SemiAnalysis Wild Datacenter Buildout Exposes AI's Labor Bottleneck

SemiAnalysis published a wild assessment of modular datacenters on July 29, tracking more than 61 gigawatts across over 1,000 sites. Its central conflict is simple. AI companies want computing capacity faster than conventional construction crews can deliver it.

The proposed solution resembles an industrial LEGO set. Factories assemble walls, electrical rooms, cooling equipment, and prepared data halls while contractors complete the site. Trucks then deliver those units for installation and final commissioning.

Yet modular construction does not remove every constraint. It moves repeatable work into factories, shortens some critical paths, and changes who controls integration. Power availability, permits, equipment supply, transportation, and testing remain stubbornly local.

That distinction separates a credible construction strategy from the industry's expanding collection of modular marketing claims. AWS, Meta, colocation operators, equipment manufacturers, and specialist integrators are all using the same label for materially different systems.

SemiAnalysis Wild Data Shows What Changed

Modular construction has moved from a specialized option to a core strategy for delivering large AI campuses.

The SemiAnalysis investigation describes more than 80 vendors operating across the modular supply chain. Its tracker covers over 61 gigawatts of planned or operating capacity at more than 1,000 sites.

SemiAnalysis estimates that modular facilities will represent more than 30 percent of live capacity by the end of 2028. That projection is proprietary analysis, not an industry census. Still, the underlying construction shift is visible across several facility layers.

Prefabrication means producing any construction element away from its final site. A modular unit is narrower: it is a self-contained block that arrives substantially assembled and ready for connection.

This difference matters because almost every modern project includes some prefabricated material. A factory-made wall panel does not turn an entire campus into a modular datacenter.

The current movement reaches much deeper. Operators can order electrical equipment on open skids, complete power rooms, cooling assemblies, rack-ready halls, or nearly complete facility blocks.

Subsystem modularization covers a limited function, such as power distribution or liquid cooling. Whole-facility modularization combines several functions within a coordinated structure.

The site itself remains largely conventional. Contractors must grade land, pour foundations, connect utilities, manage drainage, and satisfy local requirements. Those activities cannot arrive on a truck.

The shell offers the next opportunity. Precast concrete panels, standardized steel frames, and lightweight enclosures can reduce field fabrication. They also let factories work while crews prepare the site.

The systems inside provide the largest modular opportunity. Electrical gear, piping, pumps, cooling distribution units, busways, controls, and rack infrastructure contain extensive repeatable work.

SemiAnalysis estimates that a modular approach can shorten a construction window by about 36 percent, or seven to nine months. Its model also estimates an 8 percent capital reduction per megawatt.

Those results should be treated as analytical estimates, not universal guarantees. Project size, geography, design maturity, transportation, and utility schedules can substantially change the outcome.

Established vendors make similar speed claims. Vertiv says its MegaMod system can reduce deployment time by up to 40 percent against traditional construction.

Schneider Electric describes a broader range of outcomes. Its modular assessment says standardized modules offer faster, more predictable deployment, sometimes at costs comparable with conventional facilities.

These claims align on direction but not magnitude. That gap is the first sign of the market's loose terminology and inconsistent baselines.

The real change is therefore larger than a new building format. Datacenter construction is becoming a manufacturing and integration problem, with more value created before equipment reaches the site.

AI's Construction Boom Is Running Into Labor

The hardest capacity constraint is increasingly the trained workforce needed to turn electrical power into operating compute.

Datacenter demand has collided with a labor system that cannot expand on semiconductor timelines. Electricians and pipefitters need structured training, supervised experience, and often licensing before they can perform critical work.

The Bureau of Labor Statistics projects approximately 81,000 electrician openings annually from 2024 through 2034. It expects electrician employment to grow 9 percent during that period.

Those openings span the entire economy. Datacenters must compete with housing, manufacturing, utilities, renewable energy, transportation, and ordinary commercial construction for the same workers.

SemiAnalysis estimates that electricians account for 30 to 40 percent of construction labor hours on a datacenter project. Its labor model indicates that shortages become particularly acute in Texas and Ohio during 2027.

