A Gigawatt of AI Data Center Now Costs $49 Billion, Says Morgan Stanley

AI Data Center Cost, in One Number

Morgan Stanley estimates that an all-in data center built in the United States for Nvidia's Vera Rubin era will cost approximately $49 billion per gigawatt of capacity. That figure comes from a July 2026 report led by analyst Brian Nowak, and it resurfaced this week because it explains almost everything happening in the cloud market right now. The comparable figure for GB200-generation facilities was about $35 billion per gigawatt. In other words, the cost of building AI infrastructure rose by $14 billion per gigawatt in a single hardware generation, and the next generation is the one everyone is building now.

A gigawatt is an enormous amount of power, roughly enough for a mid-sized city, and $49 billion is an enormous amount of money. But the number matters because it reframes the entire AI buildout. This is no longer a story about buying faster chips. It is a story about whether anyone can afford to build the buildings, buy the power, and cool the machines at the scale the AI labs say they need.

What $49 Billion per Gigawatt Actually Buys

The all-in figure covers more than servers. Land, construction, power generation and grid connection, cooling systems, networking, and the chips themselves all count, and every one of those line items has gotten more expensive. Power is the binding constraint. A Vera Rubin-era facility needs so much electricity that developers are now building their own generation: Google's new Wyoming campus, announced November 5, includes a 702 MW fuel cell development running on natural gas with carbon capture, which Google calls the largest planned development of its kind.

Cooling is the second constraint. The denser the racks, the harder they are to keep at operating temperature, and liquid cooling adds cost and complexity that air cooling never had. Networking is the third. Moving data between thousands of chips at speeds above 800 Gbps per lane needs optical interconnects, which is exactly why Marvell paid $115 billion for Lumentum. Every layer of the stack is more expensive than the last one, and the $49 billion figure is what you get when you add them all up.

The Buildout Is Slowing in Some Places and Speeding Up in Others

The cost math is already changing behavior. Microsoft declined to expand its Texas lease with Nscale in early November, passing on an option that would have more than doubled the existing contract, and Bloomberg reported the company is tightening oversight of projects from Nebius, IREN, and Nscale. The Google Wyoming campus shows the other side of the same coin: Google bought Crusoe's data center site outright and committed $17 billion to finish it, choosing ownership over leasing.

Politics is now part of the cost equation too. Voters in Texas and Georgia elected governors on November 3 who campaigned on pausing new data center construction, and the data center pause both promised sent cloud stocks jumping anyway. Strategists read the pause as good news for the companies that already own data centers, because slower supply makes existing capacity more valuable. Amazon and Microsoft each rose more than 6% the day after the election, while the companies that build data centers and supply their power sold off.

Meanwhile, the buildout continues elsewhere. Samsung SDS broke ground on November 2 on a 120 MW AI data center in Gumi, South Korea, starting with 60 MW and targeting completion in late 2028. Supermicro and SambaNova announced a collaboration to build disaggregated inference systems pairing Nvidia GPUs with SN50 RDUs, initially planning for 10,000 chips. The industry is not stopping. It is choosing where to build more carefully, and paying more wherever it does.

Who Pays When the Math Gets This Big

The $49 billion question is who ultimately pays for all of this. The cloud companies are funding the buildout with debt, equity, and cash flow, and the scale is now large enough to move their stock prices on the news. Oracle filed to raise $20 billion for its AI cloud expansion, and its shares fell nearly 9% on the announcement. The market is no longer cheering every dollar of AI spending. It wants to see the revenue that justifies it.

That is the tension behind the Morgan Stanley number. If AI demand keeps growing the way the labs project, $49 billion per gigawatt is a rational investment in the infrastructure of the next decade. If demand growth slows, or if inference turns out to be cheaper to run than training, the companies that built at the peak will own very expensive buildings full of hardware that earns less than expected. The Microsoft Nscale pullback suggests at least one giant is already hedging that bet.

For anyone buying cloud capacity, the practical takeaway is that prices are unlikely to fall soon. The cost of building the infrastructure is rising faster than the cost of the chips inside it, and those costs land in the bill eventually. The AI era was supposed to make computing cheaper. The data centers it runs on are making it more expensive, one gigawatt at a time.