We’re looking at a projected over 500% increase in global AI compute capacity by 2028, and that translates directly into a frantic build-out of new AI data centers. This pressure is completely changing the real estate game for startups and established tech giants, forcing them to scrap old site selection playbooks. The challenge now is securing the actual physical foundation for the next wave of AI development.
Key Takeaways
- Power availability, especially access to reliable green energy, is now the deciding factor for over 60% of new AI data center sites being planned for the next two years.
- Water for cooling is a huge environmental and regulatory roadblock, with states like Arizona and California putting strict new policies on its use.
- You absolutely must be near a fiber optic network with 400GbE or higher capacity for AI workloads, a requirement that now often beats out traditional factors like how much the land costs.
- Real estate in secondary cities, especially ones with old industrial bones, gives you a major cost advantage for building AI data centers compared to the usual high-priced tech hubs.
- Local government sweeteners like tax abatements and faster permitting have a direct, measurable impact on project timelines and what it costs to run the facility long-term.
| Feature | Traditional Tech Hubs | Secondary Cities | Drought-Prone Regions |
|---|---|---|---|
| Land Acquisition Costs | ✗ High | ✓ Significantly lower | Partial (historically low) |
| Proximity to 400GbE Fiber | ✓ High (often existing) | Partial (requires assessment) | Partial (requires assessment) |
| Ample Water for Cooling | Partial (variable) | ✓ Often available | ✗ Unsustainable, restrictive policies |
| Local Government Incentives | Partial (variable) | ✓ Often strong (tax abatements) | Partial (less appealing now) |
| Access to Green Energy | Partial (variable) | Partial (requires assessment) | Partial (less appealing now) |
| Permitting Expediency | Partial (variable) | ✓ Often simplified | Partial (variable) |
2.5 Gigawatts: The Energy Appetite of AI
The amount of energy these AI data centers consume is hard to wrap your head around. By 2028, AI compute alone might need an extra 2.5 gigawatts of electricity worldwide, which is enough to power millions of homes. The search now is for reliable, sustainable, and scalable power. From my own experience helping AI startups build out their infrastructure, the first question we ask about a site isn’t about land cost. It’s “What’s the power situation, and is it clean?”
Sites that have a direct line to renewable sources, like their own solar farm or a hookup to wind energy transmission, are getting a serious premium. Just look at the new data center clusters popping up in places like West Texas, where they’ve got tons of wind power and wide-open spaces. Developers are now seriously looking at on-site microgrids and even small modular reactors (SMRs) just to get power independence and stop relying on an aging utility grid. Unlike old data center models that focused almost exclusively on the initial real estate bill, this new approach puts energy resilience first.
Water Woes: A Cooling Conundrum
After you solve for power, you hit the next wall: water. Training large language models generates a ton of heat, and the cooling systems needed are intense. Traditional evaporative cooling works well, but it drinks up incredible amounts of water. In areas already facing drought, this is a total non-starter. A U.S. Geological Survey (USGS) report showed that data centers can already use millions of gallons of water a year, and that number is going to skyrocket with AI.
This simple fact is forcing developers to completely rethink their maps, pushing them toward places with plenty of water or where they can use different cooling tech. Consider what this does to states like Arizona and Nevada, which used to be data center hotspots because of cheap power and tax breaks. Now, with water use being restricted along the Colorado River basin, those spots are looking a lot less attractive for water-guzzling AI facilities. We’re seeing a lot more interest in the Pacific Northwest or near the Great Lakes, even if power costs more, because having a stable water supply is becoming a dealbreaker. Five years ago, nobody was making this kind of trade-off.
The Fiber Frontier: Speed is Paramount
AI training and inference workloads need unbelievably low latency and huge bandwidth to function which is why proximity to high-capacity fiber optic networks is a hard requirement. Standard gigabit Ethernet isn’t going to cut it. We’re talking 400 Gigabit Ethernet (400GbE) or even 800GbE to connect everything. An industry white paper from the Optical Interworking Forum (OIF) recently confirmed that data center interconnects (DCIs) at these speeds are the bare minimum for any kind of distributed AI training.
