Okay, I am not a lawyer so I only understood about half of what just happened. But I am fairly sure, given the context, that Elon Musk’s lawyers may have just fucked up big.

Jared “James Brickhouse” Birchall, Musk’s finance guy and all-around fixer, took the stand after Musk today. Most of his testimony was dull and seemed to exist primarily to get some documents read into the record, which sucks but is a normal part of sitting through trials. But at the very end of his boring testimony something interesting happened. I believe we all got a surprise, something that rarely happens in courtrooms.

The lawyer conducting his direct examination was passed a note by another member of the team, and asked Birchall what was apparently contained in the note: was he familiar with the xAI bid for OpenAI’s assets?

“As I recall, a lawyer we were working with had asked the attorney general of California to ensure that in their fiduciary duty, proper value was being given to the assets of the nonprofit of OpenAI,” Birchall said. In his understanding, there was a negotiation “between Sam Altman and himself on both sides of the table, the for-profit and the nonprofit, attempting to discount the value of the nonprofit assets. And we made that bid in an attempt to properly account for the value the foundation had, and create a market bid that would need to be considered by the attorney general.”

Here’s some lore: in February 2025, a Musk-led coalition made a $97.4 billion bid for the nonprofit that controls OpenAI. The bid was submitted by Marc Toberoff, one of Musk’s lawyers in the current case. This bid happened as OpenAI was restructuring itself so that the for-profit arm could be cleared to go public. In Birchall’s testimony, that bid was made because Musk, Birchall, and others, thought Altman might undervalue the nonprofit as the company restructured itself. (I’m not really sure why that would be a problem for Musk and xAI, frankly, but whatever.)

The defense counsel objected, and Birchall’s rant was struck for lack of foundation. So we did this piece by piece to establish the foundation, ending with Birchall saying, again, “Sam Altman was on both sides of the table.”

On cross-examination, Bradley Wilson from Wachtell Lipton — OpenAI’s lawyers — picked the thread back up. Wilson asked how much of this Birchall had learned from sources other than lawyers. Birchall said he’d have a hard time being able to untangle that. After a few more exchanges, Wilson moved to strike all of Birchall’s testimony about the xAI bid on grounds that would not be discussed in front of the jury.

The jury got to leave early while the lawyers duked it out, and this is where it got weird. Judge Yvonne Gonzalez Rogers started asking Birchall questions herself, and it clearly was making Birchall nervous. Birchall said he doesn’t remember discussing the xAI bid with Musk or Shivon Zilis or any other principal of the Musk organization. It sure sounded like Musk’s lawyers hadn’t given OpenAI proper discovery on this topic in the depositions, and so we were doing a fast and dirty deposition with the judge right then. At one point, Gonzalez Rogers told the plaintiff’s counsel to quit coaching the witness.

Birchall said he’d spoken to the other members of the consortium about the bid, but that he wasn’t involved in discussions with Musk about when to send the bid letter. He claimed he’d heard some things from Toberoff, but that he wasn’t aware that Toberoff represented some of the other bidders. He didn’t know if xAI was aware that Toberoff represented some of the other bidders, either.

Birchall didn’t know whether other investors had firsthand information about OpenAI, he claimed. No one had documents from inside OpenAI as far as he knew. Gonzalez Rogers remained unconvinced. “I’m still struggling with how you can have conversations with these individuals to raise $97.5 billion but have no recollections even in a general sense,” she said. Birchall said he had a general sense — he called each of the people involved to see if they were interested in joining Musk on the bid.

“Why would they do that?” Gonzales Rogers asked. Birchall said these were people with whom Musk et al. had longstanding relationships. “You must have been very convincing,” she said. “You’re not very convincing today.”

Birchall said there were no numbers besides the topline one floated when he called prospective investors, and that after speaking with him, they were passed off to lawyers. He didn’t remember who chose the $97.4 billion number, and said he got it from the legal team, telling Gonzalez Rogers he didn’t get it from Musk. Gonzalez Rogers asked if that analysis was created by anyone besides Toberoff. Birchall said not that he could recall.

