Oracle is unafraid when it comes to hedging big bets.

The company is making a career out of reinventing itself on a large scale, often by spending a lot of money before competitors even know what’s coming. Now, the market is paying close attention to its latest reinvention: a big push into artificial intelligence using Nvidia’s cutting-edge chips.

Oracle’s partnership with Nvidia sounds like a dream come true for Silicon Valley and Wall Street. Oracle gets the power to rent AI computing to clients like OpenAI, and Nvidia expands its reach into yet another part of the enterprise cloud.

All involved parties are working together to build the infrastructure for what Oracle Chief Technology Officer Larry Ellison calls “the most transformative technology of our time.”

We’re not just buying chips — we’re building the world’s next supercomputer network, Ellison told investors in September. This is a generational shift, and Oracle intends to lead it.

But behind the big promises and billion-dollar commitments is a quieter story that isn’t about hype or hardware, but about math.

As Oracle gets more serious about its AI goals, people are starting to wonder how much this future will really cost, and if even the fastest-growing cloud business can avoid the laws of financial gravity.

The numbers tell a story Oracle didn’t script

It was never going to be cheap for Oracle to get into artificial intelligence. Over the past year, the company has spent billions of dollars building GPU superclusters with Nvidia hardware.

These same chips are powering the generative AI boom at companies like OpenAI and Anthropic. That spending spree makes Oracle one of the most important companies in the race for AI infrastructure.

But the race is just as expensive as it is quick. This week, a recently released report said Oracle’s AI cloud margins may be much thinner than investors thought, despite all the talk of hypergrowth.

According to internal numbers in that report, Nvidia made about $900 million last quarter, but only a small part of that went to the bottom line.

Related: Major analyst drops 5-word take on market pullback

In Oracle’s overall financials, that’s a rounding error. In reality, it shows how much it costs to build an AI empire.

Even a tech veteran like Oracle is finding that scale alone doesn’t guarantee profitability. This is due to the high power needs of data centers, the lack of chips, and the pressure from customers to lower prices.

Oracle’s message for now is to be patient: Growth comes first, then profits. In a market accustomed to quick rewards, though, patience can be the hardest thing to sell.

Stargate could change the math on Oracle’s AI margins

Stargate is Oracle’s big bet: a multi-site U.S. buildout with OpenAI and SoftBank that will be able to handle gigawatts of AI capacity and rent it out on a large scale. If Oracle does what it says it will do, the project won’t just bring in more money; it will also change the cost curve that has kept margins low.

The first lever is scale. Bigger clusters mean better use, easier scheduling, and fewer hours of downtime. These are small changes that add up across thousands of GPUs.

The second lever is mix. As capacity fills up, Oracle can sell not only raw computing power but also software, storage, networking, and managed AI services that make more money.

There is also the ability to buy things. Buying a lot of Nvidia generations at once can make them more available and lower the price difference between the list price and the landed cost.

Over time, better integration among Oracle’s database, middleware, and AI tools can raise the average revenue per customer without having to hire more people or get more powerful tools.

Related: The stock market laughed, then Palantir redefined the fight

None of this happens on its own. Capex intensity, power limits, construction schedules, and problems in the supply chain can all slow down the ramp. A delayed ramp keeps the margin needle stuck.

But if Stargate meets its goals, Oracle will finally get what its AI story has been missing: the ability to use its resources more effectively.

Why Oracle’s next move matters to Wall Street

Investors are putting Oracle’s AI story to the test to see how far it will go for growth without making money. The company’s backlog numbers are huge, its partnerships are great, and its goals are hard to ignore.

But Wall Street’s patience is beginning to run out.

People who study the stock say the real question isn’t whether Oracle can sell AI capacity, but how long it will take to make money from it at acceptable margins.

Oracle has relied on hope in the short term: Cloud bookings are up 359% from last year, and executives have promised that usage rates will rise as new clusters come online. But investors have heard this story before, and not every tech company can turn scale into profits.

A few funds have started to move money out of high-capex AI stocks and into software stocks with less debt. That dynamic puts Oracle in a tough spot: It has to show that AI revenue can grow faster than AI costs.

The stakes are high. If Stargate works, Oracle will be a key part of the world’s AI infrastructure. If it doesn’t, people might remember it as the company that helped build the future but never profited from it.

Related: Palantir’s Pentagon dream just hit a classified snag