Larry Ellison’s betting everything on OpenAI. Will it pay off or pop the bubble?

If you want to know whether the AI bubble is bursting, there’s only one publicly traded company that will tell you: Oracle.

That’s right, the database company. Oracle has burned its boats and pivoted to AI, but not in any kind of usual way. It is not a foundation model builder like OpenAI or Anthropic, obviously. It’s not quite a neocloud, though it has entered the same bare-metal business as CoreWeave. It is a software-as-a-service company that has made an audacious bet on a very specific future version of AI as Oracle’s traditional business has gracefully declined. It is significantly older than any of its AI competitors, save Microsoft, and it has decided its future involves an enormous compute deal with OpenAI, a company that does not make money.

Whether OpenAI is good for its commitments to Oracle depends a lot on how much money it can raise and how quickly it can become profitable. The risk for Oracle is that it may be sinking a lot of money into building data centers for OpenAI, only for OpenAI to be unable to pay Oracle the $300 billion it agreed to in their contract. Oracle and OpenAI did not respond to requests for comment.

But the OpenAI play — and the pivot to AI generally — suggests a specific vision: The key place to make money isn’t training foundation models. The real money is inference, or using AI models to output results on data that isn’t in the training set. So the company has looked at some startups’ businesses and decided that they are actually just features that can be added to Oracle’s existing capabilities — which is pretty much what Oracle has been up to for the entirety of its existence.

Oracle, of course, is already an enterprise business, so it has the existing relationships and large salesforce to go out there and sell its vision, one that suggests there isn’t much room for the AI stack to fragment. Rather, it will consolidate under existing players. Oracle intends to be the dominant player in that game.

Wall Street wants to bet on AI, and it can’t bet on OpenAI because it’s not public yet. So the best way to do it now is through Microsoft and Oracle. Microsoft has a more complicated business, so it’s not a pure AI bet. Oracle, on the other hand, is cleaner. That means you can take the temperature of the entire AI boom by checking in on how many people are betting Oracle won’t repay its loans on time — those are the credit default swaps. Oracle’s stock price also reacts to assorted and sundry industry events, providing a bellwether about the AI revolution — or the AI bubble, depending on how you view it.

But there’s always a tremendous gap between vision and execution, as Oracle’s history shows.

Let’s get it out of the way: Oracle founder Larry “Bad Doggy” Ellison is out of his fucking mind. He has a short attention span, a willingness to promise things his engineers have not yet built, a tremendous ego, and a competitive drive that could power every AI data center on Earth and then some. Ellison is nominally the chief technology officer and executive chairman of Oracle, and Clay Magouyrk and Mike Sicilia are nominally the co-CEOs. But Oracle has always been the Larry show, starring Larry, even when he’s busy cheating at yacht races or whatever.

Oracle’s move to focus on AI means leaving behind the high-margin, low-growth, low-capital-expenditure database business that is Oracle’s bread and butter to jump to the low-margin, high-growth, high-capex neocloud business that Oracle has taken out $43 billion in debt to build in just fiscal 2026. Why do that? Well, according to Paul Kedrosky, a longtime VC at SK Ventures, Larry got bored.

“This is the story of Larry forever,” says Kedrosky. “Whenever he left to go sailing, he’d say, ‘This company’s not as much fun as it used to be.’ The high-level take is that the orthodox company is low-growth and high-margin and makes him feel old and uncool.”

In the 1990s, one of the reasons that Oracle became a hot property was Ellison. He was among the various futurists making predictions about what the internet would do to society. In 1996, Ellison appeared on The Oprah Winfrey Show to hype what he called “the network computer.” (As part of the appearance, Oracle promised to give a network computer to each of almost 300 kids at a primary school in Menlo Park.) This was a lightweight device, even one that could be treated as a throwaway, that would connect to applications stored online. If you are thinking, Boy, that sure sounds like a modern phone, you’re right. If you are also thinking, Boy, that sure sounds like the cloud, you’re also right.

The network computer flopped. The iPhone, which kicked off the modern era of lightweight, disposable computing devices, was introduced more than a decade later, in 2007. Oracle veered away from its bold vision of the cloud, while a true believer peeled off to form his own company: Marc Benioff, who founded Salesforce in 1999. Amazon’s AWS venture into cloud computing was in 2006, a decade after Ellison had predicted that people wouldn’t need to keep software on their own computers.

