The Stargate AI data center in Abilene, Texas, US, on Wednesday, Sept. 24, 2025.

Kyle Grillot | Bloomberg | Getty Images

West Texas dust, iron-tinged and orange-red, rides the wind and sticks like a film to everything you touch. It clings to skin and the inside of your mouth, a fine grit that turns every breath into a reminder of where you are. This is the landscape where OpenAI CEO Sam Altman is orchestrating something called Stargate — a fast-expanding constellation of data centers, backed by partners including Oracle, Nvidia, and SoftBank.

Six thousand workers’ vehicles pour into the site each morning. Tires raise a constant veil of grit over a construction footprint the size of a small city — more people working this single campus than OpenAI employs across its entire payroll.

Rain comes in flashes. One minute the roads are powder; the next they’re mud — thick, adhesive, the kind that tugs at boots and gums up machinery. Then the storm moves on, the sun returns, and the surface hardens again, cracked and chalky, as if the place is trying to erase the evidence that water ever touched it.

And at dusk, the same conditions that make living there punishing turn the sky into a blaze. Shorter wavelengths fall away and reds and oranges remain.

“This is what it takes to deliver AI,” Altman told CNBC on site in September. “Unlike previous technological revolutions or previous versions of the internet, there’s so much infrastructure that’s required. And this is a small sample of it.”

A small sample: At roughly $50 billion per site, OpenAI’s Stargate projects add up to about $850 billion in spending — nearly half of the $2 trillion global AI infrastructure surge HSBC now forecasts.

The Abilene campus already has one data center online, with a second nearly complete. OpenAI CFO Sarah Friar told CNBC the site could ultimately scale past a gigawatt of capacity — enough electricity to power about 750,000 homes, roughly the size of Seattle and San Francisco combined.

“The shovels that are going in the ground here today, they’re really about compute that comes online in 2026,” she said in September. “That first Nvidia push will be for Vera Rubins, the new frontier accelerator chips. But then it’s about what gets built for ’27, ’28, and ’29. What we see today is a massive compute crunch.”

“We are growing faster than any business I’ve ever heard of before,” Altman said, squinting against the sun. “And we would be way bigger now if we had way more capacity.”

Land is cheap. Governments are willing. And the grid, for now, can be persuaded to bend.

Altman is not alone in building kingdoms.

Zuckerberg’s Hyperion and Musk’s Colossus

In the flatlands of northeast Louisiana, where soybean fields once stretched to the horizon, Meta‘s Mark Zuckerberg is erecting a four-million-square-foot monument to artificial intelligence. He calls it Hyperion, after the Greek titan. When finished, it will consume more electricity than the city of New Orleans — and cover a footprint the size of lower Manhattan.

Across the Mississippi River, in West Memphis, Arkansas, Alphabet‘s Google has broken ground on what state officials are calling the largest private capital investment in state history — a multibillion dollar campus rising from 1,100 acres of scrubland.

Thirty minutes south, on the Tennessee side of the border, Elon Musk has already begun transforming the industrial wastelands of South Memphis. His supercomputer, Colossus, was built in 122 days inside a shuttered Electrolux factory. Now he’s constructing Colossus 2, aiming for a million GPUs — and just acquired a third building to expand the complex further. To power the site, Musk bought a shuttered Duke Energy power plant across the border in Southaven, Mississippi.

In southeast Wisconsin, Microsoft is spending more than $7 billion on what CEO Satya Nadella calls “the world’s most powerful” AI data center — a facility that will house hundreds of thousands of Nvidia chips when it comes online in early 2026. And in rural Indiana, near Lake Michigan, Amazon has transformed 1,200 acres of farmland into Project Rainier, an $11 billion facility running entirely on custom silicon, built exclusively to train AI models for a startup called Anthropic.

“Cornfields to data centers, almost overnight,” Amazon Web Services CEO Matt Garman told CNBC in Seattle in October.

This is the AI boom rendered in steel and gravel — a slow carving of the country into zones of power and compute. What they’re building is not infrastructure in any conventional sense. It is the physical manifestation of a belief — that intelligence itself can be manufactured at industrial scale, and that whoever builds the biggest factory wins.

“This is the largest market in the history of mankind,” said Sameer Dholakia, a partner at Bessemer Venture Partners. “This is larger than oil, because everyone on the planet needs intelligence.”

The money

The sums involved have become difficult to comprehend.

The top five hyperscalers — including Amazon, Microsoft, Alphabet, and Meta — are on track to spend approximately $443 billion on capital expenditures this year. CreditSights projects that figure will climb to $602 billion in 2026 — a 36% year-over-year increase. Their analysts estimate that approximately 75% of that spending will go directly into AI infrastructure.

The current tech industry is among the most profitable in the history of the world, but not all of the companies necessarily have the cash on hand to cover the spend.

The debt raise has been staggering. Hyperscalers have added $121 billion in new debt this year — more than four times the average annual issuance over the previous five years, according to Bank of America. Over $90 billion of that came in just the past three months. Meta tapped the bond market for $30 billion. Alphabet raised $25 billion. Oracle just pulled off an $18 billion bond sale — and Citi says it now ranks as the largest issuer of investment-grade debt among non-financial U.S. companies.

