State of the Modular Data Center Market 2026

Updated September 27, 2026. This report incorporates 2026 market and energy data, the European Commission's September 21 data-center package, current AI rack architectures, and a fuller procurement framework.

The modular data center market is growing, but its headline revenue forecasts do not tell a buyer which facility will deliver usable capacity. In September 2026, two research firms put the year's global modular market at $35.44 billion and $41.0 billion. Their estimates cover different mixes of complete facilities, IT modules, power and cooling packages, and services. Neither number is a count of operating megawatts. For a project team, the urgent questions are whether power can be energized, whether the selected hardware has a workable electrical and thermal design, and whether factory and site work can reach acceptance on the same schedule.

This report examines the market evidence available through September 2026, the technologies changing specifications, the limits of modular construction, and the information a buyer should require before placing an order. It covers complete facilities and major prefabricated subsystems; it does not assume that every AI campus should be built from containers.

TL;DR: September 2026 findings

• Demand is substantial, but market size is a model. Fortune Business Insights estimates $35.44 billion for 2026; MarketsandMarkets estimates $41.0 billion. Their categories and forecast horizons differ, and neither figure measures installed capacity.
• Electricity, sites and equipment set the pace. The IEA estimates data centers used 485 TWh in 2025 and projects 950 TWh in 2030 in its central outlook. Its grid analysis finds that connection constraints could delay about 20% of planned global data-center capacity to 2030. A modular building cannot remove those dependencies.
• Available space remains scarce in major markets. CBRE reports 6.7% average vacancy across 16 major markets in Q1 2026. Its North American primary-market measure was 1.4% in H1. These are market samples, not global occupancy rates.
• AI changes the design unit. A rack-scale system brings its own power, cooling, weight, network and service requirements. The equipment bill of materials and acceptance test need to shape the facility before a module layout is frozen.
• Modular helps when interfaces can be repeated. Factory integration and site preparation can overlap. Grid connection, authority approvals, transport, field integration and commissioning still sit on the project schedule. The relevant comparison is cost and time to accepted operating capacity.
• Regulatory status has changed this month. EU data-center reporting already exists for qualifying facilities. The Commission published a rating-scheme package on September 21, 2026, while consulting on possible minimum performance standards. A proposed standard is not a current universal EU pass/fail rule.

What the modular market includes

“Modular data center” describes a construction and integration approach, not one fixed product. A factory may deliver an entire small facility, a repeatable IT room, a power module, a cooling plant, or separate units that become one site system. The classification matters because a market estimate can count an electrical skid and a complete IT facility in the same revenue pool even though they answer different buyer requirements.

A container is one possible enclosure, not the definition of the sector. Large AI projects increasingly combine prefabricated electrical, thermal and IT packages with conventional civil works. Conversely, a small enterprise facility may arrive as a largely integrated product. Buyers should compare the exact delivered scope before comparing advertised dollars per megawatt or months to deploy.

How large is the market in 2026?

The 2026 estimates below are useful as directional measures of supplier opportunity. They are not independent audits of sales, and they should not be averaged into a fictitious “consensus.”

The two publishers use different product segmentation, research methods and endpoints. The $5.56 billion gap in their 2026 estimates is a warning against quoting a single precise market size without a definition. Revenue also moves with component prices and project scope; it does not translate directly into data-center square metres, IT megawatts or ModulEdge's addressable market. For investment or sales planning, define geography, facility size, buyer type, power status and the work included in the contract before sizing an opportunity.

Demand is high; deliverable capacity is the constraint

The IEA's April 2026 update estimates that global data-center electricity use increased 17% in 2025 to 485 TWh. In its central projection it reaches about 950 TWh in 2030, with AI-focused facilities growing faster than the overall data-center stock. This is an electricity scenario, not a forecast of modular product sales. The IEA also emphasizes uncertainty from chips, energy equipment, grid access, financing and the economics of AI workloads.

Leasing data shows why developers want an earlier path to capacity. CBRE measured 16 GW of supply across 16 major data-center markets in Q1 2026, with average vacancy falling from 8.3% to 6.7% year over year. In eight primary North American markets, CBRE recorded 1.4% vacancy in H1 and said 80.4% of under-construction capacity was preleased. In Europe, its Q2 2026 figures forecast 13 GW of total market capacity by year-end, 20% above 2025, while new AI-related colocation signings more than quadrupled in the first half. The European figure is a forecast; the leasing series is not a census of all European demand.