So why didn’t Oracle lead both of those revolutions, if Ellison saw them coming a decade out? Well, the iPhone was a consumer product, and Oracle made primarily enterprise databases. Oracle knew how to sell to businesses — it’s why they’d so thoroughly stomped competition such as Relational Technology Inc. and Cullinet in the first place — but Ellison didn’t know how to make consumers choose to buy things rather than get forced to use it by their employer.

The failure of the network computer also made Ellison weirdly recalcitrant about the cloud. He refused to take a second crack at the idea until 2011, even mocking it as “complete gibberish.” Oracle never really recovered from its lost lead. Despite its strong enterprise software business, it lags Amazon, Google, Microsoft, and Alibaba in market share, and is barely ahead of Salesforce. Given Ellison’s competitive streak — one of his biographies is titled Everyone Else Must Fail — this has to sting. The worst part might have been losing to one of Oracle’s biggest rivals, Microsoft.

Still, the majority of Oracle’s business, as of its most recent earnings results, is “cloud and software.” The category represented 88 percent of the company’s revenue in the three months ended February 28th, which is the third quarter in Oracle’s 2026 fiscal year. (There are also hardware and services businesses, but for our purposes, they are negligible.) The majority of that is software support, which “substantially all” customers renew every year “in order to continue to benefit from technical support services and the periodic issuance of unspecified updates and enhancements” to the applications and infrastructure they also use. That brought in a shade under $5 billion in Oracle’s third quarter. The next biggest was “cloud infrastructure,” which had revenue of about $4.9 billion.

The customer support business had zero percent growth in the third quarter. Its database and applications businesses, though very profitable, aren’t growing and may even be declining, says analyst Gil Luria of DA Davidson.

The cloud business, on the other hand, is growing. It’s an “okay business, very fast-growing with low profitability,” says Luria. “Oracle cloud has single-digit margins, maybe at best teens. But they’ve been growing it very fast.”

So when ChatGPT launched the modern era of AI hype in Silicon Valley, it was inevitable that Ellison would take an interest. By February 2025, Ellison was telling former UK Prime Minister Tony Blair that AI was “a much bigger deal than the Industrial Revolution, electricity, and everything that’s come before.” In September, Oracle “shocked the markets” with a $300 billion deal with OpenAI to build data centers, one of the largest cloud deals ever. Oracle’s move into the bare-metal business — renting out servers to AI companies — can be thought of as an extension of the cloud business. Having missed the initial run on the cloud, it seems that Ellison has decided Oracle can’t be left out this time.

In some ways, Oracle was an obvious partner for OpenAI. It’s one of the few Big Boys that isn’t trying to compete with Nvidia by building its own chips — though it does have a very close relationship with AMD. But there’s one other benefit that probably tickled Ellison, a longtime Microsoft hater who even resorted to sending private detectives to sort through Microsoft’s trash: OpenAI’s biggest partner for compute used to be Microsoft. The deal was pure, flashy, competitive Ellison — and propelled Oracle’s shares to an all-time high.

But with the deal came another big personality: Sam Altman, who has a reputation in Silicon Valley as a sociopathic liar with a people-pleasing streak. OpenAI is the Sam Altman show, as became very clear in 2023 when he was briefly deposed as OpenAI’s philosopher-king. Tying Oracle so closely to OpenAI meant that Oracle was no longer the arbiter of its own fate.

And indeed, as OpenAI soon announced a series of other massive deals, Oracle’s shares fell. Now, Oracle serves essentially as a public market proxy for betting on OpenAI’s future — for better and for worse.

This time when Ellison predicted the future, he wasn’t the sole true believer, points out Nick Patience, the AI lead at the Futurum group. “It’s a more grounded bet” than the network computer, Patience says. Microsoft’s Satya Nadella and Google’s Sundar Pichai have basically the same vision. On the other hand, Ellison is “piggybacking on Sam Altman, which is probably a dangerous place to be,” Patience notes.

