What is an AI factory?

The phrase AI factory — popularised by NVIDIA and now standard industry vocabulary — is more than marketing. It describes a datacentre engineered as a production line: the raw materials are electricity and data, the machinery is racked accelerated computing, and the output is tokens — the units of text, code, images and actions that AI models consume and produce. The framing matters because it changes what the building is optimised for. A traditional cloud datacentre is designed for heterogeneous workloads, air cooling at 5–15 kW per rack, and maximum tenant flexibility. An AI factory is designed for one product at maximum yield: liquid-cooled halls at 130–600 kW per rack, a single fused network fabric, power architecture matched to the GPU roadmap, and — increasingly — a single anchor tenant per hall.

The factory metaphor also imports factory management. Output is measurable (tokens per second, model runs completed); yield is manageable; the production line is re-tooled on a known cadence; and unit economics are computable down to the token. That last step — connecting a megawatt of electricity to a price per million tokens — is what this chapter builds.

A token is roughly three-quarters of an English word, a fragment of code, or a patch of an image. Every prompt sent to a model and every answer returned is metered in tokens, and the commercial AI industry has settled on price per million tokens as its invoice line. Token volumes are the cleanest public evidence that AI demand is real and compounding. Alphabet disclosed that its systems processed roughly 480 trillion tokens per month across its surfaces in May 2025 — about 50× a year earlier — rising to more than 1.3 quadrillion per month by October 2025. Its latest disclosure uses a different measure: at Q2 2026, Gemini models processed 22 billion API tokens per minute, up from 16 billion a quarter earlier (Alphabet, 22 July 2026). That is roughly 0.95 quadrillion API tokens per month (author arithmetic) and is not comparable with the all-surfaces series, for which no 2026 figure has been published. Inference — running models in production rather than training them — has become the permanent base-load of GPU demand.

• Revenue per MW, restated in tokens. At ~700 H100-class GPUs per MW of IT load (author estimate), 1,000–3,000 output tokens per second per GPU equates to roughly 22–66 trillion tokens per MW-year (author arithmetic) — the "nameplate output" of the factory. A 2026 contract at ~$20–25m per MW per year therefore prices that output at very roughly $0.3–1.1 per million tokens of capacity, before utilisation.
• Training and inference monetise differently. Training is sold as reserved capacity; inference is increasingly sold as tokens or as dedicated always-on halls. The same physical asset can serve both, but contract shape, network design and siting optimum differ.
• Token deflation is structural; GPU-hour pricing is cyclical. Each hardware generation cuts cost per token faster than it cuts $/GPU-hour, and multi-vendor silicon and hyperscaler surplus resale add downward pressure. 2026 showed that GPU-hour prices can rise for a year or more when demand outruns deliveries; underwriting should still assume token prices fall every year the asset operates.

Training is episodic, centralised and latency-tolerant at the campus level: it goes wherever the largest coherent block of cheap power and cooling exists. Inference is permanent, distributed, and latency- and jurisdiction-sensitive: it wants to sit near users and inside legal borders. As the workload mix shifts from training toward inference (CBRE and operator disclosures through 2026 point the same way), the premium moves toward dedicated, in-country, always-on capacity with a low delivered cost per token — precisely the configuration that sovereign programmes and cheap-power regions such as the Nordics are positioned to supply. The AI factory, in other words, is not just a bigger datacentre; it is the point where energy economics, silicon cadence and national policy meet a measurable product.