OpenAI Stargate data center buildout infrastructure lead Keith Heyde

It wasn’t how Keith Heyde envisioned celebrating the holidays. Rather than hanging out with his wife back home in Oregon, Heyde spent late December visiting potential data center sites across the U.S.

Two months earlier, Heyde left Meta to join OpenAI as the head of infrastructure. His job was to turn CEO Sam Altman’s ambitious compute dreams into reality, seeking out vast swaths of land suitable for expansive facilities that will eventually be packed with powerful graphics processing units for building large language models.

“My in-between Christmas and New Year’s last year was actually mostly spent looking at sites,” Heyde, 36, told CNBC in an interview. “So my family loved that, trust me.”

His life in 2025 has only gotten more intense.

Since January, OpenAI has been quietly soliciting and reviewing proposals from around 800 applicants hoping to host the next wave of its Stargate data centers, AI supercomputing hubs designed to train increasingly powerful models.

Roughly 20 sites are now in advanced stages of diligence, with massive tracts of land under review across the Southwest, Midwest and Southeast. Heyde said tax incentives are “a relatively small part of the decision matrix.”

The most important factors are access to power, ability to scale, and buy-in from local communities.

“Can we build quickly, is the power ramp there fast, and is this something where it makes sense from a community perspective?” he said.

Heyde leads site development within OpenAI’s industrial compute team, a division that’s swiftly become one of the most important groups inside the company. Infrastructure, once a supporting function, has now been elevated to a strategic pillar on par with product and model development.

With traditional data centers nearly at max capacity, OpenAI is betting that owning the next generation of physical infrastructure is central to controlling the future of AI.

The energy needs are hard to fathom. A gigawatt data center requires the amount of power needed for some entire cities. Late last month, OpenAI announced plans for a 17-gigawatt buildout in partnership with Oracle, Nvidia, and SoftBank.

New sites will have to include all sorts of energy options, including battery-backed solar installations, legacy gas turbine refurbishments and even small modular nuclear reactors, Heyde said. Each site looks different, but together they form the industrial backbone OpenAI needs to scale.

“We’ve done this wonderful piece of bottleneck analysis to see what types of energy sources actually allow us to unlock the journey that we want to be on,” Heyde said.

A good chunk of the capital is coming from Nvidia. The chipmaker agreed to invest up to $100 billion to fuel OpenAI’s expansion, which will involve purchasing millions of Nvidia’s GPUs.

Heyde, a former head of AI compute at Meta, helped oversee the buildout of Meta’s first 100,000 GPU cluster.

In addition to power, OpenAI is assessing how quickly it can build on a site, the availability of labor and proximity to supportive local governments, according to Stargate’s request for proposal.

Heyde said the team has made around 100 site visits and has a short list of sites in late-stage review. Some will be brand new builds, and others will require conversions and refurbishments of existing facilities. Flexibility will be key.

“The perfect parcels are largely taken,” Heyde said. “But we knew that perfect wasn’t the goal — the goal for us was, number one, a compelling power ramp.”

Competition is fierce.

Meta is building what may be the largest data center in the Western Hemisphere — a $10 billion project in Northeast Louisiana, fueled by billions in state incentives. CEO Mark Zuckerberg raised the top end of the company’s annual capital expenditure spending range to $72 billion in July.

Amazon and Anthropic are teaming up on a 1,200-acre AI campus in Indiana. And across the country, states are rolling out tax breaks, power guarantees, and expedited zoning approvals to attract the next big AI cluster.

OpenAI is a relative upstart, having been around for just a decade and only known to the mainstream since launching ChatGPT less than three years ago. But it’s raised mounds of cash from the likes of Microsoft and SoftBank, in addition to Nvidia, on its way to a $500 billion valuation.

And OpenAI is showing it’s not afraid to lead the way in AI. A self-built solar campus in Abiliene, Texas, is already live.

While OpenAI still leans on partners like Oracle, OpenAI Chief Financial Officer Sarah Friar told CNBC last week in Abilene that owning first-party infrastructure provides a differentiated approach. It curbs vendor markups, safeguards key intellectual property, and follows the same strategic logic that once drove Amazon to build Amazon Web Services rather than rely on existing infrastructure.

However, Heyde indicated that there’s no real playbook when it comes to AI, particularly as companies pursue artificial general intelligence (AGI), or AI that can potentially meet or exceed human capabilities.

“It’s a very different order of magnitude when we think about the type of delivery that has to happen at those locations,” he said.

Some applicants, including former bitcoin mining operators, offered existing power infrastructure, like substations and modular buildouts, but Heyde said those don’t always fit.

“Sometimes we found that it’s almost nice to be the first interaction in a community,” he said. “It’s a very nice narrative that we’re bringing the data center and the infrastructure there on behalf of OpenAI.”

The 20 finalist sites represent phase one of a much larger buildout. OpenAI ultimately plans to scale from single-gigawatt projects to massive campuses.

“Any place or any site we’re moving forward with, we’ve really considered the viability and our own belief that we can deliver the power story and the infrastructure story associated with those sites,” Heyde said.

He understands why many people are skeptical.

“It’s hard. There’s no doubt about it,” Heyde said. “The numbers we’re talking about are very challenging, but it’s certainly possible.”

WATCH: OpenAI’s $850 billion buildout contends with grid limits