What is an AI Factory?

What is an AI factory?

An AI factory is a data center purpose-built for AI compute: designed from the GPU cluster backwards rather than from the real estate forwards, with rack densities, cooling, and networking that conventional data centers can't support. The term (used by NVIDIA, orchestration vendors like Rafay, and the neocloud operators themselves) signals a break with the traditional data center category, and operators increasingly refuse the older word entirely.

The metaphor is more literal than it looks. As one European neocloud founder describes it: the computers transform energy into tokens, and the tokens are what customers consume. Input power, output tokens. That's a factory, and it invites factory economics: throughput, utilization, and unit cost per unit of output.

How does an AI factory differ from a normal data center?

The refresh row is the quiet structural fact: an AI factory must be designed for the power draw of GPUs that don't exist yet, because each generation roughly doubles consumption per device. Facilities that can't absorb that curve are obsolete mid-depreciation.

Where do AI factories get built?

Two answers are emerging. Giant greenfield campuses (hundreds of megawatts) where grid connections allow, on timelines of many years. And retrofits: former industrial sites that already hold 10-20 MW of grid connection, converted fast. European operators lean on the retrofit path because grid queues in hubs like Frankfurt can run 6-8 years, while an old factory with existing power can be an AI factory in a fraction of the time. There's also more "stranded power" than the headlines suggest: sites with available megawatts that no existing facility can deliver at AI rack densities.

The likely end state resembles the power grid itself: a mix of giga-scale plants and distributed 20-50 MW regional facilities rather than a few centralized giants, especially for inference, which wants to sit near its users.

What do AI factory economics mean for billing?

Everything about the facility raises the stakes on the meter. Capital intensity means every unmonetized GPU-hour is expensive idleness, which is why the commercial layer stacks reserved commitments, on-demand, and spot to keep utilization sold (see GPUaaS billing). The 3-5 year refresh cycle means SKU-level pricing changes continuously as new generations land, so the price book has to be versioned, not hard-coded. And the energy-to-tokens framing becomes literal in the P&L: power is the dominant variable cost, revenue is metered output, and margin per tenant is only knowable if metering attributes both sides.

Operators selling up the stack (hosted models, per-token billing) complete the factory picture: the facility's output unit stops being the hour and becomes the token. That transition is the token factory.

FAQ

Is "AI factory" just marketing for a data center?

The rack-density numbers say no. A facility built for 15 kW racks cannot run 100 kW racks by trying harder: the power delivery, cooling, and floor design are different buildings. The new name marks a real engineering break.

How big is a typical AI factory?

There's no typical yet. The range runs from 10-25 MW regional retrofits (roughly 10,000 GPUs) to campuses in the hundreds of megawatts. The distributed end is growing fastest because it ships years sooner.

Do AI factories only do training?

No. Training favors the giant centralized builds. Inference, the operational consumption of AI, runs well in smaller distributed facilities and is the volume business over time.

Who operates AI factories?

Neoclouds primarily, plus hyperscalers building their own, and sovereign operators building national capacity. See sovereign AI billing for that tier's commercial specifics.

Related

• Neocloud: the operators

• Token factory: the output-metered version

• GPU-hour: the unit the factory sells today

• Neocloud billing: the commercial hub