Wall Street expects the pace of borrowing to accelerate.

Analysts at Morgan Stanley and JPMorgan estimate AI’s infrastructure push could drive up to $1.5 trillion in additional borrowing by tech companies in the coming years. UBS analysts forecast as much as $900 billion in new issuance coming in 2026 alone.

AWS site lead Josh Sallabedra with MacKenzie Sigalos

Katie Tarasov

“There is something inherently uncomfortable as a credit investor about the transformation of the sort we’re facing that is going to require an enormous amount of capital,” Daniel Sorid, head of U.S. investment grade credit strategy at Citi, told investors on a video call earlier this month.

You can see that discomfort in the derivatives market.

Credit-default swaps — insurance that pays out if a borrower can’t service its debt — have widened to multi-year highs for Oracle. Barclays and Morgan Stanley have told clients to buy protection, and in late October, a liquid CDS market tied to Meta began actively trading for the first time as investors rushed to hedge what’s becoming a hyperscaler debt boom.  

There’s precedent for debt-funded buildouts outrunning near-term demand. In the dot-com era, telecoms levered up to lay fiber fast. When conditions tightened, many had to restructure. The network survived — but the outcomes ranged from many early investors booking losses, to equity wipeouts.

OpenAI and the tangled web

At the center of this infrastructure arms race sits OpenAI — and a web of interlocking deals that has reshaped the competitive landscape for AI.

In the span of just two months this fall, the company announced partnerships totaling roughly $1.4 trillion in headline commitments — a figure that prompted skeptics to warn of an AI bubble and raised basic questions about whether the power, land, and supply chains exist to match the ambition.

The deals came in rapid succession.

In September, OpenAI announced a $100 billion equity-and-supply agreement with Nvidia — the chip giant taking an ownership stake in OpenAI in exchange for 10 gigawatts of its next-generation systems.

In October, OpenAI teamed up with AMD to deploy its Instinct GPUs, with the agreement structured to potentially give OpenAI a 10% stake in the chipmaker. Days later, Broadcom agreed to supply 10 gigawatts of custom chips co-designed with OpenAI. And in November, OpenAI signed its first cloud contract with Amazon Web Services, further loosening Microsoft’s once exclusive grip.

“We have to do this,” OpenAI President Greg Brockman told CNBC in October, referring to the company’s scramble to secure the raw computing power behind its ambitions. “This is so core to our mission if we really want to be able to scale to reach all of humanity, this is what we have to do.”

Nvidia is effectively financing demand for its own chips, Oracle is building the sites, AMD and Broadcom are positioning as alternative suppliers, and OpenAI is anchoring the demand. Critics call it a circular economy: capital, capacity, and revenue all recycling through the same small set of players. It works as long as growth holds — but if demand slips or funding tightens, the stress can propagate fast through a web of shared exposures.

Already, Nvidia has cautioned investors there was “no assurance” it would enter a definitive agreement with OpenAI, or complete the investment on expected terms, a reminder that headline AI pacts often start as frameworks.

Oracle’s view from the jobsite is simpler: the demand is real, diversified, and already spoken for.

“We see broad-based demand across a huge swath of the industry, so it’s not just from any one individual place,” Clay Magouyrk, Oracle’s newly elevated co-CEO, told CNBC in West Texas in September. “I don’t worry about a bubble, because I see committed demand for it.”

He described the appetite for compute as nearly limitless. “When I look at myself, when I look at my teams at Oracle, when I look at our customers, I see what looks like near-infinite demand for technology — if we can enable them to use it.”

At the DealBook Summit in December, Anthropic CEO Dario Amodei described the “cone of uncertainty” — a mismatch between long lead times and a market that can change in a quarter. Data centers take 18 to 24 months to build, and chip orders are placed years in advance, even as demand forecasts keep shifting.

“You don’t have $50 billion on you,” he said, so the financing often gets wrapped into partnerships with chipmakers or cloud providers, where “you can kind of pay as you go.”

Amodei insists Anthropic is trying to stay disciplined. “I think there are some players who are not managing that risk well,” he said, declining to share names.

The new gospel of scale

Critics question how much is firm, contracted demand versus aspirational headline math.

Gil Luria, who covers technology cycles at D.A. Davidson, points to Oracle as a test case.

“OpenAI made commitments that it’s highly unlikely they’ll be able to live up to,” he said. “Now they’re backtracking and saying these aren’t really commitments — these are frameworks. But talk to Oracle about that. Oracle thought they had a contract for $300 billion. They put that in their remaining performance obligations and made commitments to Wall Street based on that.”

Oracle stock dropped 23% in November — its worst month since 2001.

OpenAI’s Friar rejected the “circular economy” framing during the interview with CNBC in West Texas.

She compared it to the early days of the web. “When the internet was getting started, people kept feeling like, ‘Oh, we’re overbuilding, there’s too much.’ And look where we are today, right? The internet is ubiquitous. AI is going to be like that.”

