Oracle is running into multibillion-dollar cost surprises across its data center buildout, according to an exclusive from The Information. Not a rounding error. Not a quarterly blip. Costs on the infrastructure Oracle is racing to stand up are landing well above what the company planned for.
This is significant because Oracle has bet the company on exactly this buildout.
Tactical Points
- The bet is enormous. Oracle spent the last two years repositioning itself from a legacy database vendor into an AI cloud provider. That meant signing massive compute contracts, including its role in the OpenAI-linked Stargate infrastructure push, and then promising to deliver the physical capacity behind them.
- Contracts were signed before the costs were known. That’s the core problem The Information’s reporting points at. When you commit to delivering gigawatts of compute at a fixed price years out, every unplanned dollar comes straight out of your margin. There’s no renegotiation clause for “turns out transformers cost more now.”
- Oracle is the most leveraged player in the trade. Microsoft, Google, and Amazon fund their data centers out of enormous existing cash flows. Oracle doesn’t have that cushion. It’s been funding this expansion with debt, which means cost overruns hit the balance sheet harder and faster than they would at a hyperscaler with a decade of buffer.
- The status quo before this was pure optimism. For most of the past 18 months, the market treated Oracle’s backlog of signed cloud contracts as close to guaranteed revenue. Big number, big stock move. What this reporting does is put a question mark on the other side of the ledger: what does it actually cost Oracle to service that backlog?
Why the Costs Are Blowing Up
AI data centers aren’t regular data centers, and the industry keeps relearning this the expensive way. A few drivers:
- Power: Grid connections, substations, and transformers have multi-year lead times. Skipping the queue means paying for on-site generation, which is expensive and slow.
- Cooling: Dense GPU racks need liquid cooling. That’s a different building, not a retrofit of the old one.
- Labor: Electricians and specialized construction crews are the actual bottleneck in several US markets, and their rates reflect it.
- Speed premium: Building fast costs more than building well. Oracle chose fast.
What This Means for You
If you’re buying AI compute, this is a signal worth tracking. Aggressive pricing from a provider that hasn’t finished building the thing it’s selling you is a promise, not a product. Ask about delivery timelines and what happens contractually if capacity slips.
If you’re building on AI infrastructure generally, expect the cost of inference and training capacity to stay sticky rather than falling on the smooth curve everyone penciled in. The compute may get cheaper per chip. The buildings, power, and people around those chips are going the other way.
If you’re watching the AI trade as an investor or operator, this is the first real crack in the “signed contracts equal profit” story. Backlog is revenue. Backlog is not margin.
Assessment
The AI infrastructure boom has been priced on the assumption that demand is the hard part and supply is a solved engineering problem. Oracle’s experience suggests the opposite. Demand is real and enormous. Delivering against it, on budget, at this speed, is where companies get hurt.
Watch Oracle’s next earnings call closely. The questions analysts ask about capex and gross margin on the OCI business will tell you whether this is a contained overrun or a structural problem with the whole model of selling compute you haven’t built yet.
More detail on the specific numbers is available in The Information’s original reporting.