Oracle Data Centers Face Multibillion-Dollar Cost Surprises
Building an AI supercampus increasingly means paying more than you bargained for. Oracle’s attempt to salvage a proposed $165 billion project on the rocks in New Mexico is a recent example.
The cloud provider earlier this year was planning to spend billions of dollars to build its own natural-gas plants to power Project Jupiter, a 1,400-acre, two-plus gigawatt AI data center for OpenAI near the Texas border town of El Paso. But a state environmental permit application stalled over air pollution and greenhouse-gas concerns.
So in April, Oracle abruptly pivoted to powering the entire campus with Bloom Energy natural gas fuel cells, which arguably have a better shot at getting approved because they produce far fewer pollutants and slightly less carbon emissions, and they use almost no water.
The change increased Oracle’s costs. Analysts have ballparked the cost of the fuel cell microgrid, reduced in size to 2.45 gigawatts, at around $8 billion, likely a few billion more than the cost of the original gas turbines. Oracle makes up for some of that difference with better fuel efficiency from the cells.
But the change also reduces Oracle’s long-term power flexibility. Unlike gas turbines, fuel cells degrade if they don’t run continuously, which would crimp Oracle’s ability to switch to cheap solar on sunny days. New Mexico last week issued its second rejection of proposed pipeline routes to deliver fuel to the site. It is now clear that the path to a workable fuel-delivery solution could be long and costly because the state’s politics are decidedly more hostile to fossil fuels than those of neighboring Texas.
Now New Mexico’s Environment Department has announced it will hold a public hearing on the air permit Oct. 19, citing significant opposition. The New Mexico Attorney General is investigating complaints from local residents and elected officials that their names were used in letters sent to the environmental regulator in support of the fuel cells, without these people’s consent or knowledge. And Source NM, a local publication, pointed out that the fuel cells for just that facility would still emit more greenhouse gases than what the state’s two largest cities say they emit, combined.
Oracle is also in the process of discussing financing for the project that prospective investors say has more favorable terms than in its prior capital raises, although the exact terms couldn’t be learned. An Oracle spokesperson said in a statement that it’s making rapid progress on AI sites and is “confident in the returns on the capital we are deploying.”
Environmental critics are calling for a statewide data center moratorium, even though Jupiter’s construction is well underway and despite the fact that Oracle promised direct support to schools and workforce development. The company also said it will treat non-potable water that it secured outside the municipal water system to cool the data center and operate the fuel cells.
Julia Robin, the head of infrastructure planning and sourcing at Oracle’s cloud provider unit, published a letter in a local paper stating that the changes it has made to its plans show that “we’re listening and continuously improving the project.” Oracle has built excess power capacity into the design at its own expense and, when it doesn’t need it, the company would “love to send those excess electrons back to consumers in the region,” she wrote.
The extraordinary steps Oracle and other tech firms are taking to earn the social license to operate AI megacampuses reflect two things: the unprecedented resource needs of their projects and the payback they expect to earn if they can just get the locals on board. They’re arguably rich enough to swallow the costs, but they likely didn’t imagine three years ago that they would be fronting this much to upgrade local municipal water or wastewater systems or paying the entire expense of a grid upgrade.
Previously, such costs were often forced on all the electricity customers who might conceivably benefit from them.
Wisconsin Problems
Today looks a whole lot different. In Wisconsin, even though OpenAI, Oracle and Microsoft pledged to “pay their own way” by covering all power costs associated with their AI projects, the state power regulator sided with consumer advocates and sought to close some loopholes that could still leave electricity payers fronting certain costs.
The regulator’s recent ruling on transmission cost-sharing, for example, could mean Oracle, OpenAI and their development partner Vantage Data Centers would pay for the entire construction of a transmission line for their Port Washington data center campus after the companies initially expected the public to help foot some of the bill. The decision may add $100 million or more to their costs. Oracle has also sued the commission over a recent ruling that requires it to put up financial guarantees in cash or lines of credit in case of a failure, due to its relatively low credit rating compared to other tech giants. Oracle said the ruling will cost it around $100 million annually.