That forecast depends on proprietary assumptions about project schedules and reachable workers. The broader pressure is easier to establish. Multiple large campuses are being developed simultaneously in markets with finite trade capacity.

SemiAnalysis says Crusoe increased wages by 30 percent while staffing its Abilene construction site. The project reportedly required more than 9,000 workers at peak activity.

Higher wages can attract workers from nearby projects or regions. They cannot instantly create experienced electricians, commissioning specialists, or pipefitters.

The shortage also compounds itself. Pulling workers toward one large campus can delay factories, substations, transmission upgrades, housing, and other supporting projects in the same region.

Traditional construction makes that competition worse because much of the work happens sequentially. A delayed shell can hold up electrical installation, which can then postpone cooling, cabling, testing, and server delivery.

Modularization changes the sequence. Factory teams assemble repeatable systems while site contractors prepare foundations and utility connections. Parallel work replaces portions of the conventional chain.

A factory can also create a more stable labor environment. Workers remain at one controlled location instead of moving between distant jobsites with changing weather and temporary facilities.

Repetition supports specialization. One team can repeatedly wire the same power block, while another performs standardized testing before shipment. That arrangement reduces the number of field interfaces.

It does not eliminate skilled work. It concentrates selected tasks in manufacturing centers and leaves local teams responsible for placement, connection, inspection, and commissioning.

The shift therefore changes who faces pressure. Engineering, procurement, and construction contractors must develop manufacturing capabilities or coordinate with specialist integrators.

Equipment vendors must decide how much of the finished system they want to own. Selling a switchgear component differs substantially from guaranteeing an integrated power module.

Operators face their own choice. They can retain design control, outsource integration, or buy a coordinated system from one manufacturer.

Large hyperscalers possess enough engineering and purchasing capacity to pursue operator-led modularization. Smaller developers often need an integrator or original equipment manufacturer to manage more of the risk.

This is where the labor story becomes a competitive story. The scarce capability is no longer just installing equipment. It is coordinating design, procurement, assembly, logistics, and testing without recreating complexity elsewhere.

Factory Integration Beats Faster On-Site Assembly

Modular construction works when it removes dependencies, not when it merely changes where workers tighten the bolts.

The strongest case appears in electrical systems. A power module combines equipment such as switchgear, transformers, batteries, busways, and controls within a tested factory-built package.

SemiAnalysis estimates that power systems represent about 26 percent of a datacenter build by content. Moving this scope into a factory reportedly produces much of modular construction's schedule advantage.

Its model places a modular power hall at roughly 13 months to IT readiness. The comparable conventional design takes about 16.7 months.

The estimated saving comes partly from compressing mechanical and electrical fit-out from about 5.5 months to 2.5 months. SemiAnalysis also estimates a 5 percent cost reduction per megawatt.

Factory work can begin before the building is ready to receive equipment. That overlap matters more than the speed of any individual assembly task.

Testing also moves earlier. Technicians can identify wiring, control, or component problems before a module travels to a site with a crowded schedule.

Cooling systems offer a second example, although their modular value varies. A cooling distribution unit controls liquid movement between facility infrastructure and heat-producing computing equipment.

Some suppliers mount pumps, tanks, controls, heat exchangers, and leak detection on one skid. Site crews then connect power and a limited number of pipes.

Modular cooling becomes more attractive as AI rack density increases. Liquid-cooled deployments require more piping, control points, and coordination than conventional air-cooled rooms.

Factories can preassemble those repeated sections. However, site-specific climate, water, heat rejection, and equipment arrangements still affect the final system.

Prepared data halls extend factory integration closer to the servers. They can include rack positions, busways, cabling paths, containment, and cooling connections.

Schneider describes its prefabricated halls as factory-built IT modules joined at the site. Its systems target deployments in increments of 50 to 200 racks.

That approach turns the hall into a coordinated product. Operators can install servers without rebuilding every supporting interface in the field.

Whole-facility blocks go further by combining the data hall with power and cooling. They still travel as several transportable sections, despite the datacenter-in-a-box description.