“Did a lawyer tell you this was part of litigation?” Gonzalez Rogers asked. 

No, Birchall said. It was strictly a business deal.

Apparently Steven Molo, who’d been defending Musk during the deposition, had made multiple objections to questions about the deal, citing privileged communications. Business deals, apparently, aren’t privileged. But all discovery into the xAI bid for OpenAI had been blocked before the trial began. Unfortunately, by asking Birchall about the xAI deal at the very end of the direct examination, Musk’s team may have opened the door for more digging into it. You may be wondering, “opened the door to what” and your guess is as good as mine. More discovery? Maybe something about anticompetitive behavior from Musk? It doesn’t sound like it’s going to be good for Musk, I can tell you that much.

Gonzalez Rogers then asked who’d passed the note, and all the lawyers just sat there like guilty children. Finally, the guy responsible said he’d passed it, but he didn’t write it; a junior lawyer did. Who wrote it? More silence. Finally Toberoff — hardly a junior lawyer — stood up and took responsibility. Why had he done it? “I thought it was appropriate.”

“Sounds like you wanted to open the door, then,” Gonzalez Rogers said. We adjourned while she said she’d consider what to do with this testimony. She will probably rule on it tomorrow.

Correction, April 30th: It is Shivon Zilis, not Sharon Zilis.

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.

Today the first witness was sworn in in Musk v. Altman: Elon Musk. I was surprised by how flat he seemed.

This is not the first time I’ve seen Musk in court. During his defamation suit, he turned on the charm and the jury responded by finding him not guilty. Today he looked adrift and unprepared. The only times he showed real animation were when he was bragging about how much he’d done for OpenAI.

The direct examination is a way of telling a story through questions; it’s important to make the narrative clear. For a suit that accuses Sam Altman of straying from OpenAI’s mission, Musk spent a weird amount of time talking about himself, recounting his biography, and hyping up the various ventures he’s undertaken that have nothing to do with OpenAI.

For instance, he told jurors that he worked between “80 to 100 hours a week,” which was how he got so much done. It is unclear to me whether his prolific posting habits count as part of the workweek. I hope the defense asks.

We did eventually get around to OpenAI, where Musk portrayed himself as the driving force. He’d been worried about AI since childhood, and had finally felt that someone needed to prevent Google from developing it. He testified that he became involved in AI safety because he had a conversation with Google’s own Larry Page and asked, “What if AI wipes out all the humans?” Page essentially shrugged — as far as he was concerned, as long as the AI didn’t also go extinct, things were all right. “I said, ‘That’s insane,’ and he called me a species-ist for being pro-human.” So OpenAI, for Musk, was born specifically to keep Google from having too much power in AI. Petty! Musk also said that after he recruited Ilya Sutskever, then a research scientist at Google, to OpenAI that “Larry Page refused to speak to me ever again.” 

What did Musk do at OpenAI? “I came up with the idea, the name, recruited the key people, taught them everything I know, provided all the initial funding. Besides that, nothing.” He paused for laughter, and one or two people obligingly chuckled. But most of the courtroom was silent. I thought he sounded petulant. “I could have started it as a for-profit and I chose not to,” Musk said.

I do wonder how much of this the jury is following. We went very quickly through a lot of ideas, including “artificial general intelligence,” an imaginary thing that many AI researchers are nonetheless afraid of. Musk defined this as being when a computer “becomes as smart as any human, arguably smarter than any human.” (Large language models are not the same as intelligence, and AGI has been defined downward for quite some time. But whatever! This case is not about that!) 

At another point, Musk was asked to explain who former OpenAI board member Shivon Zilis was. “Shivon was the, um, my chief of staff and, uh, you know,” Musk said. One person in the gallery — presumably familiar with the fact that Zilis is the mother of a few of Musk’s kids — burst out in loud laughter. But the jury looked puzzled.