None of these indicators guarantees that a proposed site is viable. The IEA estimates that grid constraints could put roughly one-fifth of planned global data-center capacity to 2030 at risk of connection delay. A developer may own land and have a fiber route yet still lack a firm energization date, required redundancy, sufficient transformer capacity or local consent. On-site generation can change the power path, but it introduces fuel, permitting, equipment, emissions and reliability questions of its own.

Modular construction is valuable when a repeatable facility scope can proceed while those site dependencies are resolved. It cannot turn a speculative connection application into firm power. A credible programme shows separate dates for power, design freeze, long-lead orders, factory acceptance, transport, site works, commissioning and workload acceptance.

Who buys modular infrastructure, and for what job?

Cloud and GPU operators seek capacity that matches actual hardware and customer commitments. A large rack-scale AI cluster needs a different mechanical, electrical and network plan from a regional inference site. Colocation operators may buy complete rooms or just power and cooling blocks to add capacity to an operating campus. Energy and site owners may use a modular design to turn a power-secured location into an investable facility, but tenant demand and operating responsibility must be established separately. Enterprises, public bodies, industrial operators and telecom providers may need smaller local facilities for control, latency, resilience or physical access.

These buyers do not share one risk model. A developer financing a facility for a neocloud must test the tenant's credit, contracted capacity, hardware plan and renewal exposure. An enterprise buying a compact module must test whether a dedicated site is preferable to an existing-room upgrade, colocation or cloud service. A public-sector buyer may add procurement, security and jurisdictional requirements. Calling all of them “AI demand” hides the information that determines the design.

Rack-scale AI changes power, cooling and network specifications

The facility brief now often starts with a named hardware architecture. NVIDIA's GB300 NVL72 reference describes a liquid-cooled rack-scale unit with 72 Blackwell Ultra GPUs and 36 Grace CPUs. Its network and management design extend beyond the compute rack. AMD's Helios rack-scale design uses its own open-rack form factor, GPU, network, power and cooling interfaces. A facility that fits one platform cannot be declared ready for the other by changing a product label.

At design freeze, the buyer needs the selected OEM's rack dimensions and mass, service clearances, power-feed topology, transient load behaviour, network fabric and cabling routes, supply and return water conditions, pressure and flow, residual air load, leak detection, and maintenance method. The facility must also reserve space and power for networking, storage, control and support equipment. Those racks may be numerous enough to change the building layout and the IT load calculation.

The industry's installed base has not moved uniformly to extreme density. Uptime Institute's 2026 survey reports a gradual rise in typical rack densities even as more operators report peak racks of 30 kW or above. A typical rack in a mixed estate and the peak rack in a new AI cluster are different measurements. Design capacity should come from the actual equipment schedule, not a universal “AI rack” average.

Power distribution: current requirements and the 800 VDC roadmap

Higher rack demand increases current, protection and conductor challenges across the path from utility supply to the rack. The right answer depends on the selected system and the facility's conversion, redundancy and maintenance strategy. In August 2026, NVIDIA described an 800 VDC roadmap with a hybrid power rack expected in the second half of 2026 and a row power center expected in 2027. These are vendor plans and architectures, not evidence that 800 VDC is installed in every AI facility or specified for every ModulEdge product.

A buyer planning for future rack generations should ask whether space, electrical rooms, routes, protection philosophy and cooling can accommodate a later power architecture. The current phase should still be engineered to the equipment that will actually be purchased. A promise of “future-proofing” is useful only when the spare capacity, change path and disruption are shown on drawings and in the commercial scope.

Cooling: follow the heat to its final sink

Direct liquid cooling removes heat from selected components, not automatically from every device in the room. Fans, power equipment, switches and other components can still reject heat to air. A complete scheme traces the path from the chips and residual air to the CDU or heat exchanger, facility loop and outdoor heat rejection. It specifies operating and failure conditions: supply temperature, flow, water chemistry, filtration, redundancy, leak response, ambient design points and available service access.

“Closed loop” describes coolant circulation, not necessarily zero site-water consumption. Water use depends on the outdoor rejection method, climate, operating hours, water treatment and any adiabatic assistance. Likewise, heat recovery is useful only if an off-taker can accept the required temperature and seasonal profile. For project sizing, annual water or heat-reuse ratios cannot replace a peak-day connection and storage calculation. Our separate CDU sizing guide and water-budget guide show the inputs a design team should request.

