In this explainer
100 MW of AI capacity: five questions behind the number
IT load, total site power and actual consumption measure different things. Learn how to read the numbers behind an AI campus announcement.
A headline calls a new AI campus a “100 MW data center.” That sounds precise. Yet the number could describe electricity reserved for the entire site, capacity intended for computers, or an expansion that will arrive years from now.
Before comparing two projects, ask three questions: Where is the power measured? What stage has the project reached? And does the number describe a limit or actual use? The answers can change the meaning of the headline without changing a single digit.
Start at the meter, not the headline
Megawatts measure power: the rate at which energy is used or supplied. One megawatt equals 1,000 kilowatts. Megawatt-hours measure energy accumulated over time. A steady 100 MW load running for one hour consumes 100 MWh; running for ten hours consumes 1,000 MWh. The U.S. Energy Information Administration explains these units.
A capacity announcement alone cannot tell you annual consumption. For an intentionally simple calculation, assume a facility draws exactly 100 MW all year. In a 365-day year, 100 × 8,760 hours gives 876,000 MWh, or 876 GWh. If its average draw is 60 MW instead, the result is 525.6 GWh. These are arithmetic scenarios, not forecasts for a named campus.
The missing input is the actual load over time. A stated maximum does not supply that information. Nor does an annual energy total reveal the highest demand during a particular hour.
The computers are only one part of the electrical load
IT load covers computing equipment, including servers, storage and networking. Facility electricity also supports cooling and the losses involved in delivering power, among other building services. A number measured at the utility connection therefore describes a different boundary from a number measured at the IT equipment.
The Department of Energy defines annual power usage effectiveness, or PUE, as total facility energy divided by IT-equipment energy. A PUE of 1.20 means that, over the measured period, the facility used 1.20 units of energy for every unit used by IT. The extra 0.20 represents infrastructure overhead; it is not necessarily all cooling. DOE’s guidance explains the measurement.
Now compare two hypothetical announcements. Both say “100 MW,” and both use an assumed annual PUE of 1.20. Their implications differ:
| What 100 MW describes | Implied average at PUE 1.20 |
|---|---|
| Average IT power | 120 MW total facility power |
| Average total facility power | About 83.3 MW IT power |
These calculations work because both averages cover the same period and use the same boundaries. They do not establish how many megawatts of computers can safely be installed behind a 100 MW utility connection. For that, designers need peak conditions, equipment ratings, reserve margins and the intended failure and maintenance arrangements.
PUE cannot turn a site limit into a design guarantee
An efficient annual average does not guarantee that the cooling system or electrical distribution can support a particular peak. A design target is also different from a measured result after the facility has operated through changing conditions.
Uptime Institute draws a useful distinction between operational overhead and the allocation of a site’s electrical capacity. Its analysis notes that PUE does not by itself describe how provisioned capacity is divided between IT, cooling, other services and the facility’s redundancy design. Read Uptime Institute’s capacity-allocation analysis.
That means a low PUE and a large power reservation answer different questions. Neither number alone tells a prospective customer whether a specific rack can be connected, cooled and supported during maintenance. For the physical details inside the building, see our guide to a 120-kilowatt AI rack.
Signed, built, energized and occupied are separate milestones
A real example shows why the labels matter. In its Oct 8, 2026 announcement, Bitdeer described A901 in Malaysia as 67 MW of critical IT load under a ten-year data-center services agreement. It targeted initial delivery in the first quarter of 2027 and full delivery in the second quarter.
The company also said no customer offtake commitments had been signed for A901 at the announcement date. Its definition of secured capacity included owned facilities and capacity covered by executed services agreements. In this case, “secured” did not establish that all the equipment was operating or that customers had purchased the output. Read Bitdeer’s original announcement or our briefing on the agreement.
For any project, keep the milestones separate: an agreement, completed construction, an energized electrical connection, commissioned systems and occupied computing capacity. A project may achieve them in phases. Reporting a final campus target as present-day consumption skips precisely the information readers need.
Megawatts are not an AI performance score
Power is an input. Useful computation is an output. Converting one into the other requires knowing the hardware, workload and how effectively the system uses its equipment. Even two sites with identical electrical demand can be doing different work.
Our explainer on AI-cluster networking describes one reason: processors that need shared results can spend time waiting on communication. Counting electrical capacity cannot reveal that behavior. A defensible performance comparison needs a specified task and a measured result, not only a campus size.
Five questions to take to the next announcement
- Which boundary? Ask whether the figure is IT power, total facility power or a utility-service limit.
- Which stage? Separate a proposed ultimate buildout from the phase operating today.
- Which measurement? Identify capacity, peak demand or average use, and the period covered.
- Which assumptions? Ask whether PUE is a design target or measured result, and avoid using an annual average as a peak guarantee.
- Which outcome? Request evidence for the specific claim: delivered computing, electricity consumed or capacity ready for customers.
These questions also help communities examine a proposal. A developer’s eventual power target is relevant to planning, while actual annual consumption is a different question. Neither alone establishes the local water demand, electricity-bill effect or number of jobs. Those claims need their own evidence.
A useful description might read: “100 MW of planned IT capacity, with 25 MW commissioned in the first phase.” It is longer than a headline, but it tells the reader what the number measures and how much of the promise has become physical infrastructure.
.