During discussions of how best to get OpenAI the vast amounts of funding it would need for compute, there was indeed discussion of a for-profit arm of OpenAI with Musk. The strategy here, I think, was to make clear that Musk’s intentions were very different than the for-profit that came to pass. (That’s true! He did not get 55 percent equity in it, as one possible cap table suggested he should.) This all seemed pretty mushy, and we got bogged down in a discussion of what, in Musk’s opinion, a reasonable equity split between founders and funders would be; it’s hard to preempt the argument you are expecting without making it yourself. 

This is also kind of a distraction from the core point of the trial: Did OpenAI betray its mission statement and fool Musk into making a charitable donation? I agreed to a for-profit model but not THAT for-profit model isn’t a strong argument.

We’ll be back with more Musk testimony and presumably his cross-examination. If there’s a clearer story from the defense, this trial is effectively all over but the shouting. I’ve seen a strong performance from Musk on the stand before. Today he just didn’t seem dialed in. Maybe he’s grumpy about this trial because he knows he’s wasting his own time.

I admit, this is an innovation I did not see coming: Silicon Valley has invented the philosophical zombie from the classic thought experiment “lol how crazy would it be if there were a philosophical zombie.”

Until recently, the philosophical zombie was a concept closely associated with Australian philosopher David Chalmers, who defines it as “someone or something physically identical to me (or to any other conscious being), but lacking conscious experiences altogether.” Chalmers’ zombie twin is identical to him functionally and psychologically — except that he feels nothing. This is different from a Hollywood zombie, which has “little capacity for introspection and lack[s] a refined ability to voluntarily control behavior.”

So okay, Marc Andreessen is even shallower than our standard philosophical zombie — but still, I think there is a strong case he should be hunted and captured by the Stanford philosophy department so they can try their thought experiments out on him in real life. (Humanely, of course.) But I think for all of us who are interested in consciousness, Andreessen is certainly a specimen.

For those of you who are not as internet-poisoned as I am, let me recap: A video of Andreessen on David Senra’s podcast — podcasts being Andreessen’s favored form of self-disclosure — has been making the rounds. In the video, Andreessen cheerfully says he has “zero” levels of introspection — “as little as possible.” This is a positive for entrepreneurs, we are told. “And you know, if you go back 400 years ago, it never would have occurred to anybody to be introspective,” Andreessen says, thus setting himself up for thunderous dunking. 

I could point to the many ancient traditions of introspection (set out variously in the Bhagavad Gita, Plato, or Psalm 119) or note that his comments about a “guilt-based whammy” that showed up from Vienna in the 1910s and 1920s due to the work of Sigmund Freud suggests Andreessen has never met a Catholic (and is unfamiliar with major Catholic thought, to boot, as introspection is famously important to St. Augustine). I could post a portrait of René “I think therefore I am” Descartes. I could note that this is such a significant misunderstanding of Freud as to suggest Andreessen is totally unfamiliar with him. I am not going to do any of those things, as I believe they have already been done by internet commenters.

I am instead going to try to set out what happened, which is that I believe Andreessen read a book.

The reference Andreessen gives is The Mind Is Flat by Nick Chater, a professor of behavioral science at the Warwick Business School at the University of Warwick in the UK. Now personally I am somewhat skeptical of how serious a book is when its title references Thomas Friedman, especially when it is being written by some business school guy. Still, the basic thrust of the book — as far as I can tell — is a polemic against the notion of an unconscious mind.

I have not read the book, so I am relying on a lecture Chater gave at Google and reviews by his peers in order to summarize it; if this seems unfair to Andreessen, I am happy at his request to actually read the book and do this again in much more irritating detail. In Chater’s Google lecture, he begins by providing a series of optical illusions, demonstrating that our vision cannot work the way we believe it works. This is pretty convincing stuff! It turns out perception is very weird. (That noise you just heard was everyone who’s ever taken a philosophy 101 class yell “qualia” in unison. That’s the philosophy word for the “red-ness” of the color red, or the sensation of the smell of napalm in the morning.) 