Energy performance: compare measured boundaries

Uptime Institute's 2026 survey analysis reports an average annual PUE of 1.52 across surveyed facilities and a 1.36 result when facilities are weighted by capacity. That difference is a reminder that a sample average is not a design target for a particular site. PUE is total facility energy divided by IT energy over the same boundary and period. It does not by itself describe water, renewable electricity, compute efficiency or the quality of the heat-rejection design. Ask whether a quoted PUE is modelled or measured, annual or instantaneous, and what equipment lies inside the meter boundary.

Power and delivery: where factory work changes the schedule

Prefabrication can move assembly, integration and some testing away from a constrained construction site. Standard interfaces can make repeated phases easier to procure and inspect. Those advantages are strongest when rack, power, cooling, controls and network requirements are mature; transport and lifting are feasible; and factory work can overlap with civil, utility and fiber work. They weaken when every phase changes the architecture, the design remains open during fabrication, local approval requires major field rework, or common infrastructure dominates the programme.

Price needs the same boundary discipline. Normalize supplier bids for civil work, medium-voltage connection, transformers, generators, UPS, distribution, heat rejection, CDUs, fire and security systems, fiber, transport, cranes, commissioning, spares, maintenance and warranty. Treat power reservations, tenant commitments and financing as commercial conditions, not as equipment specifications. A lower module price can be the higher completed-facility price if it shifts expensive work to the owner.

Why the electrical boundary changes the business case

Consider a purely illustrative 1 MW IT deployment running at an average 75% of design IT load. If its measured annual PUE were 1.30 for the same period and meter boundary, it would use 0.75 MW × 8,760 hours = 6,570 MWh of IT energy and 6,570 × 1.30 = 8,541 MWh of total facility energy. The 1.30 value is an assumption for the example, not a prediction or a ModulEdge performance claim. At partial load, real annual PUE must be measured or modelled for that operating profile; it cannot simply be read from a peak-load brochure.

That arithmetic illustrates a contract issue. A 1 MW IT rating is not a 1 MW utility reservation, and an annual MWh estimate is not the peak import rating. Cooling plant, conversion losses, redundancy and failure states affect installed equipment and the connection application. The buyer should separately price the infrastructure, the electricity and network service, and any GPU or tenant commitments. If a vendor offers “cost per MW,” ask which MW, at what operating state, and which site works the price excludes.

Conformity and EU policy: what changed in September 2026

Modular design does not confer a certification by itself. UL 2755:2025 provides an evaluation route for prefabricated modular data-center systems and related units, including distributed power, cooling, IT and support modules that work together. Whether a particular supplier's system is certified, and what further site approvals are required, must be checked against actual documents and the destination authority.

In the EU, Delegated Regulation (EU) 2024/1364 establishes information and performance indicators for reporting by data centers with installed IT power demand of at least 500 kW. The reporting scope and responsible party must be checked for the specific installation; a 500 kW product label alone is not a legal assessment of every site.

On September 21, 2026, the European Commission published a common rating-scheme package intended to make energy and water performance more transparent and comparable. It separately opened consultation on potential minimum performance standards, with a legislative proposal planned for 2027. At this report's September 27 update, buyers should plan metering and documentation for existing reporting and follow the rating scheme's final status. They should not treat a proposed future minimum as an already applicable EU-wide PUE limit.

For any jurisdiction, ask who holds a claimed certificate, which product or site it covers, the edition and validity date, and what factory and installed-site evidence the authority will require. A component's certificate, a design intention and an accepted operational facility are three different forms of evidence.

Where the market is moving next

Several trends are clear enough to shape current specifications, even where their ultimate scale is uncertain. First, the next data-center phase is constrained by firm power, electrical equipment, construction labour, local approval and capital as much as by IT demand. Second, rack-scale systems bind compute, network and cooling into one procurement decision. Third, developers are considering smaller phases to align spending with tenant or workload commitments, but they must size shared infrastructure and plan later expansion without unacceptable outages. Fourth, operating obligations—monitoring, security, maintenance, water and energy reporting—are moving into the initial facility brief.

The 2027 watchlist should be written as questions rather than a parade of launch dates: Which announced sites obtain firm power? Which AI rack designs reach volume delivery with stable facility interfaces? How much of preleased capacity reaches commissioned operation? Which EU rating requirements take effect, and how will national rules interact with them? How do liquid-cooling operations and service costs compare at the actual site? Each answer can change a project design more than a new global market-size forecast.

