AI companies are building data centers at a pace that can make every announcement sound like proof of demand. But a planned campus is not an energized campus, an energized campus is not a fully used campus, and a signed contract is not the same as profitable usage.

This does not mean the demand is fake. AI workloads genuinely need power, networking, storage, and specialized chips. It means readers should track the full path from capital spending to paying usage.

Four stages of capacity

The first stage is planned capacity: land, permits, financing, and a construction schedule. The second is installed capacity: buildings, power delivery, cooling, GPUs, and networking are in place. The third is contracted capacity: a customer has reserved or agreed to buy access. The fourth is productive usage: workloads run, customers are billed, and cash is collected.

Each stage has a different risk. A project can be delayed by grid connections. Installed hardware can sit below target utilization. A contract can be renegotiated. A busy cluster can still lose money if power, leases, depreciation, and financing consume the margin.

Why the numbers get large

Data-center projects combine many years of spending into one headline. A campus may be designed for multiple gigawatts, while only part of it is available today. The announcement may also include partner capital rather than the operator’s own cash.

NVIDIA has said that land, power, shell construction, and capital are all critical constraints. Oracle has described large AI infrastructure projects with OpenAI. These announcements show how much coordination is required; they do not independently prove that every future megawatt will be used at attractive margins.

When I review a capacity claim, I write down the delivery window and ask whether the number means design capacity, contracted capacity, or live capacity. A 15-year lease can support planning, but it is still a long obligation that needs a customer and a working site.

I would spend 2 hours checking the project’s delivery milestones, not only the designed capacity. Ask when power is available, when hardware is installed, and when a customer can actually run work. The gap between those dates is where construction and financing risk often hides.

The utilization test

Utilization is not a single universal percentage. Training workloads may run in bursts, inference may be steadier, and reserved capacity may be paid for even when it is idle. Ask how the company defines utilization and whether it reports revenue per available GPU hour, gross margin, or customer mix.

The best evidence is a trend: installed capacity rises, customer usage rises, service revenue follows, and operating cash improves. If capacity rises much faster than usage, the company may still be investing ahead of demand. That can be sensible for a fast market, but it increases financing and execution risk.

How to avoid the two bad extremes

Do not assume every project is a bubble because spending is large. Do not assume every project is healthy because a famous customer is named. Check permits, delivery milestones, contracts, power availability, equipment ownership, and cash flow.

Also separate the economy from the stock. Real AI demand can exist while a company’s valuation expects too much growth. Conversely, a company can have a valuable long-term position while facing short-term debt pressure.

A compact research worksheet

For each announcement, record: who funds it, who owns the equipment, when power becomes available, who signed the capacity agreement, how cancellation works, and how the operator gets paid. Then compare those answers with the latest filing and update your notes after 2 reporting periods.

That process turns a dramatic headline into a testable sequence. The central question is simple: is new capacity becoming reliable service and collected cash, or is the system only producing more commitments?

Keep the worksheet after the article is published. Revisit it when the operator reports results and replace estimates with delivered capacity, billed usage, and collected cash. A dated comparison is much more useful than a permanently optimistic forecast.

Sources and limits

Read NVIDIA’s 2026 filing, NVIDIA’s infrastructure announcement, Oracle’s AI infrastructure update, and OpenAI’s Oracle cloud announcement. Project schedules and financing terms can change; verify current filings before relying on them.