Microsoft and Oracle Are Betting the House on AI Data Centers

Microsoft plans to more than triple its data center capacity, from roughly 12 gigawatts today to over 38 gigawatts by 2032, a target bigger than New York State's entire peak electricity draw. Oracle, meanwhile, is sitting on $664 billion of contracted AI infrastructure work it hasn't delivered yet, and the Pentagon has opened talks on a $5 billion loan to keep the power, chip, and construction supply chain under both companies from buckling. Three different kinds of financial stress, one bet underneath all of them: that AI demand keeps growing fast enough to make the spending worth it.

• Microsoft is targeting more than 38 gigawatts of data center capacity by 2032, over triple its roughly 12 gigawatts today, according to Bloomberg and other reports from September 10-11, 2026.
• Oracle told investors this week it holds $664 billion in remaining performance obligations, contracted cloud and AI work not yet billed, while OCI revenue climbed 121% year over year to $7.4 billion for the quarter.
• The Pentagon has opened talks on a $5 billion loan aimed squarely at the power, chip, and construction bottlenecks now threatening U.S. data center supply.
• Microsoft says it has already turned away AI and cloud business for lack of available compute, even as one analysis found Oracle now spends more building data centers each quarter than it earns.

What exactly did Microsoft and Oracle disclose?

Bloomberg and other outlets reported September 10-11 that Microsoft is planning to grow its data center footprint from about 12 gigawatts today to more than 38 gigawatts by 2032. That's not a modest expansion. It's a number that, on its own, would draw more power at peak than New York State's entire grid handles at peak. Around the same time, Oracle used its latest investor call to disclose $664 billion in remaining performance obligations, essentially a backlog of signed cloud and AI infrastructure contracts it has agreed to fulfill but hasn't yet billed. Oracle Cloud Infrastructure revenue for the quarter came in at $7.4 billion, up 121% year over year, and the company says it has delivered more than 300,000 GPUs to AI Cloud customers so far, funded in part by a $20 billion equity sale.

RelatedBig Tech's $1T AI lease bill isn't on the books

Why does tripling data center capacity matter this much?

Because Microsoft isn't chasing a hypothetical. The company says it has already had to turn away AI and cloud business it couldn't serve, simply because it didn't have enough compute sitting ready. That's a company admitting, on the record, that demand is ahead of supply right now, not in some projected future quarter. A 38-gigawatt target isn't ambition for its own sake. It's an attempt to stop leaving revenue on the table every month the buildout lags behind what customers are asking for.

What does Oracle's $664 billion backlog actually mean?

Remaining performance obligations are contracts, not cash. A customer signs up for years of committed cloud capacity, Oracle books the total value as RPO, and then has to actually build the data centers, buy the chips, and deliver the service before it can recognize that money as revenue. $664 billion in backlog against $7.4 billion in quarterly OCI revenue is a wide gap between what's promised and what's been delivered so far. One read on that gap, and it's the less comfortable one, is that Oracle is currently spending more on constructing data centers each quarter than it's bringing in. Backlog proves demand exists. It doesn't prove Oracle can finance the buildout fast enough to collect on it before the bill for building it comes due.

Who actually gets affected here?

Enterprises trying to buy Microsoft's cloud or AI capacity are the most immediate group, since some of them are the ones being turned away right now. Taxpayers are the second group, and a less obvious one: a $5 billion Pentagon loan aimed at shoring up power, chip, and construction supply chains for data centers puts public money behind what is, at its core, a private infrastructure race. There's a broader backdrop straining the whole industry too, not just these two companies: a global DRAM and HBM memory shortage is forcing chipmakers to prioritize AI customers over everyone else buying memory, which raises costs and lead times for the exact buildout Microsoft and Oracle are racing to finish.

What it means for the market

The signal for investors watching Microsoft (MSFT) and Oracle (ORCL) isn't the same signal in both cases, and that distinction is the useful part. Microsoft turning away business is a demand signal, evidence that customers want more than it can currently deliver, but it's also a capacity-execution risk: if the 38-gigawatt buildout slips, that unmet demand doesn't wait around forever. Oracle's situation reads differently. A $664 billion backlog against $7.4 billion in quarterly revenue, paired with capex that one analysis says now exceeds what the company earns in a quarter, is a real financing-risk signal, not just a growth story. This isn't investment advice, and nobody should read a buy or sell call into it. What's worth watching is simpler: whether Oracle's backlog converts into recognized revenue on a pace that matches its spending, and whether Microsoft's capacity actually lands on the 2032 timeline it's promising.

RelatedAI Datacenter Debt Isn't Subprime. It's Something Quieter.

What happens next?

The Pentagon loan is still in talks, not approved, so the first thing to watch is whether it actually gets finalized and at what terms. Microsoft's 38-gigawatt target runs through 2032, which gives plenty of time for the number to move in either direction as real estate, power contracts, and chip supply either cooperate or don't. Oracle's task is more immediate: converting more of that $664 billion backlog into delivered, billed revenue every quarter, since backlog that never converts is just a number on a slide. All of this sits inside a wider pattern in 2026 of hyperscalers taking on unprecedented debt and leaning on government involvement to fund AI infrastructure, so Microsoft, Oracle, and the Pentagon are one visible case of something happening across the industry, not an isolated arrangement.

Our take

Strip away the AI framing and this is a fairly old story: three parties making a large, leveraged bet that future demand will justify present spending, with a government loan as the safety net if the supply chain underneath them cracks first. Microsoft's piece of it is the more defensible bet, because turning away paying customers is about as clean a demand signal as a company can produce. Oracle's piece is the one worth watching more closely. A backlog that dwarfs current revenue by nearly ninety times is either the most impressive sales pipeline in the industry's history or a sign that Oracle is underwriting a lot of future promises with today's borrowed money. Probably some of both. The Pentagon's involvement is the detail that should get more attention than it has: when the U.S. government starts discussing loans to keep private AI infrastructure supply chains from breaking, that's no longer just a corporate balance sheet story.

Original analysis by GenZTech, based on Bloomberg's reporting and Oracle's investor disclosures. Figures current as of September 12, 2026.