From here, Chater proposes that the idea of an “inner self” is an illusion similar to the illusion we experience when we think we are perceiving more than one color at once. According to Chater, our minds can only do one thing at a time — that is, that we cannot have two thoughts at once, and the idea that we can is an illusion. From there, he suggests that there are no mental depths, there’s no such thing as the unconscious, and there’s no organizing principle to anyone’s mind. To the degree that an “inner self” exists, it is a kind of “tradition,” the cumulative effect of your accumulated experiences, just as a canyon exists because of the cumulative effects of water.

There are some obvious problems with this polemic, which are highlighted in the reviews. The key thing I think Andreessen is badly summarizing is an assertion Chater makes in his Google speech: The idea that the conscious mind is like the part of the iceberg above water and that there is a mysterious part of the self below the water is attributable to Sigmund Freud, and this conception of the “larger self” is wrong. There’s only the part above the water.

Here’s Andreessen’s read of Chater: “TLDR There is no inner self, you’re chasing an imaginary concept, the end.”

In one of several follow-up posts to the interview on X, Andreessen went on to something even sillier: “You’re a 15 second sliding context window with the working memory of a goldfish, your long term memory is mainly fake, and it’s a minor miracle you can get out of the door in the morning.” It’s impressive in the sense that almost every single part of Andreessen’s sentence is wrong.

I will begin, pedantically, with the goldfish, which actually have impressive memories, like, years’ worth. But let’s get to the point: “your long term memory is mainly fake.” I suppose this depends on what Andreessen means, doesn’t it? On the one hand, any cradle Catholic can do an “Our Father” or a “Hail Mary” basically at will for the rest of their lives, even if they haven’t set foot in a church in 20 years. There are other indications that memory is in some sense “real”: Small studies suggest that musical memory is preserved even in Alzheimer’s patients. Finally, most of us have attended a play, where actors recite speeches at will. How this is possible if long-term memory is fake, I don’t know. On the other hand, long-term memory can be unreliable — it is possible even to implant false memories. But “mainly fake”? I hope Andreessen stretched before that reach.

Obviously, if you’re a “sliding 15 second context window” it would be impossible to have consistent thoughts, preferences, and beliefs over time — except, of course, people do. Personality may change over a lifetime, but there is often stability for years or decades.

As for “it’s a minor miracle you can get out of bed in the morning,” well, I have to assume this reflects Andreessen’s own mornings. (Also, any parent of a toddler can tell you that gross motor skills are independent of sophisticated thought, memory, or introspection.) This is the only part of the sentence I can’t really argue with — I personally don’t have much trouble getting out of bed, but I can’t speak for him.

At this point, a reader may be tempted to point out to me that Andreessen is trolling, and I am taking him far too seriously. But unfortunately for everyone, Andreessen has waded into an interest of mine, making it intersect with my actual job. He has been posting about his theory of mind for days now — a repeated behavior, which suggests it must be important to him. 

So since we’re doing quasi-scientific polemics, I’ll throw my own in. I’ve been wondering for a while why many of our Silicon Valley overlords behave like they’ve repeatedly sustained traumatic brain injuries, and I think I’ve come to a conclusion. One thing we know from a variety of studies is that cognition doesn’t decline for some people as they age — those people tend to have stayed in school longer and tend to have greater intellectual engagement. We also know that if you don’t use one of your languages on a regular basis, you start to forget it — even if it was your first language. Writing by hand is better for memory; predictive text leads to boring writing. 

I am now going to extrapolate from this: If you stop using some skill, mental or physical, you lose it and its benefits. We’ve already seen signs of this with heavy users of AI, along with AI psychosis, which may occur because chatbots are too sycophantic. (I have pitched my editors on attempting to give myself AI psychosis, but they seem to think I am making a joke.) So I’m going to suggest that overreliance on AI is bad for you — which suggests a corollary: Whatever is happening to normal people as a result of AI overreliance has already happened to the ultra-wealthy.

