Why a profitable Oracle scares investors

The puzzle

Profitable, cash-rich, and down 56%

In fiscal 2026 (to May 2026) Oracle earned $17.1B of net profit and held $31.3B of cash. Yet its share price fell about 56% from its September 2025 peak, because it also owes $130B.

In 2025 it was the hot AI stock: the video says it rose 87% and the company was briefly worth about $900B. So how does a money-making company become the scary one? The answer starts in the 1970s.

The problem

Three desks, three copies of Warren

In 1970s banks, every department kept its own files. Warren moves from Goa to Mumbai and tells the savings desk. His loan letters and card bills keep going to Goa. Asking "which Mumbai customers have a loan and a card?" meant weeks of custom programming.

This wasn't just an IT headache. The loan desk couldn't see Warren's healthy savings, so it never offered him a home loan. If he started missing card payments, it could still approve a loan and lose money.

The building block

One shared table every desk reads

In the late 1970s Larry Ellison, Bob Miner and Ed Oates (with Bruce Scott) shipped the first commercial SQL database and named it Oracle. Every desk now reads one record, so Warren is one customer. Banks, governments and big companies bought it.

The old business model

Customers bought machines, Oracle sold copies

A bank that bought Oracle also bought its own servers, built air-conditioned rooms, paid the power bill and hired engineers. Oracle only shipped software. Making one more copy costs almost nothing, so every new customer was nearly pure profit. That was the rule for three decades.

The rival idea

Amazon rented machines by the hour

In 2006 Amazon Web Services did the opposite. It bought millions of computers so customers needed none. Before, Warren's online store had to guess: buy for 10,000 users and crash at a million, or buy for a million and pay for idle machines. The cloud let him rent servers like electricity, paying only for what he used.

A two-person startup suddenly had a giant company's computing power with no upfront spending. The video names Airbnb, Slack and Pinterest as companies that grew this way.

The hidden gold mine

Renting machines turned out hugely profitable

Bezos kept AWS's numbers private for years, so rivals underestimated it. In 2025 AWS alone made $45.6B of operating profit, more than double Oracle's $20.6B in fiscal 2026.

Amazon first reported AWS's numbers separately in 2015, nine years after launch. On a full-year basis AWS's 2025 margin was 35% ($45.6B on $128.7B of sales); the video's 39% is a recent quarter.

The miss

Oracle laughed at the cloud, then paid

Larry Ellison openly mocked cloud computing: "all it is is a computer attached to a network." Old customers stayed because moving databases is costly and slow, but new companies started in the cloud. By fiscal 2016 Oracle's new licence sales fell 15% in a year.

By then Amazon, Microsoft and Google were the cloud giants. Oracle had cloud products before 2016, but the video's point stands: it took the cloud seriously late.

The jackpot

RPO is just an order book

A builder that signs contracts for 10 buildings hasn't built or earned anything yet. The promised money is its order book. Oracle calls it remaining performance obligations (RPO). On 9 Sep 2025 it reported $455B. Analysts expected about $180B.

Where did roughly $300B extra come from? The Wall Street Journal reported that OpenAI agreed to rent about $300B of computing power over about five years, starting in 2027. Oracle didn't name the customer.

The stretch

A $60B-a-year promise from a $12B company

$300B over five years is about $60B a year. At the time, OpenAI's revenue was about $12B a year, and Oracle's own operating cash flow was $21.5B. OpenAI promised Oracle alone five times its revenue, every year. Both companies were stretching far beyond their size.

How the market reacted

The stock paid for money not yet earned

The next day, 10 Sep 2025, Oracle stock jumped about 36%, its best day since 1992. Larry Ellison's fortune rose more than $100B and he briefly passed Elon Musk as the world's richest person. Oracle hadn't received a single dollar from the deal: share prices pay for what investors expect a company to earn tomorrow.

The flip

From copying code to pouring concrete

To rent computing power to OpenAI, Oracle needs land, buildings, power and expensive Nvidia chips first. Software kept about 70¢ of each dollar as gross profit (the video's figure). Oracle says its AI data centres will keep about 30–40¢ over a contract's life.