Oracle’s OpenAI deal was basically kismet. Oracle had been working on a data center in Texas for Elon Musk, a friend of Ellison’s, who made an abrupt about-face when he decided his company xAI could build his own data center faster. Just as Musk left Oracle in the lurch, a LinkedIn message from OpenAI infrastructure chief Peter Hoeschele arrived in the inbox of a sales leader at Oracle, Bloomberg reported. The resulting deal was significantly larger than the one Oracle had been discussing with Musk, with options to expand it further.

To fulfill the deal, Oracle will build five very large data centers. “All told, they’ll require millions of chips and consume 4.5 gigawatts of power — more than all the homes in Chicago,” Bloomberg wrote of the deal. Oracle is planning to build them with an initial completion date in 2027, though according to Bloomberg, that has already slipped to 2028. It’s a more aggressive bet than any other major company has made on AI, and one that the less reckless — or perhaps, less desperate — Microsoft shied away from.

Oracle’s previous chief executive officer, Safra Catz, was skeptical of the financial benefits of the cloud. It had lower margins and required costly data centers, Bloomberg reported, citing employees who’d heard Catz’s reservations. She was replaced last year, shortly after the OpenAI deal, by Magouyrk and Sicilia, who previously ran Oracle’s cloud business and applications. In the announcement, Ellison, unsubtly, is quoted saying that “Clay and Mike committed Oracle’s Infrastructure and Applications businesses to AI.” Oracle had burned the boats.

OpenAI, for its part, needs Oracle for its investment-grade credit rating, notes Stijn Van Nieuwerburgh, a professor at Columbia Business School. OpenAI doesn’t have one, and couldn’t support the necessary compute buildout on its own. Effectively, OpenAI is renting Oracle’s creditworthiness.

Of the hyperscalers, however, Oracle has the lowest credit rating. It also has the greatest debt load, even before the infrastructure buildout came into play. What’s more, when I say “Oracle is building data centers,” I am doing a little sleight of hand. Unlike Google and Meta, Oracle doesn’t actually build its own data centers. It’s leasing data centers that other companies are building on its behalf. In Abilene, Texas, that’s Crusoe, with whom Oracle has signed a 15-year lease; Oracle committed to paying more than a billion dollars a year despite Crusoe’s relative lack of experience, according to SemiAnalysis., an industry newsletter.

Crusoe’s inexperience is a relatively minor risk compared to the much larger one Oracle is taking on OpenAI, which is by far the biggest customer represented in Oracle’s remaining performance obligations (RPOs), which represent how much money Oracle is slated to earn from its existing contracts. Of the $553 billion in RPOs that Oracle reported in its most recent earnings release, more than $300 billion is OpenAI. So how’s OpenAI doing?

In November 2025, Sam Altman sent a memo to OpenAI in which he wrote, “I expect the vibes out there to be rough for a bit.”

The vibes had been rough for a bit before the memo, actually. There was the whole thing where Altman got booted and then reinstated. Meta raided OpenAI for talent; several key OpenAI players such as Ilya Sutskever, Bob McGrew, and Mira Murati went on to found their own companies. Anthropic, one of OpenAI’s biggest competitors, is also composed of former OpenAI talent. The executive reshuffling is basically constant.

That’s not all on the corporate chaos front. Because OpenAI is trying to go public, it had to repeatedly renegotiate its deal with Microsoft. Not only is Microsoft freed up to partner with other companies, like Anthropic — the revenue-sharing agreement between the two companies ends in 2030 (rather than whenever OpenAI hits a development milestone) and the total payments are capped. “OpenAI products will ship first on Azure, unless Microsoft cannot and chooses not to support the necessary capabilities,” Microsoft announced. But OpenAI can now partner with other cloud providers to try to build its enterprise business.

OpenAI’s approach to AI is decidedly unfocused, especially in comparison to Anthropic. There’s something of a profit panic as the company tries to figure out how to make money from its scattershot AI enterprises, while Claude Code and Cowork from Anthropic emerged as the winners for enterprise AI spending. Anthropic doesn’t have image generation or video generation products. It has instead stayed laser-focused on the enterprise market.

OpenAI has, by contrast, chased the consumer market, exposing itself to other kinds of risk in the process. The company is facing multiple lawsuits from people who say ChatGPT encouraged loved ones to kill themselves — and, in some cases, others. This has, perhaps predictably, resulted in political pressure.

