Stargate is often described with one enormous number: 10 gigawatts of planned AI infrastructure. That number is easy to repeat and easy to misunderstand. It is not a promise that 10 gigawatts of chips are already running, and it is not the same thing as 10 gigawatts of electricity delivered tomorrow.

The useful question is simpler: what work has to happen before a power target becomes useful compute?

The short answer

OpenAI’s infrastructure plan treats power, land, permits, construction, networking, chips, and operations as one system. A 10-gigawatt target signals a very large build-out, but the target still has execution risk. Readers should separate announced capacity from energised facilities, installed hardware, and revenue-producing workloads.

OpenAI says its plan depends on partners and additional sites, with power and permitting among the practical constraints. That makes the announcement important, but not proof of immediate earnings.

What 10 gigawatts measures

A watt measures the rate of energy use. A gigawatt is one billion watts. In a data-center discussion, the number usually describes a facility’s electrical capacity rather than the number of GPUs. A site needs power for accelerators, CPUs, memory, storage, networking, cooling, lighting, and backup systems.

The difference between utility capacity and useful compute matters. A site can have a grid connection while construction, transformers, cooling equipment, or servers are still unfinished. It can also have servers installed while the software stack and customer workloads are not ready.

I use a four-step mental model when reading these announcements: contracted power, permitted construction, energised capacity, and productive utilization. Those four stages can be separated by 2 years or more, depending on location and equipment.

Why the build is difficult

The bottleneck is not only the GPU. Large campuses need substations, high-voltage equipment, fiber, water or alternative cooling, buildings strong enough for dense racks, and a workforce to operate them. A delay in one dependency can leave expensive hardware waiting.

There is also a financing question. A developer may fund a site with debt, equity, leases, customer commitments, or a mixture. Each structure changes who carries the risk if demand is slower than expected. A headline that combines a long-term ambition with a financing announcement should be read as a plan, not as recognized revenue.

How developers should interpret it

For software teams, more infrastructure can mean more model availability, lower queue times, and new agent products. It does not automatically mean every API becomes cheaper. Providers still price for scarce hardware, network egress, support, and the cost of keeping capacity available during peaks.

For investors, ask four questions:

  1. How much capacity is operational today?
  2. Who owns the buildings and hardware?
  3. Are customer commitments take-or-pay, or can customers walk away?
  4. What utilization and power cost assumptions make the project profitable?

Those questions reveal more than the gigawatt number alone.

Sources and limits

OpenAI’s compute infrastructure explanation describes the partner ecosystem and the need to secure land, power, permits, and staffing. Its scaling update provides additional context on planned compute and partner commitments. NVIDIA’s financing announcement shows why financing is becoming part of the infrastructure story. Announced capacity is not a guarantee of completion or profit.

The practical takeaway is to track each site through the four stages rather than treating 10 gigawatts as one finished asset. That habit prevents both hype and needless cynicism.

A useful reporting habit

When a company publishes a capacity target, record three dates: announcement, expected delivery, and expected utilization. Then revisit the claim when each date arrives. This makes it possible to distinguish a delayed project from a cancelled one and a finished building from a productive cluster.

The same discipline works for readers who are not investors. If you are choosing an AI API, ask how much of the provider’s capacity is live, which workloads it serves, and what fallback exists. A large infrastructure plan can improve future availability while still leaving today’s developer with queues, regional limits, or changing prices.

Spend 2 hours checking the latest permitting, construction, and utilization evidence before treating the target as operating capacity.