So, a site might have cheap land and all the power in the world, but if it doesn’t have a direct on-ramp to major fiber routes, it’s dead on arrival. I’ve personally seen great sites get vetoed because the cost to run new high-capacity fiber was just too high. This reality forces developments to cluster around existing internet exchange points and telecom hubs, even when the land is more expensive there. The scale and speed that AI demands makes the old “we can just build fiber out to it” argument fall apart. The budget and timeline for that kind of trenching can kill a project before it even starts. Many startups underestimate this and get a nasty surprise when they see the connectivity bill.
Secondary Cities: The New Growth Zones
While the big hubs like Northern Virginia and Silicon Valley are still data center magnets, the sky-high costs for land, power, and labor are pushing AI infrastructure builds into secondary and tertiary cities. Places like Columbus, Ohio, or Des Moines, Iowa, are suddenly looking very appealing. These spots offer much cheaper land, a recent Cushman & Wakefield report found land can be 30% to 50% cheaper than in prime markets, plus lower property taxes and easier permitting.
Many of these secondary cities also have existing industrial infrastructure, like old power substations ready for an upgrade and a workforce from manufacturing that can be retrained for data center jobs. These locations come with their own set of challenges, of course. Is it harder to convince top-tier specialized talent to move there? Sometimes. But the bottom-line economics, combined with local governments desperate for tech investment, make them a very smart play for startups trying to scale AI operations without blowing all their cash on a plot of land.
Incentives and Regulations: The Local Impact
You absolutely can’t ignore the effect of local government incentives and the regulatory climate when picking a site for an AI data center. States and towns are in a bidding war for these projects, offering everything from property tax abatements and sales tax exemptions on equipment to fast-tracked permitting. The Data Center Frontier published a study showing these incentives can knock up to 15% off the total cost of ownership over a decade. That’s real money.
Just as important is working through the local red tape. Some counties have zoning laws or environmental reviews that can bog a project down for years, while others, hungry for investment, will roll out the red carpet with “fast-track” programs. My team has burned countless hours digging into specific municipal codes because a friendly regulatory environment can be the difference between breaking ground in 18 months or waiting three years. The ease of doing business and having a predictable timeline are key, sometimes even more so than the direct financial perks. Overlooking these local details is a classic, and expensive, misstep.
Winning in the AI infrastructure race means getting site selection right, and that now requires a laser focus on power, water, fiber, and friendly local politics. The companies that figure this out first will be the ones who build the compute foundation for what’s coming next in AI innovation.
What is the most critical factor for AI data center site selection today?
Power. Hands down. Because of AI’s immense energy needs, having access to a reliable, scalable, and increasingly green power source is the top consideration, often more important than land cost or local tax breaks.
How does water scarcity affect AI data center development?
It’s a huge problem. The high-density hardware for AI generates so much heat that the cooling systems need massive amounts of water. This makes drought-prone regions less viable and forces developers to either find new locations with ample water or spend a lot more on water-efficient cooling tech.
Why are secondary cities becoming attractive for AI data centers?
They’re a bargain. Secondary cities offer much lower costs for land and taxes, and the permitting process is often way easier than in the major tech hubs. They also tend to have existing industrial infrastructure and local governments that are eager to cut deals for new investment.
What role do fiber optic networks play in AI data center site selection?
They’re a go/no-go factor. AI training and inference demand extremely low latency and massive bandwidth, so a site is simply not viable without direct access to 400GbE or faster fiber routes, no matter how good its other features are.
Can local government incentives truly impact AI data center operational costs?
Absolutely. Incentives like property tax abatements and sales tax exemptions on gear can dramatically lower the total cost of ownership for a data center over its lifetime, making a huge difference to the long-term financial viability of a project.