A procurement test for a real project

Issue one common project brief to every bidder. It should state the workload and owner, intended hardware, initial and future IT load, site and available power, climate and water conditions, availability objective, deployment phases, operator, and required acceptance milestone. Ask each supplier to respond with:

1. A capacity schedule: nominal, peak and continuous IT load; facility input; thermal load; liquid and air split; outdoor heat-rejection duty; and failure-state ratings.
2. An interface schedule: rack geometry, weights and service clearances; voltage, power feeds and protection; network paths; coolant temperatures, flow, pressure and quality; and monitoring points.
3. A scope matrix: supplied, integrated, installed and excluded work for every package from utility connection to rack and from chip heat to outdoor rejection.
4. A programme with dependencies: design freeze, long-lead purchases, factory tests, authority approvals, site readiness, transport, SAT, integrated testing and operating handover.
5. A normalized commercial offer: equipment, transport, civil work, commissioning, spares, maintenance, warranties, change process and payment milestones.
6. Proof: drawings, calculations, component and system credentials, inspection/test plans, previous relevant acceptance records where shareable, and named responsibilities for open items.

A supplier can still be considered if a detail awaits the final OEM bill of materials. The quotation should identify that assumption, who resolves it, and how a change affects price, capacity and schedule. Buyers need to tell a mature design from a provisional promise.

Where ModulEdge fits

ModulEdge develops modular facility designs for different IT loads and site conditions. Its current portfolio spans standard smaller systems, a 500 kW IT design, and a high-density modular AI building. The offer can include engineering and integration of project power, cooling and control systems; the included and excluded interfaces belong in the project scope.

The MDC-500 specifies 500 kW of IT design capacity across ten 42U rack positions. Its documented architecture uses air-cooled IT with active rear-door heat exchangers, a chilled-water loop and external modular chillers. The MDC-2000 is a prefabricated modular building configurable for NVIDIA GB300 NVL72 and AMD Helios-class infrastructure. The current public GB300 layout has 16 compute racks and 16 network/service racks. A previously issued 16-plus-8 reference model calculated 2.328 MW peak IT load; that value is not a revised load schedule for the current 16-plus-16 layout or for an AMD configuration. Rack population, electrical rating, cooling interfaces and outdoor plant are finalized against the selected equipment and site.

ModulEdge manufactures switchboards and related electrical and control assemblies. Specialist OEMs manufacture items such as UPS systems, generators and transformers; ModulEdge can select, engineer, assemble, integrate and test those items within a defined project power architecture. These statements describe product scope and current offer. Project credentials, delivery schedules and measured performance require evidence for the specific configuration and site.

For an initial fit discussion, share the site location, documented usable power and target date, intended equipment and rack count, availability objective, cooling and water conditions, and whether you need infrastructure supply, equipment integration or a tenant-ready facility. ModulEdge can then identify a relevant platform, the design work needed and the boundaries that remain outside it.

Frequently asked questions

How big is the modular data-center market in 2026?

Published estimates cited here range from $35.44 billion to $41.0 billion for 2026. They are publisher models with different category definitions, not measured installed capacity. Compare the included solutions, services, geography and forecast date before reusing either figure.

Does modular mean a shipping container?

No. A modular facility can be a compact enclosure, a prefabricated building, an IT or data-hall block, a power or cooling module, or a hybrid campus. The delivered boundary is more important than the external shape.

Is a modular data center always faster or cheaper?

No. Factory and site work can overlap when interfaces are stable, but grid connection, permits, long-lead equipment, transport and commissioning can control the finish date. Compare priced, accepted operating capacity on the same scope and timeline.

Can one facility design support every AI rack platform?

No. Rack-scale systems differ in geometry, electrical topology, cooling temperatures and flow, network fabric, service access and support racks. Platform configurability requires configuration-specific engineering and acceptance.

Does liquid cooling eliminate air cooling or water use?

No. Some equipment still rejects heat to air, and site-water use depends on the outdoor heat-rejection method and climate. A complete design identifies both heat paths and peak as well as annual water requirements.

Does UL 2755 or factory testing certify the finished site automatically?

No. A relevant product evaluation and witnessed factory tests can support approval, but the actual certification scope, installation, local authority and site acceptance still have to be checked.

What changed in EU data-center policy in September 2026?

The Commission published a rating-scheme package on September 21 and opened consultation on potential minimum performance standards. Existing reporting obligations under the 2024 framework remain separate. Buyers should verify the package's final legal status and applicable national rules before specifying compliance.

What should a buyer ask for first?

A documented site and workload brief: usable power, target hardware, rack schedule, cooling and water conditions, operating model, deployment phases and acceptance date. It lets suppliers price the same project boundary and show which assumptions are still open.