Somewhere right now, an AI project is about to fail before it even starts—not because the model doesn’t work, but because nobody asked the facilities team what it would cost to cool it. The moment it happens usually looks the same: a late-stage quote that nobody budgeted for, arriving after every other decision has already been made. For one company, it was project-ending: $2.8 million to pipe water into a building never designed for GPU workloads.

Everything else had checked out. The strategy, the budget, the hardware. None of it mattered. You can’t negotiate with infrastructure.

This AI infrastructure planning gap is one that most organizations are just beginning to understand—and it starts with who is making the decisions. AI procurement is still largely handled within IT departments, with facilities treated as an afterthought. But you can’t install a software strategy into a building that can’t support it. Enterprise AI isn’t a technology challenge. It’s an organizational one. Here’s what it means in practice.

Where AI infrastructure planning meets physics

The first thing most organizations discover is that AI infrastructure behaves differently from traditional IT—not as a matter of degree, but kind.

Most commercial buildings aren’t designed to support the power densities required for Modern GPU clusters.  At lower densities, existing air-based HVAC systems may suffice. But as workloads scale, air cooling becomes insufficient, and liquid cooling becomes necessary. That means piping water into facilities that were built for people, not processors—with all the structural, mechanical, and cost implications that follow.

Energy demand compounds the problem. GPU clusters don’t run intermittently. They operate continuously at a sustained high capacity, producing loads that can rival those of small industrial operations. For most commercial facilities, the existing electrical infrastructure simply can’t handle the demand.

Built for people, not processors

The environment makes this harder. Unlike hyperscale data centers, most commercial and industrial facilities have evolved over decades. Buildings constructed years apart often have very different electrical capacity, cooling capabilities, and physical constraints. Even structural considerations come into play.

That last point is underappreciated. High-density compute systems are heavy, and when you introduce liquid cooling, the added weight of water becomes an engineering constraint. Enterprise AI needs a foundation that can support its weight, withstand its heat, and meet its power demands.

This is why the facilities conversation can’t happen after the procurement decision. By the time teams select the hardware and schedule delivery, the window to identify these physical constraints—and budget for them—has often already closed.

The readiness gap

The deeper problem is organizational, not physical. Many companies have decided they want to deploy AI without defining what, specifically, needs to run locally—and why.

This matters because the answer shapes every infrastructure decision that follows. For many workloads, cloud remains the most practical starting point. It abstracts infrastructure complexity, scales on demand, and doesn’t require a facilities audit.

Edge deployment—running AI locally, within your own environment—introduces significant additional complexity and should be driven by clear requirements such as:

• Latency: where real-time processing is non-negotiable and cloud round-trips introduce unacceptable delay;
• Data sovereignty: where regulatory or contractual obligations prevent data from leaving the site; and
• Operational resilience: where local uptime is independent of network connectivity

Without clarity on these requirements, organizations default to edge deployments that aren’t justified by the use case—and inherit AI infrastructure planning costs that weren’t anticipated in the original business case.

AI infrastructure planning: From one-off to repeatable

Early AI deployments tend to be treated as singular projects. As organizations move beyond those initial pilots, challenges with scale begin to surface. Each facility imposes different constraints, and what worked in one building might not translate to the next.

This is where standardized deployment becomes critical. Standardized deployment upends the traditional planning process by prioritizing infrastructure capacity before any team makes a procurement decision. They follow a consistent model that helps facilities teams assess readiness from the beginning: verifying transformer capacity, confirming structural limits, and evaluating water availability before a project reaches the point of no return.  

By using a consistent assessment process with defined requirements, the variation between facilities becomes a known variable rather than a late surprise. You find the $2.8 million problem in week one, not week twelve.  

But standardization alone isn’t enough. Before committing to physical changes, organizations need to be able to test against them. Using digital modeling tools, infrastructure and facilities teams can run scenarios together—simulating power loads, cooling requirements, and structural constraints before a single piece of hardware is ordered. What begins as a one-off deployment becomes a repeatable, predictable process that moves risk from the physical world to the digital one, where surprises aren’t as costly.

Four questions for your facilities team

Ask the hard questions first, the ones that treat infrastructure as a priority:

1. What is our actual usable power capacity, not what’s on paper, but what we can safely draw today?
2. What headroom do we realistically have for new loads?
3. What would it take (and cost) to liquid cool even a single rack?
4. If we had a $1M budget to prepare for AI, where would you invest first?

The $2.8 million pipe quote wasn’t a facilities failure. It was a process failure. And it’s one that could have been caught months earlier if the right people had been in the room.

Enterprise AI doesn’t need better technology. It needs better organizational design. The companies that figure that out first won’t just deploy AI faster. They’ll deploy it in buildings that can support it.

Thinking about Enterprise AI? Start by understanding what your environment can support. Explore how simulation, visibility, and standardized infrastructureapproaches can help you move from pilot to scalable deployment with confidence.