Transportation becomes part of the engineering. Designers must account for road dimensions, bridge limits, lifting capacity, module weight, delivery routes, and assembly tolerances.

A system optimized for shipping may not be optimized for future hardware. Fixed dimensions can limit rack arrangements, cooling layouts, or later retrofits.

This is why the best modular design is not necessarily the most complete module. It is the design that transfers repeatable work without locking the operator into the wrong assumptions.

Reference designs help address that problem. A validated architecture coordinates compute, networking, power, cooling, controls, and structural requirements before each site receives local adjustments.

Factories then produce repeatable units from that architecture. Site engineers adapt the foundation, utilities, permitting documents, and environmental controls.

The mechanism resembles manufacturing more than traditional contracting. Design decisions occur earlier, procurement becomes coordinated, and quality depends on repeatable processes.

Late changes become more expensive under this model. A custom request can affect several modules already moving through production.

Operators must therefore freeze designs sooner. That requirement can conflict with rapidly changing accelerator systems, networking standards, and cooling requirements.

The central bargain is clear. Standardization gains time by reducing field decisions. It also reduces the freedom to redesign the facility after manufacturing begins.

AWS and Meta Are Simplifying Different Layers

AWS and Meta show that modularization is a strategy, not a single product category.

SemiAnalysis describes AWS Project Houdini as a program that prefabricates white-space infrastructure. White space is the area containing racks and their immediate power, cooling, and network connections.

The program reportedly prepares data-hall skids away from the site. SemiAnalysis says this can reduce the period before servers arrive from months to weeks.

AWS retains substantial control over the design and equipment selection. Cupertino Electric reportedly supports the program as a design partner, while AWS manages procurement at hyperscaler scale.

This operator-led model preserves flexibility over components and architecture. It also places inventory, scheduling, integration, and quality risk directly on the operator.

AWS has also developed a newer modular design that SemiAnalysis identifies as SAMDC. The analysis describes narrower, repeatable structures coordinated around their internal systems.

A narrower hall reduces structural spans and unnecessary building volume. It can also lower the number of interfaces that field crews must complete.

Meta attacks the shell more visibly. Its Prometheus campus in New Albany, Ohio, uses aluminum-framed, fabric-clad structures that resemble large industrial tents.

SemiAnalysis reports that each structure covers approximately 125,000 square feet. Satellite observations showed eight structures standing by April 2026, following the project's July 2025 announcement.

That timeline does not represent eight completed datacenters. The enclosure is only one layer within a functioning facility.

Power distribution, cooling, network connections, controls, server installation, testing, and commissioning must still follow. Utility readiness can remain the controlling schedule.

The fabric enclosure also carries different durability and flexibility assumptions from a conventional concrete building. Faster weather protection does not automatically provide equal lifecycle performance.

Meta's approach illustrates design simplification rather than complete modularization. It reduces the shell's material and construction demands so crews can focus on the systems that make compute operate.

AWS moves more integrated work into prefabricated halls. Meta removes complexity from the building around those systems.

Neither strategy proves that one modular format will dominate. Both show hyperscalers redesigning facilities around speed instead of adapting conventional buildings.

Colocation providers face a different problem. They serve multiple customers whose equipment, redundancy, security, and deployment schedules may vary.

A hyperscaler can standardize around its own requirements. A colocation operator must preserve enough flexibility to support customers that have not arrived when construction begins.

Equipment vendors offer another route. Vertiv, Schneider Electric, Eaton, and Flex can supply several layers, including power modules, cooling systems, prepared halls, and integrated blocks.

SemiAnalysis estimates that Vertiv's potential content rises from about 3.5 million per megawatt in traditional projects to roughly 7 million per megawatt with modular solutions.

That figure represents SemiAnalysis modeling, not a guaranteed contract value. It captures why vendors want to own more integration.

A supplier that once sold components can now package engineering, assembly, controls, testing, and commissioning. More responsibility produces more revenue per project, alongside greater execution risk.

Integrators sit between operator control and vendor ownership. They can assemble owner-selected equipment while providing manufacturing capacity and production discipline.