Also there’s a lawsuit from Elon Musk over OpenAI’s for-profit arm that — if Musk wins — may threaten its public offering.

This degree of chaos does not inspire confidence in Altman’s management skills. And OpenAI, like all AI labs, is a money furnace. Recently, OpenAI projected that it will spend $665 billion by 2030 — $111 billion more in cash burn than it previously predicted. That’s not all. Its gross profit margins last year were lower than the company predicted, as it had to buy last-minute compute to meet demand. OpenAI projects it will be cash-flow positive in 2030, two years later than its rival Anthropic. Both companies are threatening to go public this year.

The thing that’s driving up OpenAI’s costs is inference, the very thing Ellison is betting on. This is probably a positive sign for Ellison’s intuitions about AI use, but it might not be the best thing for his partner. OpenAI has promised $1.4 trillion in its contracts.

OpenAI recently raised $122 billion. “That could last them a few years,” Luria says. “I am 100 percent sure they can get to $1.4 trillion? Probably not, but they do have money now, and that makes a difference for Oracle.”

Luria’s skepticism is understandable. Take Stargate, the flashy data center project that OpenAI announced, Altman standing shoulder to shoulder with Ellison and Donald Trump. The joint venture hasn’t hired staff and isn’t developing any data centers for OpenAI, The Information reported in February, describing Stargate as a “shelved idea.” Earlier this month, The Information discovered several Stargate leaders, including Hoeschele, ditched OpenAI; they washed ashore at Meta.

OpenAI didn’t get its planned 10GW of data center capacity from Oracle and SoftBank last year, either. Part of the problem for OpenAI is that its credit wasn’t as good as Oracle’s — so OpenAI just made its deal with Oracle directly. The two companies also made an unusual arrangement where if there was a delay or the project came in over budget, OpenAI and Oracle would share costs. (They also both benefit if things are under budget.)

As for the other member of the project, SoftBank has its own OpenAI agreement, which has led to slapfights over who controls the 1GW facility in Milam County, Texas. OpenAI has signed a long-term lease with subsidiary SoftBank Energy, which would develop and own the data center. Between the complications with Oracle and SoftBank, it seems likely that OpenAI doesn’t have the money or the compute to achieve all of its ambitious goals.

Despite all that, Altman is targeting a public offering by the end of 2026, The Wall Street Journal reported in January. That may be in doubt now, after the Journal also reported OpenAI missed revenue and user growth targets. “Chief Financial Officer Sarah Friar has told other company leaders that she is worried the company might not be able to pay for future computing contracts if revenue doesn’t grow fast enough,” the Journal wrote, citing anonymous sources. In response, Dealbook analyst Harrison Rolfes issued a note explaining that OpenAI likely won’t go public this year.

OpenAI is racing to beat Anthropic and Musk’s xAI, which has been subsumed by SpaceX and is trying to IPO this summer. SpaceX is seeking a valuation of $1 trillion or more. OpenAI’s most recent funding round gave it a valuation of more than $850 billion. Granted, valuations are more art than science, but it’s also the case that if OpenAI were to IPO with that valuation, it would be valued at 28 times its projected 2026 revenue. By way of comparison, Nvidia — a company that is making an actual profit from the AI boom — is valued at 12 times its projected 2026 revenue. OpenAI’s rich valuation, the company’s long way to profitability, and its aim at consumers rather than enterprises may lead some investors to sit out the IPO, The Information reported.

If and when the mandatory paperwork associated with an IPO filing is made public, we’ll get a sense of exactly how challenging the environment is for OpenAI — but it seems that OpenAI doesn’t have an easy road to do everything it’s trying to do and still stay solvent.

OpenAI is a flaky partner for Oracle. While that’s the largest challenge for Oracle’s data center buildout, it’s not the only risk. Payments on both Oracle’s bonds and its data center leases will occur on a fixed timeline. Anything that slows the buildout threatens Oracle — money will be flowing out on schedule, but if the build doesn’t happen on time, there may not be enough money flowing in.