Consider: I spend a lot of time standing in line (at the supermarket, at the post office, at a bakery, at a coffee shop). While I’m standing there, I can either be annoyed at how long it’s taking me to get to the front of the line or do something else: goof around on social media or retreat into my mind palace, where I might daydream, review events from earlier in the day, or think about something interesting, like whether it matters if I don’t have free will. I have a lot of opportunities to hang out in my mind palace, because I do a lot of mundane chores, like grocery shopping, laundry, and picking up after myself.

The ultra-wealthy don’t have to do any of this. They pay people to do it for them — cooking, cleaning, shopping, you name it. There are no periods of the day where they are forced to figure out how to entertain themselves in the face of repetitive chores. If you don’t use it, you lose it — so maybe a lot of them lose introspection. In fact, they can offload any cognitive activity they don’t enjoy! That’s what subordinates are for, isn’t it? And whatever gets consistently offloaded, they lose.

But, worse, they don’t notice they’re losing any skills because they are constantly surrounded by people on their payroll, or people who want something from them. People who are, you know, aggressively financially motivated to flatter them and agree with them. If that yes-man tendency in chatbots is what’s driving AI psychosis or social skill atrophy in normal people, then the same thing has already happened to the very rich.

So I am inclined to take Andreessen at his word. He’s having a little meltdown, as he sometimes does, first of all because people are making fun of him online — a thing he doesn’t encounter on a regular basis outside social media due to the yes-man phenomenon of wealth, and which is for this reason much more upsetting for him than it would be to a normal person. He has wholesale adopted Chater’s thinking, or at least a version of it, and it probably felt revelatory and correct to him because his own mind is shallow. (He doesn’t strike me as the kind of guy who’s thought about the hard problem of consciousness much, you know?) I’m not convinced he knows he was reading a polemic.

But second, I also believe him when he says he doesn’t introspect. Because we know he agrees with Chater, we can just use Chater’s definition of thought here, bracketing whatever problems that definition may have. Because introspection under Chater’s definition is just another kind of perception, Andreessen has just functionally told us that he is missing qualia. We’ve got ourselves a philosophical zombie right here, and possibly a methodology for creating them at will. Someone should get on the horn to Australia — Chalmers is going to be so excited.

I am also going to presume that in his posts, Andreessen is describing himself. But obviously there’s some overlap with a thing that Andreessen invests in, which is to say AI — which also has no long-term memory, no actual perception (and no qualia), and no ability to get out of bed. It is possible that Andreessen likes Chater because Chater suggests there is no depth to humans, which means that AI “intelligence” is very close to human intelligence, which means AI can replace humans at their jobs and Andreessen’s investments will make a boatload more money. 

So, inspired by Vanity Fair, I thought it might be fun to interview Andreessen via ChatGPT — an a16z investment — since he seems to be implying that humans aren’t much different than AI. (Andreessen’s firm is also invested in xAI, but I think asking Mecha Hitler to impersonate Andreessen is probably too rude, even for me.) I asked ChatGPT, roleplaying as Andreessen, why he didn’t introspect; it told me “I wouldn’t actually say that,” before going on at length about why people who want to be entrepreneurs should use it but also “validate everything against external feedback.” But hey, if improvising nonsense is all Andreessen does, this is as much of an answer as the real guy can give us.

So I decided to ask ChatGPTeessen another question: What did Andreessen mean when he wrote “Introspection = neuroticism x narcissism x thumbsucking”? Here’s the answer:

This answer, much like the first one, doesn’t sound stylistically like Andreessen: It’s not succinct or tendentious enough. On the other hand, it’s an improvising 15-second context window with no introspection telling me what Marc Andreessen thinks, which is what Andreessen also claims to be. And the gist does sound like his thinking, doesn’t it? If there’s no “inner self,” what’s the difference?

So I guess we’ve all learned something important: Marc Andreessen, a philosophical zombie, can be easily replaced with AI. I bet that’s great news for a16z’s investors, since it means Andreessen no longer must be paid — and it frees him up to be the subject of some philosophy experiments. (Well, unless the neuroscientists get to him first, I guess.) I look forward to finding out what it’s like to be a philosophical zombie — or maybe that it’s not like anything to be a philosophical zombie.