Modular AI Data Centers: Seven Trends Shaping Supply, Density and Capital

1Few units are being built

The modular story is often told in gigawatts. The factory reality is still counted in units per month.

Specialist makers are only now opening dedicated plants. The planned output figures they have published are modest, and most "hundreds" or "thousands of units" claims are stated targets rather than production. Large industrial groups do have scale, but they mainly sell power and cooling modules and prefabricated halls rather than ready-to-run AI containers — and their order books stretch well into 2027.

Why it matters: a mid-size order of 10–20MW is roughly 6–20 units of 1–1.8MW — potentially months of one factory's output. Delivery slots, not just price, become the negotiating point.

Labour is a second bottleneck. Factories are limited by skilled workers as well as floor space. In a 2026 global survey, nearly 90% of respondents cited shortages in mechanical, electrical and plumbing trades, and over 70% of markets reported contractor capacity as tight or overstretched. This cuts both ways for modular: moving work from the site to the factory is one answer to the shortage (one vendor estimates prefabrication cuts on-site labour by 20–40%), but the factories compete for the same electricians and technicians.

2Rack power decides which modules last

Each NVIDIA generation pushes more power into each rack. A module sized for today's chips may need replacing, not upgrading, for the next.

Most AI modules announced to date are rated at roughly 60–175kW per rack — enough for Blackwell, short of Vera Rubin. A few vendors now claim more: Armada's Leviathan uses 200kW racks, HPE states up to 400kW per rack for its AI Mod POD, LG CNS has announced a 4MW module built around Vera Rubin, and Hyperfusion and Aleria target Q1 2027 for a 10MW Vera Rubin site in Abu Dhabi. None has yet been delivered at Rubin-class density.

3Chip readiness comes from the kit inside, not the brand on the box

Whether a module can run NVIDIA's next chips depends on its power and cooling equipment — and that equipment mostly comes from a handful of industrial suppliers.

• Eaton — its Beam Rubin DSX platform is part of NVIDIA's Vera Rubin DSX AI Factory reference design (March 2026), alongside Schneider, Vertiv and Siemens.
• Schneider Electric — published a Vera Rubin NVL72 reference design (188/227kW per rack), an 800V DC prototype for Rubin Ultra and a 2.5MW coolant distribution unit.
• Delta Electronics — prefabricated AI module with 800V DC in-row power up to 660kW and liquid cooling up to 3MW, stating Vera Rubin support (June 2026).
• Vertiv — prefabricated platforms up to 10–12.5MW per block; a Rubin-specific design (OneCore Rubin DSX) sits in NVIDIA's DSX blueprint, though no per-rack figure is published yet.

By contrast, most container makers describe their units as "Rubin-ready" on their own authority. For a buyer, the practical test is simple: ask whose power and cooling equipment is inside, and whether it appears in NVIDIA's reference designs. NVIDIA's DSX POD (October 2026) widens that circle to 17 partners, including Armada, Crusoe, Bull, HPE and the large server makers; partnership signals roadmap alignment, not a commissioned unit.

4New sellers are arriving from next door

The supplier base is widening fast, and much of the new supply comes from adjacent industries.

Why it matters: more choice for buyers, but also more overlap — a supplier that runs its own capacity may compete for the same customers it sells to.

5Investors are funding, not buying

Capital is flowing into the category at rising valuations; consolidation is happening one layer down, in cooling.

• Specialists raise rather than sell. Armada raised $230M at a $2B pre-money valuation (May 2026); Crusoe raised $3.9B at a $30.9B valuation (Sep 2026), with modular units one part of a larger campus business.
• Industrials buy cooling. Ecolab acquired CoolIT for $4.75B (closed July 2026); Trane Technologies acquired LiquidStack (March 2026); Johnson Controls led Accelsius's $65M round (January 2026).
• Site owners buy modules and lease them back. Energy Vault is buying ten Crusoe Spark units for its Snyder, Texas campus and leasing them to Crusoe as powered shell (first phase 8MW, target operation Q1 2027).

Why it matters: modules are becoming a financeable asset class in their own right — owned by whoever controls the power and the land, not only by the operator.

6Regulation is catching up

Modular units are subject to the same reporting and permitting rules as any data center — and Europe is starting to shape the category directly.

• EU reporting is already in force. Any data center with 500kW or more of installed IT power must report energy efficiency (PUE), water use (WUE), energy reuse and renewable share each year (first reports September 2024, then every May). Vendor efficiency claims will become checkable against these filings.
• France is preparing a framework. In July 2026 the French government commissioned GIMELEC and France Datacenter to map industrial capacity for prefabricated and modular digital infrastructure in France and Europe, assess dependencies, and identify industrial, regulatory and skills levers. The final report is due in November 2026.
• Permitting is often the longest step. Environmental assessments, water rights and permits for backup generation can take longer than the grid connection itself — modular speed only helps once the site is cleared.

7France and Europe: demand is arriving faster than local supply

France has attracted some of Europe's largest AI data center commitments. The French factory capacity to build modules for them is only now being planned — and a government mission is mapping the gap.

Demand is large and near-term.

• SoftBank's France programme (announced May 2026): 5GW of AI capacity across three northern sites — Bosquel, Bouchain and Loon-Plage near Dunkirk — with up to €75B in total, of which a €45B first phase (3.1GW by 2031). EDF is a partner; Schneider Electric is the strategic power and prefabrication partner.
• Mistral AI's own 44MW cluster (~13,800 NVIDIA GB300s, announced March 2026) is hosted by Eclairion at Bruyères-le-Châtel (Essonne), a modular campus with 60MW operating and 120MW planned by end-2027, with pods supporting 30–200kW per rack. Mistral plans a Vera Rubin deployment in 2027 (reported at GTC Berlin).
• Public buyers are moving too: the Côte-d'Or department has run a container AI data center (104 NVIDIA GPUs) from French startup Hivenet's Policloud range since January 2026.

French and European supply, as publicly announced

What it means

• A timing gap. France's largest projects need prefabricated modules in 2026–28; the flagship French factory starts delivering in 2028. Until then, modules for French sites will largely come from Spain, elsewhere in Europe, or from US and Asian makers.
• A density gap. Among public announcements reviewed for this paper, French module makers serve edge and enterprise needs, with one exception at HPC scale: Bull, which built JUPITER's modular data centre in Germany. None has yet delivered a megawatt-scale AI module at Vera Rubin density.
• A sovereignty question. Public and regulated buyers increasingly want French or European supply. The GIMELEC / France Datacenter mission — mapping French and European prefabrication capacity, dependencies and industrial levers, with its report due in November 2026 — is the natural forum for closing these gaps.
• A skills gap. AI data centers are an industry in their own right, and much of the work is skilled manual trades rather than software: electricians, cooling technicians, commissioning engineers, electrical-equipment manufacturing (Upside NAI France, 2026 white paper). In the US the effect is already measurable: data center construction reached an annualised rate of more than $75B in July 2026, nearly 60% higher than a year earlier, and The Economist estimates AI has created around one million US jobs since mid-2023, against roughly 200,000 AI-related layoffs, spanning data center trades as well as AI professional roles. In France the link is not yet documented at that scale, but construction hiring is already strained, with 65% of planned hires judged difficult and electricians among the most sought-after trades. Skills are one of the levers the GIMELEC / France Datacenter mission was asked to examine.