Friar said equity is too expensive, so OpenAI is preparing to take on debt for the first time to finance expansion. The company has reviewed more than 800 potential sites across North America — weighing land, substations, and transmission capacity.

And like much of the industry, OpenAI is looking at every viable power source — renewables, gas, and even nuclear — as utilities and tech companies chase always-on power that wind and solar can’t reliably provide themselves.

“The real bottleneck isn’t money,” she said. “It’s power.”

That demand isn’t fading. In late December, SoftBank’s Masayoshi Son agreed to pay $4 billion for DigitalBridge, a firm that invests in data centers. To fund the deal — and his $40 billion commitment to OpenAI — Son sold down SoftBank’s entire stake in Nvidia. He later told a forum in Tokyo that he “was crying” over having to sell the shares.

The scarce asset now is energized real estate — and the ability to plug in at scale. Power like that is regulated and permitted, which means the buildout also depends on Washington.

OpenAI has lobbied the Trump administration to expand the CHIPS Act tax credit to cover AI data centers — though when its CFO floated the idea of a government “backstop” for infrastructure loans at a Wall Street Journal event in November, the backlash was swift enough that she walked it back within hours. Altman took to X to insist the company does not “have or want government guarantees.”

The companies aren’t waiting for Washington. They’re borrowing, building, and betting that the economics will catch up — because so far, every time they’ve scaled, the models have gotten better. That pattern is the industry’s founding conviction: more compute produces more capable systems. It’s why startups that have never turned a profit can still command valuations in the hundreds of billions.

The wager isn’t only that training ever-larger models will keep producing step-change intelligence. It’s that the payoff is now spilling out of the lab, as those models are put to work across the economy — answering customers, writing code, routing claims, drafting contracts, compressing weeks of work into hours. That’s inference: not training the model, but the everyday usage that turns models into products.

Inference is where the hype has to convert into margins, and it’s also where the compute bill never stops: each new user, workflow, or agent adds recurring demand, not a one-time training run. That’s why the buildout has started to look less like a moonshot and more like a utility race, with companies scrambling to secure the power and capacity to serve what they expect will be always-on intelligence.

“We have continued to be surprised, even as the people who pioneered this belief in scaling laws,” Daniela Amodei, Anthropic’s president and co-founder, told CNBC during a sitdown at the company’s headquarters in San Francisco. “Every year we’ve been like, ‘Well, this can’t possibly be the case that things will continue on the exponential,’ and then every year, it has.”

Anthropic’s revenue has jumped tenfold, year-over-year, for the last three years. In 2025 alone, the startup’s valuation surged from $60 billion to a funding round currently underway that could put it north of $300 billion.

The reckoning

Dario Amodei, Daniela’s brother, believes we are approaching something like “a country of geniuses in a datacenter” — AI systems that can perform at the level of Nobel laureates across every domain. He believes that threshold could come as soon as next year.

But he’s also sounding alarms.

“Look at entry level consultants, lawyers, financial professionals, many of the white collar service industries, a lot of what they do, AI models are already quite good at without intervention,” he told 60 Minutes. “And my worry is that it’ll be broad, and it’ll be faster than what we’ve seen with previous technology.”

That belief is driving the industry’s spending binge — but skeptics worry the buildout becomes a debt-fueled overreach, ending in a familiar cleanup: bankruptcies, fire sales, and shareholders wiped out.

Matt Murphy, a venture capitalist at Menlo Ventures and an early Anthropic investor, frames it differently.

“I’ve been in the venture business for 25 years,” Murphy said, “I’ve seen the cloud wave, the mobile wave, the semiconductor wave. This is the mother of all waves.”

Eerial shot of Open AI Stargate I (Abilene)

Courtesy: OpenAI

Stand back far enough and a new geography comes into focus.

Zuckerberg’s Hyperion. Musk’s Colossus. Altman’s Stargate. Amazon’s Rainier. Google’s archipelago of compute clusters. Each one a monument to a different vision of the future — and each one anchored to the same constraint: power.

Data centers are rising near plants and transmission lines, in places with cheap land, willing governments, and grids that can be pushed to expand. And the towns around them are now showing up in investor decks, earnings calls, and trillion-dollar projections.

Analysts tell CNBC the stakes are bigger than stock prices. Either this year marks the beginning of a transformation as profound as electrification and the internet, or it marks the peak of a bubble that future historians will study as a cautionary tale.

Altman hears the doubts — but he rejects the notion that the buildout has gone too far.

“People will get burned on overinvesting,” he told CNBC in September. “And people also get burned on underinvesting and not having enough capacity.”

“Smart people will get overexcited, and people will lose a lot of money. People will make a lot of money. But I am confident that long term, the value of this technology is going to be gigantic to society,” added Altman.

For now, the construction continues. The trucks kick up dust. The transformers hum. And across the American heartland, the factories of a new age take shape.

WATCH: Microsoft sees 10x return on OpenAI stake after restructure

The post The year AI tech giants, and billions in debt, began remaking America appeared first on Our Story Insight.