In partnership with
Two announcements this month used one word, AI factory, for two different things, and the money followed each. On 7 October, NTT DATA opened an AI factory in Munich. It is a live client test environment inside an existing building, the Munich 2 data centre, on NVIDIA Blackwell processors. On 1 October, JERA, Dell and RHAELM signed a memorandum of understanding for a 400 MW AI data centre. It sits on land next to JERA's Chiba power station, near Tokyo. Its total capital, land to compute, is expected to exceed £11.3bn ($15bn, €13.2bn). Dell's "AI Factory" is its compute layer.
"Data centre" is the umbrella asset class. Common usage treats AI factories/hyperscale data centres as almost one thing. "Hyperscale" says whose capacity it is and whom it serves. "AI factory" names a class of data centre: one built to produce AI. An AI factory and a conventional data centre share a shell and a labour bill. At the same capacity they differ by more than ten times in rack density, and that gap carries a price. A figure that says "data centre" without naming the class is a figure a buyer cannot place.
The most common rack density has only just passed 11 kW
NVIDIA defines an AI factory as "a new class of infrastructure built to manufacture intelligence". In its words, "Data centers once stored files. Now, AI factories produce tokens."
In the Uptime Institute's 2026 global survey, the most common rack density passed 11 kW for the first time. Without the high-density sites the average is 7.8 kW. Uptime's chief technical officer, Chris Brown, uses the class name himself. "You have the artificial intelligence training world, AI factories, which are driving rack densities up faster than anywhere else." Data Center Knowledge, reporting the survey, describes two markets at once. One is AI training clusters. The other is enterprise workloads on central processing units (CPUs), which still hold most installed capacity.
Uptime and the IEA give three axes and only one holds the new class
Resilience comes first. The Uptime Institute's Tier Standard rates and certifies site infrastructure from Tier I, basic capacity, to Tier IV, fault tolerant. That axis owns the word "tier", and this analysis uses it for nothing else.
Deployment model comes second. The International Energy Agency (IEA) counts three types: enterprise, colocation and server provider, and hyperscale. Synergy Research puts hyperscale operators at 48% of worldwide capacity, non-hyperscale colocation at 20% and enterprise at 32%. That share counts whose capacity it is, not what its halls hold.
Workload density comes third: air-cooled conventional halls, through high-density accelerated halls, to the liquid-cooled AI factory. The new class lives on this axis alone. A Tier IV facility may contain no AI hall at all.
Rack density differs by more than ten times and the power profile follows
Against that 11 kW, Schneider Electric's reference design for NVIDIA GB300 NVL72 clusters runs to 142 kW per rack. Trade press reports NVIDIA showing a 600 kW rack for its Rubin Ultra generation, due late 2027.
At these loads heat leaves through liquid, direct to the chip. Researchers at Microsoft, OpenAI and NVIDIA describe training clusters of tens of thousands of GPUs (graphics processing units). Their power swings in step as the job moves between compute and communication. If those swings harmonise with the critical frequencies of utilities, the paper warns, they can damage grid infrastructure. A conventional hall draws a comparatively steady load.
A conventional facility sells availability, hence the tier. An AI factory sells throughput, measured in tokens. Siting therefore turns from latency to power. RHAELM says land next to JERA's power stations "can shorten the path to power" and bring capacity online sooner.
An AI hall builds 7 to 10% dearer and waits on the same transformer
Our August analysis of build costs drew on Turner and Townsend's index. It finds electrical systems are 54% of an air-cooled build and mechanical systems another 22%. A high-density AI hall costs only 7 to 10% more to build than a conventional one. Both wait on the same transformers and grid connections.
That index puts Tokyo at £11.4m ($15.2m, €13.4m) a megawatt and Dublin near £7.5m ($10.0m, €8.8m). Cushman & Wakefield's 2026 guide puts a US and Canada new build at £13.2m ($17.6m, €15.5m) a megawatt, chips excluded. The UAE's Stargate campus is costed above £22.6bn ($30bn, €26.4bn) for 5 GW. None of these figures names the class inside the hall.
Turner and Townsend put the build premium at 7 to 10%. The rack density gap is more than ten times. The cooling plant and the grid contract follow the density, not the shell.
When your organisation prices a data centre build, do its figures name the facility class?
The EU tender and the Singapore allocation now name the class in their own terms
The European Union coined its own class name. The EuroHPC tender for up to seven AI gigafactories, which we reported last week, closes on 12 November. Winners are expected to begin operations within 18 months of selection. Singapore's 200 MW allocation on 21 August came with conditions. Every winning proposal runs liquid cooling and fully energy-efficient IT equipment. An allocation that writes "liquid cooling" into its conditions has named the class without using the word.
Cassava Technologies' first AI factory in Africa, on NVIDIA's platform, had its first Cape Town deployment in place by May. A 20 MW Johannesburg site is planned, with Nigeria, Kenya, Egypt and Morocco to follow.
Three disciplines follow and the first is naming the class on every figure
First, name the class with every figure. A build cost, a water draw, a power profile or a cost of delay specific to AI halls says so. A general figure is never presented as AI-specific. In Scotland, CX Tech has proposed a 1 GW campus of eight data centres at the former Chapelcross nuclear site. It is costed at £16.5bn ($21.9bn, €19.3bn). Public reporting names no class, so its cooling plant, power contract and grid timeline cannot be read from the figure.
Second, classify on three axes, never one. State the tier, the deployment model and the density class separately. "AI factory" answers the third question only.
Third, price the inside of the hall apart from the shell. A schedule or a model can carry a class column and flag every figure that arrives without one. The estimator, the information manager and the grid engineer decide what the tag changes.
The class decides three line items for investors
Density decides the cooling capital, since liquid plant is specified per rack. Power profile decides the grid contract, since a load that swings needs grid flexibility or on-site smoothing. Siting decides the connection timeline, where the interconnection queue sets the pace, as we showed in September.
The immediate priority is a class column in every portfolio schedule. Tokyo against Dublin means something only when the class matches. Our Atlas pairs the terms, AI factories / data centres, because the map holds every class.
In closing…
A data centre stores. An AI factory manufactures. They share a shell, a labour bill and a wait for the same transformer. At the same capacity they differ by more than ten times in rack density. The cooling plant, the power contract and the siting follow. When a number says data centre, ask which class. A figure that does not say hides which price it carries.
Next week: Africa's data centres stopped waiting for the grid. Everyone else is next.
Your SOC 2, handled end to end with AI agents.
Enterprise buyers won’t put your product near their customer data without a SOC 2 report. Twenty-one state privacy laws now hold them accountable for the vendors they share data with, so their diligence lands on you.
Sprinto gets you audit ready in 14 days, across three working sessions. AI agents connect your stack, collect the evidence auditors ask for, and close the gaps as they appear. You approve, they execute.
Sprinto also answers the security questionnaires your buyers send, using the same evidence base, so security reviews stop holding up deals.
No compliance hire. No consultant. Your auditor signs off.