Margin is only half the change. Software needs no upfront factory. A data centre costs a fortune before the first rent arrives, and its chips wear out.

The cash gap

Spending ran $29B ahead of cash

In fiscal 2026 Oracle spent $55.7B on construction and equipment, then $28.5B more in the next three months. Over those 15 months its business brought in $55B of operating cash. The gap was filled with new bonds, plus $19.9B of new shares sold in mid-2026.

The mechanism

Build now, earn later: the J-curve

The video's toy case: spend $100 building something that earns $30 a year in rent. Your running total starts at −$100 and climbs $30 a year. You're underwater for over three years, and the line traces the letter J.

Investors ask three uncomfortable questions. What if the customer pays a year late, while interest on the borrowed $100 keeps running? What if rent drops to $20, pushing break-even from about 3 years to 5? What if the chips are outdated in four years and must be replaced?

Try it

Bend the J-curve yourself

You borrow $100 to build at year 0 and pay interest on it every year. Rent arrives at the end of each year once the customer starts paying. Toy assumption: $50 of the $100 is chips, re-bought each time they age out. The dashed line is the video's simple case: $30 rent, no interest, no delay.

The warning signs

The chips moved faster than the concrete

Doubts grew about whether OpenAI could pay $60B a year. In Dec 2025 Bloomberg reported some OpenAI data centres slipping to 2028; Oracle denied it. In Mar 2026 OpenAI dropped plans to expand Oracle's flagship Abilene, Texas site. It wanted newer Nvidia chips, in other locations.

By September 2026 the stock was about 56% below its peak. A data centre is built for one generation of chips. If the customer would rather have the next one, the building risks being out of date before it pays back.

The other side

The optimists have real numbers too

OpenAI closed a $122B funding round at an $852B valuation in March 2026. The video says its annualised revenue raced from $5.5B to nearly $70B. Oracle's order book kept growing, to $664B by August 2026, with new customers beyond OpenAI.

So why the nerves? Because an order book is still a promise, and the cost of filling it is being paid now, much of it with debt. The next panel stress-tests that.

The stress test

Interest only goes up; profit can fall

Oracle's operating profit runs at about $27B a year, and about $5.7B goes to lenders as interest. The video applies three shocks: borrow $40B more at 7%, refinance cheap old debt at today's rates, then let profits slip to $20B from idle buildings and price cuts.

Step 1 is simple: $40B × 7% = $2.8B more a year. Step 2: in 2021 Oracle borrowed at just 1.65% on notes due 2026. Its February 2026 bonds paid 4.55–6.85%, and about $17B of debt comes due within two years. The red bars only grow because a loan's interest is fixed once you borrow; only profit can shrink.

The video adds ~$1B in step 2 (17 × 6.5%) and lands at $9.5B. The extra cost is really the difference in rates: $17B × (6.5% − 1.65%) ≈ $0.8B, so ~$9.3B. "Profit" here is annualised operating income from Oracle's Aug 2026 quarter ($6.7B × 4), not the $17.1B net profit.

What lenders are saying

Insuring Oracle's debt now costs 4×

A credit default swap (CDS) is insurance on a loan: the insurer pays out if the borrower defaults. Insuring $1M of Oracle debt for a year once cost about $5,000. It reached about $12,000 in the 2020 COVID panic. In 2026 it passed $20,000 (200 basis points), a record.

A basis point is 0.01% a year, so 200 bp on $1M is $20,000. Rising insurance prices don't mean Oracle will default. They mean lenders now see a real chance it could.

Field guide

The words in every AI-spending headline

So what

Three questions for any AI build-out story

• Who is paying, and can they afford it? A $60B-a-year promise from one customer whose revenue was $12B is concentration risk, however fast that customer grows.
• How long is the J-curve, and what does waiting cost? Delays and rent cuts stretch the time underwater; borrowed money charges interest the whole way.
• Will the asset outlast the debt? Buildings last decades, chips a few years, and bonds run 3 to 40 years. Ask which wears out first.

Whether the skeptics or the optimists are right, the video says only time will tell. Sources: Oracle FY2026 results ↗, Oracle Q1 FY2027 results ↗, Oracle Q1 FY2026 results ↗.