Oracle's $23.7B Cash Burn Is the Real AI Story

Think of Oracle right now as a publican who's just been told the Guinness lorry will arrive every hour for the next decade. Demand is real, the orders are signed, but the cellar needs rebuilding, the taps need replacing, and the bill for all of it lands before the first pint is poured. That's the shape of Oracle's fourth quarter: a $638 billion order book and a $23.7 billion hole where the free cash flow used to be.

The Numbers

The headline beat is real. Oracle posted Q4 revenue of $19.2 billion, up 21% year over year, against a Wall Street expectation of $19.09 billion. Non-GAAP earnings came in at $2.11 a share versus consensus of $1.96. GAAP earnings were $1.45. By any normal reading of an earnings print, that's a clean beat on both lines.

The cloud breakdown is where the AI story actually lives. Cloud revenue hit $9.9 billion, up 47%. Infrastructure as a service, the bit that matters for GPU workloads, grew 93% to $5.8 billion. SaaS came in at $4.1 billion, up a more pedestrian 10%. Legacy software revenue was $6.8 billion, down 2%, which Oracle attributes to customers moving to the cloud, a polite way of saying the on-prem licence business is being cannibalised by its own faster-growing sibling.

Then there's the number that reframes the whole quarter. As Constellation Research reported, Remaining Performance Obligations at quarter end were $638 billion, up 363% year over year. That's not a typo. RPO is essentially contracted revenue Oracle hasn't recognised yet, and it just quadrupled.

Oracle was unusually direct about where it came from. Most of the Q3 and Q4 RPO increase was large-scale AI contracts in which customers either prepaid Oracle for GPU purchases, or in some cases bought the GPUs themselves and handed them to Oracle to operate. The prepaid and customer-supplied hardware portions of those contracts now total $75 billion.

That last detail is the guts of it. Customers are functionally co-financing Oracle's AI build-out. Some are writing cheques upfront. Others are showing up to the data centre with their own H-series silicon and asking Oracle to plug it in and run it. The Multicloud AI Database product grew 404% in the quarter, which tells you the data layer is being dragged along by the compute layer.

And then, the bill. Free cash flow for fiscal 2026 was negative $23.7 billion. Oracle attributes this to investments supporting the cloud infrastructure business. In plain English: the GPUs, the buildings, the power, the cooling, and the network gear are being paid for now, while the revenue gets recognised over years.

What's Actually New

The last cloud cycle, the AWS-Azure-GCP one, was financed largely off internal cash flow. Hyperscalers built capacity, signed customers, and recycled the operating use. The capex was huge but it was self-funded. Margins compressed for a few quarters and then expanded again as utilisation caught up.

What Oracle is showing this quarter is a different financing model entirely. When customers prepay for GPUs, or supply their own, the economic relationship inverts. Oracle is no longer the bank. The customer is the bank, and Oracle is the operator. That $75 billion of prepaid and customer-supplied hardware is a structural admission that GPU supply is so constrained, and AI training budgets so enormous, that the people who want compute are willing to fund the build.

Anyone who has tried to get a meaningful H100 or B-series allocation in the last 18 months knows why this is happening. The bottleneck isn't dollars, it's silicon, power, and a data centre with the right ratio of megawatts to square feet. If you're a frontier lab or a hyperscaler-adjacent AI company, prepaying Oracle to jump the queue is rational.

The second new thing is the scale of the forward book versus the run-rate. Oracle's trailing annual revenue is in the mid-fifties of billions. RPO is now $638 billion. That's roughly a decade of revenue contracted forward, dominated by a small number of very large AI deals. Backlog used to mean three quarters of visibility. Now it means a decade. Whether you find that comforting or terrifying depends on your view of AI demand durability in 2028 and beyond.

The third genuinely new thing is the Health business. Oracle expects double-digit growth in fiscal 2027 driven by a new Cerner hospital and clinic patient care management system. That's a long way from the GPU story, but it matters because it's a reminder that Oracle still has an enterprise software franchise underneath the AI infrastructure trade. Anyone who has integrated against Cerner's old APIs in a hospital setting will tell you a credible rewrite is overdue.

What's Priced In for AI Development

For engineering teams planning AI workloads, most of this print confirms what was already obvious. GPU capacity will be contractually locked up years in advance. If you're a fintech CTO or an ad-tech platform lead thinking you can show up in Q2 2027 and rent serious training compute on demand, the $638 billion RPO number is your answer. The good seats are sold. What's left is spot capacity at spot prices, with all the unpleasantness that implies.

What's priced in: that hyperscalers and large AI labs will continue to absorb most of the GPU supply, that multi-year reserved commitments will dominate, and that anyone wanting frontier compute pays upfront. None of that is news to anyone who has tried to host or fine-tune a large open-weights model in the last year.

What's not priced in, I'd argue, is the financial fragility this creates for the operator. Oracle's negative $23.7 billion free cash flow isn't a one-off. It's the cost of being the venue for the AI buildout. If even one of the very large customers behind those prepayments hits trouble, or quietly renegotiates, the operating use works in reverse. The market is treating the RPO number as quality earnings deferred. It's really a bet that the counterparty list stays solvent through the next downturn.

For platform engineers, the practical read is that vendor concentration risk is back. If your AI roadmap depends on a single cloud's GPU allocation, you're now exposed to that cloud's customer concentration too. The boring bit nobody wants to talk about: capacity planning for 2027 needs a backup vendor in the architecture diagram.

Contrarian View

The consensus take on this print is that Oracle has finally cracked it, that Larry Ellison's late-career pivot to AI infrastructure has worked, and that the negative free cash flow is a feature, not a bug, because every dollar burned today buys a dollar of contracted revenue tomorrow at attractive economics.

Here's the contrarian read. When customers are supplying their own GPUs to a cloud provider, that's not a sign of provider strength. That's a sign that the customer doesn't trust the provider to procure fast enough, and is willing to take on the supply chain risk themselves to get the workload running. The cloud is being reduced to a colocation facility with extra steps. Margins on operating someone else's hardware are not the margins of a software business.

The other quiet worry is software revenue down 2%. Oracle says it's cloud cannibalisation, and that's probably true. But the IaaS growth is overwhelmingly AI training and inference, not the steady enterprise app modernisation that produces sticky 20-year customer relationships. If the AI capex cycle cools, Oracle is left with a smaller legacy software business and a vast, depreciating GPU estate. That's a real risk the $638 billion RPO headline obscures.

Key Takeaways

• RPO of $638 billion is real but concentrated: Most of the 363% jump came from a small number of very large AI contracts. Counterparty risk is now Oracle's biggest single exposure.
• Customers are co-financing the buildout: $75 billion of contract value is prepaid or customer-supplied hardware. That's a tell about GPU scarcity, not Oracle's balance sheet strength.
• Negative $23.7 billion free cash flow is the price of admission: Engineering leaders evaluating Oracle as an AI infrastructure partner should model the buildout cycle, not the current quarter's print.
• IaaS at 93% growth confirms the GPU bottleneck: Reserved multi-year capacity commitments will dominate frontier compute access through 2027 and likely 2028.
• Don't ignore the Health pivot: Cerner's new patient care system driving double-digit Health growth in fiscal 2027 is the non-AI ballast in the story, and it's underdiscussed.

Back to the publican. The cellar is being rebuilt, the lorries are queued up outside, and a handful of very large customers have prepaid their bar tab through 2035. That's a brilliant position to be in, right up until you remember the rebuild was paid for on credit and the customers who prepaid all work in the same industry. The pints will pour. Whether the cash flow ever comes back positive is a different question, and one this quarter doesn't answer.