There are a lot of things that could potentially slow Oracle’s mad dash to build data centers: increasing objections from communities near data centers, supply chain risks from Trump’s war on Iran, and energy risks for the same. What’s more, Oracle has data centers in the now-destabilized Middle East, which could lead to surprise costs that have nothing to do with OpenAI, but nonetheless make it harder for Oracle to pay its bills.

Increasingly, communities are objecting to the mad rush to build AI data centers, so much so that 11 states are considering moratoriums. In the case of one of Oracle’s attempted data centers in Doña Ana County, New Mexico, the attempt to quickly build a data center by buttering up local officials — without even really consulting the community the facility would be built in — is now facing several lawsuits from a local environmental group. The data center, dubbed Project Jupiter, would emit more greenhouse gases than the state’s two largest cities, Albuquerque and Las Cruces, combined. The state’s land commissioner has rejected an application for a segment of gas pipeline to power the data center. Though construction has already started, it’s still awaiting two air quality permits — and the decision, with an original deadline of April 22nd, has been delayed several months to allow for a public hearing as opposition to the data center has mounted.

Another Oracle-OpenAI data center, in Port Washington, Wisconsin, has similarly drawn pushback from locals. Several protesters were arrested at a city council meeting in December, and one was dragged out for chanting “Recall” at the mayor. Construction on this data center is also underway. Among the people who oppose the data center is the comedian Charlie Berens, who has 3 million subscribers to his YouTube channel of mostly Midwestern humor. This data center project also faces lawsuits from locals, including a challenge to tax incentives for the project worth nearly half a billion dollars. An investigation is now taking place about whether meetings that pushed the data center development forward violated open records laws; Port Washington has also been accused of not turning over public records in response to a request. The construction itself, going around the clock to avoid delays, has also irritated neighbors; new rules limit construction time.

There are other signs that the OpenAI alliance may be shaky. OpenAI declined to expand the Abilene, Texas, data center it partners with Oracle on, possibly because it doesn’t have the newest clusters of Nvidia chips. It wasn’t just OpenAI who didn’t want to work with Oracle on this; lenders didn’t want to finance an expansion with Oracle as the tenant, according to The Wall Street Journal. Banks have reportedly grown wary of Oracle debt as private credit investors have gotten anxious about their funds. Should Oracle require more money, it may be harder to find.

And now that Ellison’s pal Donald Trump has started a war in Iran with no end in sight, new risks are stacking up for data centers broadly — including Oracle’s. The New York Fed has said that supply chains are facing mounting pressure. In particular, Iran’s blockade on the Strait of Hormuz is a threat to the global helium supply; helium is used in manufacturing semiconductors, and there is nothing that can replace it. There is also an aluminum crisis; the material is used in data center server racks and cooling units. Data centers had already driven up the price of aluminum, while also making it more difficult to manufacture aluminum in the US by increasing energy prices.

Speaking of energy prices, there is one more obvious problem that has been created by the Iran war: more expensive energy. Infrastructure damage from the war, along with the closure of the Strait of Hormuz, has sent prices up without spurring new drilling. That may mean increases in the price of other kinds of energy as people seek out alternatives. Expensive energy could turbocharge objections to data center buildouts, as well as making those builds more pricey. It may also force purveyors of compute to raise their prices, which could make AI even more expensive than it already is.

On top of all of that, Oracle’s existing data centers are also in greater jeopardy than before Trump went to war. A Dubai data center has been hit by debris already. Iran claims to be proactively targeting data centers from US companies, including Oracle, and has struck several Amazon facilities. Should Iran take out more Oracle hubs, the company may be forced to rebuild those at the same time as it attempts to build its AI centers — leading to a cash crunch. It may also limit revenue.

An increase in the cost of construction and energy, a decrease in revenue, or both could make it much harder for Oracle to meet its lease and debt obligations, which are fixed. Repayment of the $43 billion Oracle raised in fiscal 2026 is split up over a series of years, starting in 2029 and ending in 2066, with a total effective interest rate of 4.9 percent.

In the next five years, Oracle will have to repay the $9.5 billion it borrowed in fixed-rate notes, plus another $500 million in floating-rate notes, plus interest. This suggests an aggressive view of what the company can accomplish in five years, since we know Oracle isn’t making enough cash right now to cover its operating expenses and capital expenditures. To turn that around, Oracle has to build data centers, fast, to turn its theoretical future revenue into actual money.

