is a variable, verification is a constant.
The silence in the code is the loudest warning sign. Here, the silence is the absence of any discussion about the grid connection timeline in the project pitch decks.
The Contrarian Angle: What the Bulls Get Right
It is easy to be a bear on infrastructure. But the counter-narrative has merit. The bulls are not ignoring the energy problem. They are buying the solution to it.
This energy crisis is creating a massive investment cycle in the energy sector. The demand for power generation, storage, and grid equipment is a direct result of AI. This is not just a cost center for AI. It is a profit center for the firms that can supply the power.
The tech giants are not sitting still. They are signing power purchase agreements for renewables. They are exploring small modular reactors. Microsoft's deal with Constellation Energy for nuclear power is a signal. Google's investment in SMR technology is another. These are not public relations stunts; they are hedging strategies for an energy-constrained future. They are acknowledging the constraints and securing supply.
So, the bulls are correct that the energy constraint is also an investment opportunity. The problem is the time lag. This is not a 6-month fix. The time frame for building new energy generation and grid capacity is measured in years, often a decade. The mismatch between the speed of AI deployment and the speed of energy deployment is the fundamental variable.
The bulls ignore the "energy justice" component. The load from these centers can drive up local electricity prices for residents. The environmental impact of increased carbon emissions. This is a social cost that is not in the spreadsheet. And in a bull market, the cost of the externality is often not priced in. It is a hidden liability.
The Takeaway: The New Due Diligence
The next phase of the AI trade is not about the model. It is about the grid connection. The market is currently pricing in unlimited compute. It is not pricing in the physical limits of the power grid. The data suggests the market will correct this.
The question is not whether AI will run out of chips. The question is whether it will run out of power. The due diligence for any AI infrastructure project must now include the power procurement timeline, the grid connection queue, and the long-term energy price curve.
The people who say "code does not care about your roadmap" are right. But the code does not care about your transformer shortage. The code cares about the electrons. If they are not there, the model is just a static file.
We are entering a phase where the energy variable determines the winner. The technology is now secondary to the energy procurement. The economics beats engineering in the long run. And in this case, the grid is the only arbitrage.