The $3 Trillion Mismatch in AI

I got curious about a simple question: if everyone's racing to build AI infrastructure, what's actually the slowest part of the chain? The answer surprised me. It's not chips. It's not even the buildings. It's the wire between the building and the grid.

TSMC can ramp a new chip node in 6-12 months. A hyperscale data center can go from dirt to operational in 12-18 months if you throw enough money at it. But getting that building connected to the electrical grid? Four to eight years. Sometimes longer.

That mismatch is creating a crisis that most people don't know about yet, but they're already paying for it.

In Western Maryland, residential electricity bills jumped $18 a month after the 2025-26 PJM capacity auction. In Ohio, it was $16. These aren't people who own data centers or trade crypto or train large language models. They're just people who live near the infrastructure that does.

The reason their bills went up is straightforward: data center demand is driving capacity prices through the roof. PJM's capacity auction saw a $9.3 billion price increase, and that cost gets spread across every ratepayer in the region. The family in Cumberland, Maryland paying an extra $216 a year is effectively subsidizing Microsoft's next GPU cluster.

This is the part of the AI boom nobody's writing about. Not the chips, not the models, not the fundraising rounds. The electricity. And specifically, who pays for it when the grid can't keep up.

Here's the scale of the problem. The national interconnection queue — the line of projects waiting to connect to the grid — now totals 1,400 GW of generation capacity. That's larger than the entire installed generating capacity of the United States. There are 10,300 active projects waiting. The median time from application to commercial operation has doubled over the past two decades, from under two years to more than four.

But the national numbers, as bad as they are, obscure what's happening in specific regions. Texas is the sharpest example.

ERCOT's large-load interconnection queue — mostly data centers — went from 63 GW in December 2024 to 226 GW by November 2025. That's a 259% increase in twelve months. Data centers account for roughly 73% of those requests.

To put that in perspective: ERCOT added about 23 GW of new generation capacity over 2024-2025. The queue is asking for 226 GW. Of those 226 GW in requests, only 7.5 GW have actually been connected or approved. Another 128 GW haven't even submitted their interconnection studies yet. The gap between what's being demanded and what the grid can deliver isn't closing. It's accelerating.

Northern Virginia tells the same story from a different angle. It's the largest data center market in the world — more than 300 facilities contributing $9.1 billion a year to the state economy. Data centers already consume 25% of the region's electricity supply. And Dominion Energy, the local utility, is telling new projects to expect five to seven years for a grid connection. That timeline is expected to get worse, not better.

Think about what that means. A company can order chips, build a facility, install servers, hire staff — and then sit there, dark, waiting half a decade for the utility to connect them. The building exists. The hardware exists. The power doesn't.

Now here's where it gets interesting. Every proposed solution to this problem has its own version of the same timeline mismatch.

Small modular reactors are the favorite talking point. Google signed a 500 MW deal with Kairos Power. AWS cut three separate nuclear deals. The DOE fast-tracked 10 companies to achieve reactor criticality by July 4, 2026. But no data center in America is powered by an SMR today. The first advanced reactor — Oklo's 50 MW Aurora unit at Idaho National Lab — isn't expected until 2027 at the earliest. Scaling from one prototype to hundreds of deployed reactors is a decade-long project. SMRs are a 2030s solution being marketed as a 2025 answer.

On-site solar and batteries sound practical until you do the math. A 100+ MW hyperscale data center running AI workloads needs power 24/7 at near-perfect reliability. AWS put a 5.8 MW rooftop solar array and 2.5 MW battery on one facility. That's a rounding error against the building's actual demand. You'd need to massively oversize both the generation and storage capacity to make it work, and the upfront cost at that scale is prohibitive. Solar helps at the margins. It doesn't solve the core problem.

Grid reform is the one area where there's real progress, and it's still not enough. FERC Order 2023 replaced the old serial study process with cluster studies. PJM processed 14.3 GW of projects in 668 days under the new system, compared to five-plus years under the old one. That's a genuine improvement. But here's the catch: the reform sped up the paperwork. It did not speed up the physical construction of transmission lines, which still takes four to eight years. You can approve a project faster, but you can't string wire faster. The bottleneck just moved from the filing cabinet to the field.

Every workaround has a timeline problem. And every timeline problem ultimately means the same thing: the grid can't keep up, demand keeps growing, and the cost of that gap gets passed to ratepayers.

Goldman Sachs projects a 165-175% increase in global data center power demand by 2030. McKinsey sees US data center demand tripling from 25 GW to over 80 GW in the same period. S&P Global forecasts US data center utility power demand hitting 134.4 GW by 2030, up from 61.8 GW in 2025.

These aren't speculative numbers from AI hype merchants. These are from institutions that move capital based on their projections. And they all point in the same direction: demand is growing exponentially while supply grows linearly.

The optimists argue this is just how infrastructure works — demand spikes, utilities adapt, the market sorts it out. The ITIF published a paper titled "The United States Needs Data Centers, and Data Centers Need Energy, but That Is Not Necessarily a Problem." The argument is that history shows utilities respond to demand.

But this isn't normal demand growth. ERCOT's queue didn't grow 5% or 10% in a year. It grew 259%. The interconnection queue nationally is larger than the entire existing grid. We're not talking about incremental load growth that utilities can plan for over a decade. We're talking about a step-function increase in demand hitting infrastructure that moves on geological timescales.

The unfairness of this is what I keep coming back to. The companies driving this demand — Microsoft, Google, Amazon, Meta — are among the most profitable enterprises in human history. They're racing to build AI infrastructure because the financial returns are enormous. But the cost of upgrading the grid to support that infrastructure doesn't come out of their margins. It comes out of capacity auctions and rate cases and utility bills. It comes out of the family in Ohio paying an extra $16 a month so that a data center down the road can train a model that will generate billions in revenue for a company in Redmond or Mountain View.

There's a term for this in economics: a negative externality. The cost of AI infrastructure is being externalized onto ratepayers who get none of the upside. The data center boom is a private wealth engine running on public infrastructure, and when that infrastructure can't keep up, it's the public that pays — both in higher bills and in the opportunity cost of a grid that could have been upgraded for resilience, renewable integration, or basic reliability instead.

Silicon moves in months. Steel and copper move in years. And the bill for the gap between them lands on the kitchen table of people who never asked for any of it.

Sources: Lawrence Berkeley National Lab "Queued Up: 2025 Edition," PJM Interconnection Reform Progress Reports, Goldman Sachs "AI to Drive 165% Increase in Data Center Power Demand by 2030," ERCOT System Planning Update (December 2025), McKinsey "AI Power: Expanding Data Center Capacity to Meet Growing Demand," S&P Global US Data Center Power Forecasts, Bloomberg reporting on Dominion Energy data center wait times, FERC Order 2023.