The bond market has reflected uncertainty about Oracle’s plans. In December, Oracle’s investment-grade notes were trading like junk bonds, because investors feared data centers would be delayed, according to Bloomberg. Also in December, Oracle’s credit default swaps (CDS) — a kind of insurance in the case of a default — got expensive. But by February, investors were feeling better after newly announced financing plans suggested that Oracle would avoid having its credit downgraded. After Oracle’s strong earnings were announced in March, the five-year CDS got cheaper — then, later that month, hit an all-time high, suggesting investors were nervous again.

“Oracle’s CDS has become the credit market’s proxy for AI risk,” John Lloyd, global head of multisector credit and a portfolio manager at Janus Henderson Investors, told Bloomberg.

In April, another $14 billion of bonds for an Oracle data center were issued in a special purpose vehicle, keeping the debt off Oracle’s balance sheet.

Investors have reason to be nervous. More than half of the data centers scheduled to be built this year may be delayed by equipment shortages — or even canceled. While some builders — like Oracle’s partner Crusoe — are refurbishing old transformers or relying on other strategies, uncertainty around the AI buildout has been rising.

It’s not all bad news for Oracle. For instance, ByteDance has been renting chips from Oracle to circumvent export prohibitions of Nvidia’s most advanced chips. According to The Information, ByteDance has become one of Oracle’s largest cloud customers. Oracle also has a $2 billion stake in TikTok’s newly spun-off US operations and hosts all of the company’s user data.

What’s more, SemiAnalysis has suggested that in addition to deals in Northern Virginia, ByteDance is a major customer of Oracle’s in Southeast Asia. As ByteDance is planning to grow in Southeast Asia, Europe, and Latin America, Oracle will benefit, according to SemiAnalysis. “The scale of the Oracle and Bytedance partnership remains under the radar,” SemiAnalysis noted, rating Oracle’s GPU service as Gold on its most recent ranking chart.

Oracle had strong results in its most recent earnings, too. The company did better than expected at keeping its costs low. It also showed strong growth in its cloud infrastructure business, and 90 percent of its database projects were on or ahead of schedule. “A strong record of on-time delivery is evidence of solid execution,” wrote Luke Yang, an analyst with the financial firm Morningstar. Still, Yang said that there was a lot of uncertainty around Oracle, since the AI landscape changes quickly.

But more significant than the bare-metal business may be Ellison’s vision of private AI, deployed within databases Oracle already runs. Sure, Oracle has talked about efficiencies from using AI coding tools. That’s not really the big play, though.

Oracle already has sensitive data for a number of businesses, including healthcare records. Having an AI software stack means being able to deploy AI agents into that data to better organize it — with fewer concerns about leakage than there would be with general-purpose third-party LLMs. “Training AI models on public data is the largest, fastest-growing business in history,” Ellison said in December, on an earnings call. “AI models reasoning on private data will be an even larger and more valuable business. Oracle databases contain most of the world’s high-value private data.”

There’s reason to believe that training is no longer going to be the same kind of growth industry for AI bare-metal providers; inference will be. After all, one line of thinking goes, the big LLMs have already scraped everything available on the web. But that doesn’t really matter for the application of AI to businesses — inference is what they’d want anyway. Maybe they don’t have everything a business might need yet, says Patience, the analyst with Futurum. But it’s clearly where Ellison is heading.

“I don’t think he’s early this time,” Patience said, suggesting this moment was unlike the network phone. “A lot of people would have to be completely wrong, so he’s more protected.”

Oracle is also a go-to vendor for the Trump administration, Patience points out. Oracle has, for instance, just won a contract with the Centers for Medicare and Medicaid Services to modernize the agency’s data. It also just won a similar contract with the Air Force. And now US government customers can use a number of Oracle services, including its generative AI.

This has, perhaps understandably, freaked people out. Oracle has a history of unauthorized data collection — in 2024, it settled a class-action lawsuit claiming that the company illegally compiled “digital dossiers” including where people browsed online, bought gas, banked, and ate, and sold the information to marketers. The same year, Ellison also suggested, in an analyst meeting, that AI and surveillance will make sure that “citizens will be on their best behavior because we’re constantly recording everything that’s going on.”

