Some shows take a decade to find their center of gravity. While AI Infra Summit has been around for about that long, it seems to have found its own over the last year. The conference drew about 8,500 attendees, more than double last year’s turnout, and it was clear nobody anticipated this.

Sessions were packed, hallways were jammed, and finding the right conference room for the right keynote was nearly impossible due to a lack of signage. I started looking for the NVIDIA keynote on the “mainstage” 10 minutes before its scheduled start. I found it five minutes before the 20-minute session ended. In my defense, the mainstage was not where one would expect the mainstage to be.

Apparently, the show is moving to San Jose next year.

I don’t write any of this to complain. AI Infra Summit caught fire. These are good problems to have. If you spent time at OCP Summit in its early years, the feel here is familiar. Real energy. Real technical conversation. Organization still catching up to the crowd. This combination is a good sign, so long as the organizers make good use of the next 12 months.

I especially like how the program is built. The tracks map to the AI chain rather than to vendor categories: compute, AI datacenter, data movement, data and models, and physical AI. That framing is honest about how this industry actually works. An AI deployment is a chain running from a substation through a rack through a network fabric out to a robot on a factory floor, and the weakest link sets the pace for everything else.

So, organizing a show this way forces the entire AI value chain, the power people, the silicon people, and the robotics people into the same hallway.

We need more of this.

Where Was All the News?

I heard grumbling all week that nothing major came out of the show. Google, NVIDIA, AMD, Amazon, and Intel all took the main stage, and the flow of announcements still felt lighter than the crowd size suggested it should have been.

Three things are going on.

First, this year was bigger than vendors planned for. Marketing calendars get locked months in advance, and they were built around last year’s attendance.

Second, the old habit persists. Companies still hold their best material for their own events, where they control the stage, the narrative, and the follow-up. That instinct made sense when third-party shows were small. It makes less sense now.

Third, and this is really the first point again: vendors didn’t anticipate the coverage. I heard from more than one person that more press credentials were issued here than at the OCP Global Summit. This tracks with what I saw on the floor.

So the “no news” read is kind of backward. For me, the news is the show itself.

The Session

As part of the conference, I had the chance to moderate a session on the AI Datacenter track with Kevin Janda, corporate vice president of Cloud Operations and Innovation at Microsoft. Kevin’s been in the datacenter business a long time, and what makes him interesting is that he doesn’t separate the business case from the engineering case. Siting, power procurement, capital staging, and community relations.

He treats it all as one problem. Because it is.

We spent the session on what has quietly become the hardest part of the AI buildout in the United States. Not the chips. Not the capital. Getting a community to say yes.

Some of These Concerns Have Merit. Some Don’t.

The opposition to datacenters is built on a stack of concerns, and the industry has done itself no favors by treating the whole stack as noise.

Before getting to the concerns, it helps to size up what we are actually talking about. The US has more than 3,000 operational datacenters, with more than 1,500 more in development. Virginia leads with 685 total facilities, followed by Texas at 466, California at 331, Illinois at 262, and Georgia at 235.

Put that against the rest of the world, and the concentration is striking. The US hosts close to half of all the datacenters on the planet, on the order of ten times the count of the next country on the list. Germany and the UK sit in the 500s. China is lower still by facility count. Whatever we decide to do about datacenter growth, we are deciding it for most of the world’s capacity.

The geography of these datacenters is pretty interesting. Today, 87% of existing datacenters sit in urban areas. Of those being planned, 67% are headed for rural areas, and 39% are landing in counties that don’t currently host a single datacenter. And 75% of the planned buildout is going to the South and Midwest.

So this industry is about to introduce itself to a whole lot of communities that have never hosted one. That’s the backdrop for everything that follows.

Back to the concerns. Power consumption is the strongest point of the argument. To be clear, datacenters aren’t taking over the grid. Berkeley Lab put US datacenter consumption at roughly 4.7% of national electricity in 2024, with a projection of about 11.8% by 2030. Climbing fast? Yes. But a long way from the apocalyptic framing.

But consumption isn’t what shows up on a household bill. Market structure is. In PJM, which manages the power grid and wholesale electricity market across much of the eastern United States, capacity prices rose from $28.92 per megawatt-day in the 2024/2025 auction to $329.17 per megawatt-day for 2026/2027. And PJM’s independent market monitor attributed 63% of the increase in the 2025/2026 auction to datacenter demand. So when someone in Loudoun County says their bill went up because of datacenters, they aren’t wrong.

Water looks very different at the national level. US datacenters consumed roughly 17 billion gallons on-site in 2023, compared with agricultural irrigation, which is measured in the tens of trillions. But the national number hides what matters locally, and the Environmental Law Institute notes that two-thirds of datacenters built since 2022 sit in water-stressed regions. A national average is cold comfort to a county in the middle of a drought year.

There’s also a historical element to this that’s easy to overlook. Look at Lowell, Massachusetts, where residents sued over a datacenter expansion and the city council passed the state’s first datacenter moratorium (a judge paused the expansion in August). For context, Lowell was one of America’s big industrial cities, built around textile mills along the Merrimack River.

Those mills created jobs and prosperity. They also helped leave the Merrimack badly polluted.