Not every effort to satisfy the public has to cost more. On the Texas side of El Paso, just over the border from Oracle’s Project Jupiter, Meta may have made a cost-effective choice that is also less water-intensive.
Meta’s McCloud power project is partnering with local utility El Paso Electric to install 813 gas generators from power provider ERock for a 225 megawatt microgrid while it waits for the utility to build a substation to connect a Meta data center to the grid. Meta will pay 100% of the costs for two to five years as a “bridge” to getting grid power. Then the utility, which owns the generators, is proposing to use the generators as a low-cost “swing” power source in times of peak demand. These engines pollute more than large gas turbines but don’t use water and won’t need to run continuously in the future.
Another promising example involves Google’s projects powered mostly by renewable energy in the Texas Panhandle, which could also end up being cheaper per megawatt than huge off-grid gas projects. Lazard’s widely-watched report on the long-term costs for different power sources confirmed that solar is still cost-competitive even when supplemented with batteries and gas—as Google’s Intersect subsidiary plans to do—and despite inflation in the solar supply chain.
But there’s plenty of reasons to believe we’re only beginning to see the true cost of AI data center projects come to light.
Mystery Bacteria
In Wyoming, Meta is scrambling this month to explain why a rare and deadly bacteria found its way into Cheyenne’s wastewater system after it flushed out some pipes in a cleaning procedure. Its contractors had configured the pipes for a future “closed loop” cooling system that is becoming a standard feature of AI projects. It’s still a mystery where the bacteria came from, but the municipal utility alleges it came from Meta’s site and has now prohibited the plant from discharging fluids into its wastewater network. Meta may now need to build its own collection systems.
Meta is appealing the decision, saying in a statement that subsequent, independent tests detected no bacteria and the utility hasn’t shared its own testing data with Meta or the public. “We are committed to being a good neighbor in Cheyenne, which includes helping care for the local watershed,” a Meta spokesperson said.
Launched in the past few years, closed loop water systems were considered a major improvement to AI data centers because they barely use any water once the system is filled. But they require regularly draining small amounts of water into wastewater systems, and the Meta incident could increase calls for stepped-up treatment and monitoring. Water monitoring firm Ketos said AI facilities risk surprising municipal wastewater systems by discharging water with high concentrations of corrosion inhibitors, metals, and biocides and other substances that build up as the water is reused over and over.
To be fair, Ketos suggests wastewater discharge from AI facilities is a highly solvable problem if the water is properly monitored and treated. A Meta spokesperson said its data centers “operate under local sewer use permits that define acceptable discharge parameters, and our flows will fall within those limits.”
Near-Junk Credit Rating
Even before the closed-loop question surfaced, water sourcing and treatment in drought-stricken areas was emerging as a new, non-trivial expense. As we recently reported, Lancium, a power developer working with Oracle and other tech firms, appeared at the Texas legislature to pledge it would be a good water steward by drilling for and treating non-potable water in the Texas Panhandle that wouldn’t compete with farmers’ supply.
The next concession tech firms may agree to could be pricier still: capturing carbon emissions from all the new gas they’re burning.
Do these ballooning expenses make a project only borderline attractive? So far, I still hear an optimistic answer: even though the cost to build and power a gigawatt of AI is bordering on at least $60 billion, Nvidia servers powered by its Grace Blackwell chips, renting out at $3.50 an hour, could generate around $12 billion to $13 billion a year, including operating costs. That’s a relatively quick payback, assuming the chips’ value holds up the way it has in recent years.
But based on the surprises that keep coming, it’s a good bet tech firms will need to keep padding their cost estimates.
It’s probably no coincidence that last week, S&P Global lowered Oracle’s long-term issuer credit rating to just one level above non-investment grade, otherwise known as junk. The ratings agency cited “significant upfront capital investments and long-term data center leases, both of which we have continually underestimated.”
