AI Factory MW vs Data Center IT Load

An AI infrastructure announcement can contain several megawatt figures without any of them meaning the same thing. “AI factory capacity” may describe a computing platform, an initial GPU deployment, the critical IT load inside a facility, the utility connection available to the site, or a long-term portfolio ambition. Treating those figures as interchangeable creates false comparisons and weak investment decisions.

The practical answer is simple: every reported MW figure should be attached to a defined boundary, a date and a delivery status before it enters a market model. PowerlandMap separates platform targets, deployed AI capacity, IT critical load, facility electrical capacity and contracted power because each answers a different question.

Why AI factory capacity needs a measurement boundary

The problem is visible in recent public announcements. On 9 September 2026, ZanKore said it had secured a senior term loan of up to $3.1 billion to fund an initial 100 MW deployment of NVIDIA AI infrastructure in Indonesia. The company also referred to a target of 1 GW of NVIDIA DSX AI Factory capacity and approximately 200 MW of initial capacity in the first half of 2027. Those are useful facts, but they are not automatically equivalent to 100 MW or 200 MW of commissioned data center IT load.

The earlier Ooredoo announcement of 6 August 2026 described a 1 GW platform target and approximately 200 MW of AI capacity expected in the first half of 2027. The ZanKore financing announcement of 9 September 2026 then tied financing to an initial 100 MW deployment. Reuters independently reported the financing on the same date. None of those sources publicly identifies the exact sites, utility import capacity, IT critical load or commissioning profile behind every figure.

This is not a criticism of the announcements. Platform sponsors communicate strategy, financing and commercial scale; developers, utilities and investors need a different level of precision. The analyst’s job is to preserve what was reported while avoiding a conversion that the source does not support.

Five different MW figures can coexist

Platform target

A platform target is the broadest number. It describes the scale a company intends to develop across time, geographies or several facilities. It may include phases that do not yet have land, permits, grid access or customers. It is valuable as a statement of ambition and procurement direction, but it should not be counted as current supply.

PowerlandMap records such figures as reported targets and connects them to the evidence and observation date. This is the same discipline explained in our analysis of what “up to 1 GW” means for data center supply: a ceiling is not a delivery schedule.

Initial AI infrastructure deployment

An initial deployment figure can refer to the power envelope assigned to GPU systems, an aggregate across several facilities, or a commercial tranche of the wider platform. The most important question is the denominator: is it compute equipment power, critical IT load, or facility input?

If the announcement does not say, the figure should remain labelled “AI infrastructure capacity” rather than being silently converted into IT MW. That preserves comparability and keeps the PowerlandMap methodology auditable.

Critical IT load

Critical IT load is the power delivered to servers, storage and network equipment. For data center supply analysis, it is often the most useful operational denominator. Yet even this figure can be ambiguous when a project mixes high-density liquid-cooled halls, conventional cloud space and support areas.

An AI facility may also change density or rack configuration without changing the total critical load. GPU generation, rack density and cooling architecture therefore belong beside IT MW, not inside it. Users comparing projects through PowerlandMap coverage should be able to see which capacity is reported and which remains unverified.

Facility electrical capacity

Facility electrical capacity includes IT load plus cooling, electrical losses and ancillary systems. Depending on design and reporting practice, it may be expressed as utility import, transformer capacity or a planning figure. It cannot be treated as IT MW without a documented conversion basis.

Power usage effectiveness can help explain the relationship, but applying an assumed PUE to a press-release number is not enough. The operating condition, redundancy philosophy, climate and phase all matter. A credible market model leaves the cells separate until compatible evidence exists.

Contracted and deliverable power

The final boundary is the power that can actually be delivered at the required date. A grid agreement, a reserved connection position and energised capacity are different milestones. On-site generation or storage may support resilience, but it does not automatically replace firm utility capacity for the whole load.

This is where megawatt comparisons become development analysis. The PowerlandMap product links reported capacity to projects, sources and readiness signals so that users can distinguish an announced programme from a site capable of commissioning equipment.

A practical decision framework

Before comparing two AI infrastructure announcements, ask five questions. What physical or commercial boundary does the MW figure describe? Is it one site, several sites or an entire portfolio? What date does the figure apply to? Which delivery milestone supports it? Is the number reported directly, estimated by an analyst, or inferred from another metric?

The answers produce a more useful classification:

• Reported target: publicly stated ambition, not yet treated as supply.
• Committed phase: financing, procurement or contractual evidence exists, but delivery readiness may still be incomplete.
• Development capacity: named site with credible land, permitting and power milestones.
• Commissioned IT load: operational critical capacity with a clear measurement boundary.

This hierarchy is intentionally conservative. It prevents a portfolio target from appearing beside a commissioned facility as if both were immediately available. It also helps investors connect capacity claims to the financing evidence discussed in our guide to reading data center debt signals.

What the ZanKore example tells decision-makers

The public facts are material. ZanKore has announced up to $3.1 billion of senior debt for NVIDIA infrastructure, an initial 100 MW deployment in Indonesia, a broader first-half 2027 capacity statement of approximately 200 MW, and a 1 GW platform target. Ooredoo, Indosat Ooredoo Hutchison, Nokia and NVIDIA bring capital, telecom, networking and compute capabilities.

The inference is narrower: the programme should generate meaningful requirements for data center space, power, cooling, connectivity and phased delivery across Southeast Asia. But the public sources observed on 13 September 2026 do not yet permit PowerlandMap to assign exact sites, electrical import capacity, IT MW or ready-for-service dates. Those fields remain open rather than estimated.

That distinction matters to several audiences. Investors can avoid capitalising an ambition as delivered capacity. Operators can assess whether a deployment is additive supply or a customer requirement already allocated to existing facilities. Utilities can focus on named connection requests rather than portfolio headlines. Developers can identify where site and power readiness evidence is still missing. End users can compare timing, not merely scale.

The PowerlandMap view

Matthieu Gallego’s view: the market should stop asking whether an AI project is “100 MW” and start asking “100 MW of what, where, and by when?” A precise answer is more valuable than a larger headline. In my experience, most development risk sits in the interfaces between compute procurement, facility design, grid delivery and commissioning—not in the headline number itself.

For PowerlandMap, the right response is not to discard ambitious figures. It is to preserve each one with its original wording, source date and capacity class, then update the record as sites and milestones become public. That creates a decision-grade timeline without pretending that uncertainty has disappeared.

What to watch next

The next useful evidence for ZanKore will be named facilities, utility or energy counterparties, site-level capacity definitions, permitting milestones and a phased commissioning timetable. Customer commitments would clarify demand, while procurement and construction awards would reveal delivery progress. Any of those developments could materially change how the 100 MW, 200 MW and 1 GW figures should be interpreted.

PowerlandMap will continue to separate reported capacity from verified supply across its global market coverage. Investors, developers and operators can request access to evaluate project evidence, or book a demonstration to see how the methodology works on live infrastructure programmes.

*Observed by PowerlandMap on 13 September 2026. Source publication dates: Ooredoo, 6 August 2026; ZanKore and Reuters, 9 September 2026. All capacity descriptions above retain the terminology of the cited public sources unless explicitly identified as analysis or opinion.*

— Matthieu Gallego