That’s not to suggest a modern datacenter is anything like a nineteenth-century textile mill. It isn’t. But communities have long memories. When a large industrial operator says the environmental impact is understood and manageable, history gives people a reason to ask: how do I know you’re right?

So the concerns land differently depending on which one you pick up.

And here’s the part I think the industry keeps getting wrong: it doesn’t matter much. Even when a concern is exaggerated, dismissing it is a losing strategy. Correcting someone’s math at a planning commission meeting has never once produced a permit. The community isn’t asking to be educated. It’s asking what it gets and what happens when something goes wrong.

That question has an answer. We’ve had it for 20 years.

Quincy Is the Proof

During our conversation, Kevin pointed to Quincy, Washington, and it’s the right example because it’s the longest-running one we have. Microsoft built its first datacenter there more than two decades ago. What did that turn into?

NPR reported that datacenters now cover roughly 75% of Quincy’s property tax revenue, which funded a high school, a police station, a fire station, and a library. The local poverty rate more than halved over ten years, falling to 13% in 2023. Seasonal agricultural workers moved into year-round employment. Microsoft marked the anniversary with $210,000 in grants to local organizations and has since published Community-First AI Infrastructure commitments that put numbers behind paying its own way on electricity, replenishing more water than it uses, and adding to the local tax base.

Quincy also had advantages most communities don’t: cheap land, existing fiber, and hydropower off the Columbia River.

So Quincy isn’t a template you can stamp onto a site in Georgia. What it is, though, is a demonstration that the partnership model works when it’s real.

Partnership Is the Whole Thing

So let’s strip away the community relations rhetoric for a minute. For any of this to work, four conditions have to hold.

1. Actual partnership. The community has standing to shape the project rather than react to it.
2. Real consumption numbers, published on a schedule, whether or not they flatter the operator.
3. Committed investment, with tangible benefits and measurable results. Commitments have to land as benefits the community can see, on terms it can measure. A promise with no measurement attached is indistinguishable from a promise nobody kept.
4. A published plan for when constraints appear. Not if. When. And published means findable: posted where any resident can read it, not buried in a regulatory filing. What happens when the aquifer drops, when the substation is oversubscribed, when a drought year hits? Redesign, resequence, delay, or walk away?

Let me get concrete about the transparency piece, because it’s where most operators hide behind process. Transparency here means a small number of specific things: metered water withdrawal and consumption reported at the site level rather than rolled into a global sustainability figure, the actual load curve the facility places on the local substation, the contractual mechanism by which the operator pays for the grid upgrades its load requires, and a jobs count that separates construction headcount from permanent operations headcount.

Nobody wants to publish that fourth one, because the permanent number is small. Publish it anyway. The community already suspects the number is small, and confirming their suspicion honestly costs less than getting caught inflating it.

That’s the pattern across all four conditions. Every one trades short-term messaging comfort for long-term standing. It seems like the industry has spent two years optimizing for the first.

Delay Compounds

This isn’t theoretical. Data Center Watch counted at least 75 projects worth roughly $130 billion blocked or delayed in the first quarter of 2026 alone, roughly matching all of 2025 in three months. Network World reported estimates that 30% to 50% of large-scale capacity slated to come online this year will slip. Permitting isn’t the only cause. But community opposition has become another constraint on an industry already short on power, equipment, and skilled labor.

Here’s what seems to get missed in the coverage. Delay compounds. A project that sits through a 12-month zoning fight doesn’t restart where it stopped, because the long-lead items around it keep moving. Substation-class transformers are quoting 75 to 110 weeks, and generator step-up units 100 to 150 weeks or more. And grid connection runs on its own clock entirely, with active projects in PJM’s datacenter load-growth zones waiting three to four years.

Miss your slot in either queue, and you’re not 12 months late. You’re years late.

And the demand behind those projects doesn’t wait. Every quarter of delay is an opening for investment, innovation, and capability to develop somewhere else, including places whose values we don’t share.

Ireland already ran this experiment. Grid constraints in Greater Dublin led to a de facto moratorium on new datacenter connections, operators paused or canceled projects, and the grid operator later warned the government of a possible mass exodus of datacenters from the country. The capacity didn’t disappear. It moved.

Where This Lands

Datacenter buildouts are going to continue. Yes, they’ve already slowed. And they’ll slow further in some markets. But they aren’t stopping, because the demand driving them isn’t a fad.

So we have a choice about the next five years. We can accept that reality and manage the growth so that communities benefit from it. Or we can keep creating villains and heroes, with the technology industry playing the villain and politicians somehow cast as the hero. What world are we living in when politicians are the heroes?

The villain story may be satisfying, but it doesn’t solve anything. It doesn’t lower bills. It doesn’t protect water. It doesn’t create jobs. It creates delay.

And delay is the one thing that can close the gap those earlier numbers describe. Close to half the world’s datacenters, ten times the count of the next country on the list. That’s a commanding lead, and it reflects decisions made years ago. Keep litigating every proposed build, and it tightens. The gap doesn’t close because somebody else sprints. It closes because we stopped running.

Quincy took 20 years to become an argument. We don’t have 20 years to run that experiment again. The operators who figure out how to compress that timeline by leading with transparency and measurable commitments rather than being forced into them later are the ones still building in 2030.

The rest will be in court.