If you use generative AI, you know that it can seem like magic. Chatbots and multimedia models can effortlessly conjure up poems or high-res videos at the snap of a finger.

But AI models’ speedy outputs and sleek interfaces mask the enormous amount of physical infrastructure behind them—and as AI continues to grow, the data centers and power plants that AI is built on are starting to get widespread attention outside of the industry.

Earlier last week, I took a train to Orangeburg, New York, a sleepy lower Hudson Valley suburb just 25 miles from Fortune’s newsroom in downtown Manhattan. I was there to visit one of a wave of AI infrastructure projects popping up across the country—Orangeburg is the future home of data center company DataBank’s newest site, named LGA3.

DataBank already operates two data centers in the New York metro area: one in Newark, New Jersey, and one in Chelsea, Manhattan. But LGA3 will be by far its biggest site—a $250 million, 200,000-square-foot facility drawing up to 45 megawatts of energy to power five massive data halls packed to the gills with computer chips.

The facility won’t open until next year, but tenants have already been booking space—most notably New Jersey-based AI startup CoreWeave, which recently secured an eye-watering $19 billion valuation and has already reserved almost half of LGA3’s capacity.

“The explosive growth in artificial intelligence has required a complete reevaluation of traditional data centers to meet demand for next-generation compute requirements, and this new data center campus provides some of the most advanced new technologies that will allow us to deliver for our customers,” CoreWeave founder Ben Venturo wrote to me in a note.

My taxi from the train station took no less than four wrong turns as we wound our way past farmhouses and office parks to the LGA3 construction site, wedged between an electrical substation and the New Jersey state line. Hopping out of the car, my first impression was the sheer size of the building. LGA3 looked about the size of a New York city block, a massive, single-story hall with high ceilings—I wouldn’t be surprised if you could fit a commercial jet inside.

“Addicted to technology”

DataBank CEO Raul Martynek greeted me on the way in, wearing clear-rimmed glasses and a lavender button-down. Martynek has been in the internet infrastructure industry for decades, almost since the advent of the commercial internet in the 1990s. He’s been with Databank since 2017, overseeing the company’s 69 data centers across the U.S. and U.K. Martynek told me that he hasn’t seen an explosion in demand for digital infrastructure like the one AI is creating since the dot-com bubble of the late ’90s.

“Humans are addicted to technology, period. And ultimately, for the data center sector, what we do is we enable humans to deploy more technology,” Martynek told me. “If you deploy more technology, you need more fiber, more cell towers, more data centers. And for this particular phase that we’re in with AI, data centers are the bottleneck.”

A huge increase in demand from AI has catapulted data centers into front-page headlines. Martynek explained to me that most things we do online nowadays—from accessing images on our phones to scrolling social media to prompting ChatGPT—involve physical hardware more than we realize. Wi-Fi routers and cell towers are constantly sending signals through underground fiber optic cables to data centers and remote servers, accessing stored information and keeping the internet humming.

“The internet is a network, right? Information gets sent out over fiber optic cables as photons. And they travel around the world at close to the speed of light,” Martynek said. “I was hanging out with a network guy last night saying, ‘What do you do?’ He said, ‘We’re plumbers, right?’”

And these days, being a plumber is a good business. Exponential increases in the amount of data being generated for and by the internet over the past 20 years—and expectations that AI will only speed things up even more—mean that space to store all that information is in high demand.

“This device didn’t exist before 2007,” Martynek told me, pointing to his iPhone. “So think about how much content and how many applications have been created [by it.] All that stuff ends up in a data center…That’s the physical ecosystem.”

AI boosts need for building space

Data centers might not be the sexiest projects, but a surge in demand from AI companies is bringing in big money, heavy press coverage, and some of the biggest names in construction. DataBank alone has spent around $4.5 billion on data center projects since 2016. Tishman Speyer, the real estate company building LGA3, is one of the highest-profile names in the business: it worked on the World Trade Center and Chicago’s John Hancock Center, and it owns Rockefeller Center, too. A low-slung data bank next door to a suburban Little League baseball complex might seem an odd addition to its portfolio, but it’s betting that data centers will prove to be just as important as skyscrapers.

When I first got the invite to visit, the location surprised me. New York? Home to some of the highest real estate and energy prices in the country? Wouldn’t it be cheaper to build this in the middle of the desert, where land is cheaper and there’s access to bottom-dollar renewable energy?

But Martynek explained that for many customers, it’s just not practical to be located thousands of miles away from one of the most important parts of your business. New York is one of the country’s largest data center markets, with around 800 megawatts of capacity currently online, much of it catering to finance and tech companies who depend on nearby computing capacity to build and trade around the clock.

“It’s not practical for a data center to be in the middle of nowhere—there’s too much latency,” Martynek said, referring to delays in the response time between computers and offsite data centers. “Too many things can happen along the way.”

“Data centers have tended to cluster around metropolitan areas,” he continued. “New York has always been a pretty big data center market. That’s really a function of the population and a function of the businesses—if you’re JPMorgan, you don’t want your data center in Omaha.”

Public policy is also a factor. New York Governor Kathy Hochul unveiled the state’s Empire AI initiative earlier this year, which earmarked over $400 million to fund, among other things, infrastructure such as data center projects.

Donning a hard hat and reflective vest, I walked around the half-built structure with two construction managers. They pointed out the airplane hangar-sized area where the banks of computer chips would eventually be installed, along with the ventilation and water cooling to keep them from overheating.

Once construction is finished up, CoreWeave and DataBank’s other customers will start installing their chips, and DataBank expects the facility to be up and running in full by early next year. Once it’s online, CoreWeave will start leasing out its computing capacity to tech startups and other AI companies. Martynek told me DataBank hasn’t had any trouble finding customers.

“We signed the contract with CoreWeave last year. This building didn’t even exist then—it was just dirt. That’s how in-demand this product is,” Martynek said. “There’s a frenzy.”

As the saying goes, strike while the iron’s hot—DataBank is already putting together plans for another site right next door, LGA4. Next time you’re driving around town, keep an eye on the nondescript buildings in your area: The AI data center boom might be closer than you think.