He described a world in which police officers, wearing Oracle body cams, tell their cameras they need two minutes to go to the bathroom. “We’ll turn it off,” Ellison says. “The truth is, we don’t really turn it off. What we do is, we record it so no one can see it. No one can get into that recording without a court order.” And an AI is always watching, Ellison says. “These are the kind of next-generation systems we can build using AI.”

And of course, databases are central to the vision. “We need to unify all the national data, put it into a database where it’s easily consumable by the AI model, and then ask whatever question you like,” Ellison said in another speech at the World Governments Summit. “Right now, countries’ data is fragmented.” Ellison’s prophecy is, effectively, government by database. AI tools and government contracts may make it competitive with Palantir, the current AI standard-bearer in assembling government data so the secret police can stalk their victims.

This vision is, obviously, bad news for democracy, but it’s great news for Oracle! I find myself curious about whether the company will simply use its private enterprise data to help add, hmmm, efficiencies to its government efforts, making it easier for surveillance-minded authoritarians to track citizens. Oracle has been wooing authoritarian governments, including China, and has suggested that “pretty much every government is going to want a sovereign cloud and a dedicated region for that government.” And because Oracle is so boring, most people may not even recognize it as a threat.

Oh, and there’s one more thing. Besides the risk of Oracle snooping into your business, there’s also the possibility of regulatory capture — that is, because Ellison is so tight with the Trumps, what remains of the government watchdogs won’t stop it. That means there’s no one to prevent Oracle doing assorted dirty deeds — but presumably not dirt cheap. This may create some downside risk if, say, Democrats ever win back power, but perhaps Ellison is betting that if he deploys his technology correctly, that simply won’t happen.

Which brings me, finally, to Oracle’s core competency: lock-in. A lot of companies remain on Oracle databases because it is difficult and expensive to relocate. If Oracle’s inference is good enough, the company basically becomes Hotel California for anyone who’s put data there — because to leave is to leave the inference behind.

Despite some fairly serious risks to Oracle — largely from its OpenAI deal, and to a lesser extent from the war in Iran — the company may be positioned to succeed. The degree to which you have faith in Ellison’s vision is also the degree to which that vision is disquieting.

Oracle’s AI buildout doesn’t necessarily make a lot of financial sense; Ellison may royally piss off his shareholders by the time this is all through. But making financial sense has never been Ellison’s strong point. And besides, what’s he going to do, let Microsoft beat him to the hot new technology?

In Oracle’s most recent earnings, the AI buildout shows up on the balance sheet as capital expenditures. Its most recent earnings were for the third quarter of fiscal 2026, which ended in February. Oracle spent $39 billion on capital expenditures, more than three times as much as in the previous fiscal year. As a result, the company now has negative free cash flow.

Oracle also told investors that it expected to spend a total of $50 billion in capital expenditures in 2026, and that it was forecasting $67 billion in revenue. In fiscal 2027, Oracle expects $90 billion in revenue. The company did not say how much it forecast in capital expenditures.

Morgan Stanley analysts suggest Oracle will need “$100 billion or more for 2027 and the first half of 2028,” according to The Wall Street Journal.

“Oracle is about execution right now,” says Luria. “The number one thing is the ability to build data centers and deploy capital to create data centers.” The thing to watch for is whether Oracle can get financing.

But even if everything goes smoothly on the building front, Ellison still has to deal with Sam Altman. If OpenAI’s chaos gets too out of hand, it may suck in Oracle, too.

What happens to Oracle if OpenAI shits the bed? One possibility is that it sacrifices its software stock premium and gets priced like a utility, which is effectively what the bare-metal business is. Like the telecoms from the ’90s internet boom, it (and its bellwether status) fades in significance as the AI-native companies that survive an AI bubble bursting eventually emerge from the wreckage to reshape our society however that may go. It’s not impossible that the company goes bankrupt, if enough things go wrong at the same time. Because for Oracle to be the dastardly surveillance company of Larry Ellison’s dreams, it has to nail the timing. And that’s never been his strong suit.