The $300 Billion Bet That Made Oracle the AI Boom's Weak Link

On September 10, 2025, Oracle announced a contract to sell OpenAI roughly $300 billion of computing power over about five years — the largest cloud deal in history by a wide margin. The stock jumped 42% that day, Oracle's market value briefly traded past $1 trillion, and for a moment its chairman, Larry Ellison, was reported to be the richest man in the world.

A year later Oracle trades near $145, down roughly 60% from its 52-week high of about $346 and off 25% for the year. The greatest supply commitment in the history of cloud computing produced the worst shareholder experience in Oracle's modern history. The deal did not fail. The question is whether the market was right to decide that the deal, and the debt that came with it, has turned Oracle into the AI boom's weak link.

The deal front-loads the risk

The economics of the arrangement are the whole story, so they are worth spelling out. OpenAI committed to buy about $60 billion per year of compute from Oracle across five years, beginning when the infrastructure comes online in 2027. To deliver that, Oracle had to build the capacity first — buying GPUs, land, and power years before a dollar of that contract shows up as revenue.

That is a standard cloud build-out pattern, but the scale is not standard, and neither is the financing. In fiscal 2026 (the year ended May 31), Oracle spent $55.7 billion on capital expenditures, up 162% from the prior year, and burned $23.7 billion of free cash flow — a number the company had essentially never produced before as a negative. Management now guides to roughly $70 billion of capex in fiscal 2027 and is raising $40 billion more through a mix of debt and equity, including a $20 billion share sale.

The load-bearing fact, though, is concentration. Oracle's backlog of contracted future revenue, its remaining performance obligations, surged 363% to $638 billion over the past twelve months. But as analysts pointed out at the Q4 report, more than half of that backlog is a single customer: OpenAI.

Why this is different from the other hyperscalers

Every hyperscaler is spending enormous sums on AI data centers. Amazon, Microsoft, Google, and Meta have each made tens-of-billions-of-dollars commitments. The reason the "weak link" label attaches to Oracle specifically is that it alone has placed the majority of its contracted future revenue on one payer — a payer whose own revenue still trails the size of its compute bill.

As of early 2026, OpenAI's annualized revenue was around $25 billion against a compute commitment of roughly $60 billion per year — an enormous gap, even after OpenAI closed a $122 billion funding round in March 2026 at an $852 billion valuation. In other words, Oracle has built its near-term balance sheet on a single customer's ability to keep raising money and keep growing fast enough to justify bills it currently cannot pay out of its own revenue.

The credit market has noticed. In July 2026 S&P downgraded Oracle's rating to BBB-, one notch above junk, after heavy AI spending drove debt higher and cash flow negative. Oracle has committed around $250 billion to AI-focused data centers.

This is what separates Oracle from its peers: it is not front-loading capex into a diversified backlog or a broad customer base. It is betting the single largest part of its future on one customer's ability to pay, at a scale that exceeds that customer's own revenue.

The execution has cracked, not collapsed

It would be wrong to write Oracle off as an operating failure. The fundamentals beneath the financing still grew. In its fiscal Q4 reported in June, revenue rose 21% to $19.18 billion, cloud infrastructure revenue jumped 93% to $5.8 billion, and non-GAAP EPS of $2.11 beat estimates. The stock fell 11% anyway, because the market was repricing the financing, not the product.

There are also real cracks in the delivery story. Stargate — the broader OpenAI-Oracle-SoftBank build-out — has been reported to be floundering amid disagreements among the partners over control. In March 2026, OpenAI and Oracle abandoned a planned 600-megawatt expansion at the flagship Abilene, Texas campus, a reminder that giant interconnection queues can shrink as fast as they inflate.

Where that leaves the judgment

The original worry — that Oracle had become the fragile link in the AI chain — has already been priced in hard. At roughly $145, Oracle trades at about 25 times trailing earnings and 6 times sales, down from a peak valuation that carried it past a trillion dollars. The market has cut it in half partly because the near-term return curve is genuinely ugly: negative free cash flow, a credit downgrade, heavy dilution coming.

But the risk is not resolved, either. It now hinges entirely on the 2027 ramp: whether OpenAI starts paying Oracle's contracted amounts on schedule, whether utilization of all that built capacity is high, and whether the deal's terms survive as AI demand and model economics shift. If OpenAI stumbles, Oracle's stranded capex risk is real and concentrated. If OpenAI's compute consumption keeps growing as quickly as it has, Oracle's 2027 inflects from a cash-burner into a cash collector.

That is the contract both ways, and it is why I would not call Oracle either a clear buy at the bottom or a short at these levels. The thesis that made Oracle an AI hero in September 2025 is intact. The question a holder or a watcher has to answer is narrower: whether the risk that remains — one customer, one ramp, one downgrade cycle — is the risk you are being paid to take, versus what the rest of the AI trade offers. The supply commitment was always a dual signal. It still is.