The Foundation: Queue and Lock-in
The decisions that will most limit a company's AI over the next two years don't go through any committee. They form on their own, in two places that almost never make it onto the roadmap.
The energy that powers the model, and the vendor that hosts it: neither looks like an AI decision. Both are.
The number that reaches the board, and the wrong conclusion drawn from it
The first one arrives disguised as sustainability. Someone shows how much data centers consume, everyone makes a concerned face, and the meeting moves to the next topic.
The number is usually right. The conclusion drawn from it isn't.
You're not going to run out of energy. You're going to run out of a place in the queue.
Global consumption isn't the constraint that decides your roadmap. Nobody allocates capacity based on a planetary average. What decides is something else: according to the Energy and AI report from the International Energy Agency, nearly half of US data center capacity is concentrated in five regional clusters, and half of the data centers under construction in the country are being built inside those same already-saturated clusters.
Resource constraint and queue constraint look like the same problem from a distance. They're opposites, and the difference only shows up once it's late.
Resource gets solved with money: you pay more, you buy more. Queue gets solved with lead time, and money barely helps. Whoever arrived later waits — including your cloud vendor, who's also in line to connect new data centers to the local power grid.
The layer that stopped virtualizing
Around 2014, I was put in charge of leading the build of a private cloud in Brazil. The decision came from above and wasn't mine; my part was making it happen on top of an operation that couldn't stop.
The press at the time highlighted that you could stand up a server in days instead of months. That was true, and it was about the physical build. The gain that actually changed the business was different: whoever needed an environment started allocating their own machines by software, on the spot, without waiting on anyone.
Months, then days, then immediate. That was the trajectory an entire generation of executives lived through, and it taught a lesson that gets in the way today: that capacity shows up when you ask for it.
Now the constraint has moved down to a layer that doesn't virtualize. Substations, transformers, and environmental permits aren't provisioned by software, and their timelines aren't counted in days. The yardstick the previous generation learned — request and receive — no longer applies at the base of the stack, exactly where nobody's looking because everyone already forgot that layer existed.
The second dependency, and it lives in-house
There's a second decision with the same signature, and this one doesn't depend on any energy provider.
Consolidating vendors is usually the right call. A scattered base gives you no scale, no negotiating power, and costs more to manage than it gives back. So far, no mistake.
What consolidation charges you is optionality, and it charges silently. While everything works, the cost doesn't show up in any report. The price only reveals itself the day the company needs to leave, switch cloud providers, switch model vendors, switch data platforms — and by that day the price is already set, without anyone ever having signed off on that number.
Consolidating isn't the mistake. The mistake is consolidating without having calculated what it costs to undo.
The two dependencies, energy and vendor, share the same signature: nobody chose either one deliberately, they formed as a byproduct of correct decisions made one at a time. And both only show their real price at the worst possible moment, when there's no more time left to negotiate.
What this puts on your desk
Two questions separate who's prepared from who just looks prepared.
When capacity gets scarce, is your company at the front of the queue or behind it? That doesn't get resolved on the day scarcity arrives. It gets resolved in contracts and capacity commitments locked in months or years ahead, and most companies find out the answer too late, once the competitor has already secured the space.
If you had to switch vendors in twelve months, what breaks, what does it cost, and who does it? If nobody at the company can answer all three parts of that question, the answer is already the diagnosis: optionality was sold off without anyone ever signing that sale.
Where this fits in the rest of the building
This is the foundation of the whole building, and the reason it comes before any content floor is simple: decisions about where data lives, how models get trained, and how agents operate all depend on a physical capacity that was already committed before any of them were even discussed. Whoever decides the data floor or the strategy floor without knowing what queue the company is in, and with how much optionality left, is deciding on top of a foundation that may not exist by the time it's needed.
Where this piece's yardstick comes from
The text above uses two questions, what queue the company is in and what it costs to switch vendors, as a shortcut for a broader criterion that also organizes the rest of this series.
The first axis is the reach of the error: a capacity commitment, a vendor dependency, and an effect that has already crossed the company's front door can't be evaluated by the same yardstick, because the cost of undoing each one is a different order of magnitude. The second axis is the control left on the human side after the decision has been made, which ranges from being able to reverse it at any moment to only noticing the effect after the fact.
The rule linking the two fits in one line: the control that remains has to be greater than or equal to the commitment's irreversibility. Consolidating vendors without having calculated the cost of leaving is exactly the case where that inequality breaks — and it breaks silently, because nothing in the contract warns you.
The four reach bands, the ceiling for each one, and the questions that make the yardstick bite are published, with dates, on the method page.
The data on US data center capacity concentration comes from the Energy and AI report by the International Energy Agency (IEA), published in 2025. The private cloud build scene is from 2014, and the decision to build it came from the organization's leadership at the time, not from the author, whose role was to execute it on top of a continuously running operation.