Oracle's $455bn order book

Oracle's remaining performance obligations (RPO), the measure of contracted but unbilled customer backlog that software analysts trust most, rose by 36 per cent year-on-year to a record $455 billion. The figure shows how much companies around the world want enterprise cloud infrastructure and specialised AI computing capacity.

The market's response was extraordinary for a company of Oracle's size. The shares jumped by more than a third in the session after the results, their biggest one-day gain in decades, and the rise briefly made Larry Ellison, the co-founder and largest shareholder, the richest person in the world. The company also set out a forecast that its cloud infrastructure revenue would climb to $144 billion within five years, from a small fraction of that today.

Who signed the contracts

The backlog is large but not broad. Management said the jump came from a handful of multibillion-dollar contracts with a few customers, and the largest of them is widely understood to be OpenAI. The Wall Street Journal has reported that OpenAI agreed to buy roughly $300 billion of computing capacity from Oracle over about five years, starting from 2027. Neither company has confirmed the figure, but it would account for a large share of the total.

That link runs through Stargate, the joint venture announced at the White House in January by OpenAI, SoftBank, Oracle and MGX to build AI data centres in America. Oracle agreed in July to develop additional capacity for OpenAI under that umbrella, with the flagship site at Abilene, Texas. A backlog concentrated on one customer is only as sound as that customer's ability to pay, and OpenAI, for all its growth, still spends far more than it earns and depends on fresh rounds of outside funding.

A backlog that needs building

Shareholders cheered the backlog. Credit analysts looked at what it will cost to deliver. A $455 billion backlog is not cash in the bank. Oracle has committed to building dozens of gigawatt-scale data centres and filling them with tens of billions of dollars of advanced GPUs, and it still has to secure grid connections that are in short supply.

The timing mismatch is the heart of the problem. Oracle pays for land, buildings, power equipment and chips up front, while its customers pay for capacity as they use it, over contracts running several years. Revenue is recognised only as the service is delivered. In the meantime the company carries the cost of the construction, and if any customer delays or renegotiates, the assets are already built.

Some of the load is being shifted to partners. Specialist developers such as Crusoe are building sites like Abilene and leasing them to Oracle, which then installs the computing equipment. That keeps construction debt off Oracle's own balance sheet but replaces it with long-term lease commitments that credit analysts treat much like borrowing. The obligation is the same whichever column it sits in.

Borrowing to build

To turn contracts into billable cloud revenue, Oracle must spend heavily, and much of the money will be borrowed. With corporate borrowing costs high, it will need to keep issuing debt, which will strain its credit metrics and reduce how much of its earnings turn into free cash.

Oracle starts from a more stretched position than its rivals. Microsoft, Alphabet, Amazon and Meta fund their data centres largely from operating cash flow and carry some of the strongest credit ratings in corporate America. Oracle already has a large debt load from years of share buybacks and its purchase of Cerner, the health-records company, in 2022, and its bonds sit in the lower half of the investment-grade scale. Capital spending has already absorbed most of its operating cash flow, so every additional data centre from here requires external finance.

The ratings arithmetic is unforgiving. Agencies judge a company like Oracle largely on debt relative to earnings before interest, tax, depreciation and amortisation. Borrowing rises immediately when construction starts; the earnings from a new data centre arrive only once it is finished, filled and billing. For two or three years leverage goes up before it can come down, and an issuer already near the bottom of investment grade has little room to absorb that.

Lessons from the fibre boom

The nearest precedent is the telecoms build-out of the late 1990s. Carriers such as Global Crossing and WorldCom borrowed heavily to lay fibre against forecasts of explosive internet traffic. The traffic did eventually arrive, but too late for the companies that financed the capacity, and a wave of defaults followed in 2001 and 2002. The lesson was that demand can be real and the financing still fail, if the timing is wrong and the balance sheet is thin.

The differences matter too. Oracle's customers have signed binding contracts rather than issuing forecasts, and the largest cloud buyers are among the most profitable companies in the world. The chips and power equipment have resale value, and demand for AI computing currently exceeds supply. A slowdown would hurt Oracle's margins long before it threatened its solvency.

Who gains and who carries the risk

The winners from the order book are easy to name: Nvidia and other chip suppliers, makers of electrical equipment and cooling systems, construction firms and the utilities that sell power to the sites. Shareholders have captured the value of the backlog immediately through the share price. Bondholders, by contrast, get a fixed coupon and bear the downside if a customer walks away, which is why credit investors tend to demand wider spreads just as equity investors are celebrating.

What would change the call

The signals are straightforward. Oracle's next quarterly results, due in December, will give the first reading on how fast capital spending is rising against the new backlog. Watch the size and frequency of Oracle's bond sales over the coming months, any change in rating agency outlooks, and whether the backlog converts into billed revenue at the pace management has promised. OpenAI's own fundraising matters as much as Oracle's, since the customer's ability to pay underpins the contract. Oracle's $455 billion order book is strong evidence of demand for AI, but delivering it will turn a software company into a debt-funded infrastructure business, and the rating agencies will probably grow uneasy well before the data centres are finished.