This competitive map matters because the winner may not be the company with the best individual component. It may be the organization that manages interfaces across several components.

That shift places pressure on conventional contractors. Their field expertise remains necessary, but customers increasingly want predictable manufacturing alongside jobsite execution.

It also pressures smaller modular vendors. A good cooling skid must fit a larger electrical, structural, controls, and commissioning plan.

The LEGO metaphor hides this difficulty. Real modules are not interchangeable plastic bricks. Every connection carries requirements for power, water, data, safety, control, and physical tolerance.

The Modular Label Hides Uneven Risks

The modular market is advancing faster than its definitions, making vendor comparisons unusually difficult.

The semianalysis wild thesis starts with a terminology problem. Vendors use modular to describe products ranging from an equipment skid to an entire prefabricated facility.

Those offerings address different constraints. Comparing their schedule claims without examining scope creates a misleading result.

A factory-built electrical room might shorten one critical path. It cannot resolve delayed grid interconnection, land approval, fiber access, or local permitting.

A rapid enclosure can protect crews from weather sooner. It cannot guarantee that cooling equipment or transformers arrive on schedule.

A complete facility block transfers more work into the factory. It also creates greater transportation, coordination, and single-vendor exposure.

Marketing percentages often use different starting points. One supplier may measure manufacturing time, while another measures the period from design approval to commissioning.

Project maturity also affects the baseline. Repeating a proven module should outperform a first deployment that requires new engineering, tooling, and certification.

Factory capacity can become another bottleneck. Moving work away from the site does not create unlimited assembly lines, skilled technicians, or component inventory.

Concentrated manufacturing introduces geographic risk. A disruption at one factory can affect several projects that share the same design.

Large modules must reach their destinations. Heavy-haul permits, road access, cranes, staging areas, and delivery sequencing can all constrain deployment.

The design-freeze problem is equally serious. AI hardware evolves quickly, while a large facility can remain in planning and construction for years.

A module designed for one rack density may require modification when customers adopt hotter accelerators or different cooling connections. A standardized system can then become a standardized mismatch.

Reference designs reduce coordination errors, but they do not remove this timing risk. Operators need clear upgrade paths for power, cooling, networking, and controls.

Quality claims deserve similar scrutiny. Controlled factory conditions support repetition and testing, but poor processes can repeat defects across many units.

Field-built systems can be inspected as they develop. Factory systems require buyers to understand vendor testing, documentation, traceability, and acceptance procedures.

Integration responsibility must also remain explicit. When several suppliers contribute modules, someone must guarantee that their controls and physical interfaces work together.

The operator-led model keeps that accountability close to the owner. It requires deep engineering talent and purchasing leverage that few companies possess.

An integrator-led model consolidates coordination but adds another organizational boundary. Performance depends on the integrator's authority over suppliers and design changes.

An equipment-vendor model offers a more unified package. It can also deepen dependency on one supplier's product roadmap, factory capacity, and service organization.

Cost claims face the same qualification. SemiAnalysis estimates modular construction is about 8 percent cheaper per megawatt, while Schneider describes comparable costs in some scenarios.

Both positions can be reasonable because schedule compression has value beyond construction spending. Earlier operation can produce revenue or strategic capacity sooner.

Yet speed has little value if electricity is unavailable. A finished module waiting beside an unenergized site remains idle capital.

Land represents another tradeoff. Single-story halls and lightweight structures often need more acreage than dense, multistory facilities.

That design works in regions with available land. It fits less comfortably in constrained urban markets where operators must maximize capacity per acre.

Durability also matters. Lightweight shells can deliver fast enclosure, but their maintenance, weather resistance, and retrofit options require long-term evidence.

The cautious conclusion is not that modular claims are empty. It is that every claim needs a defined scope, baseline, and responsible party.

Buyers should ask what moved into the factory, which schedule segment changed, and which constraints stayed on-site. They should also identify who owns final system performance.

What the Next Three Months Should Reveal

Three signals will show whether modular construction is relieving the AI infrastructure bottleneck or merely relocating it.

The first signal is factory throughput for integrated power and cooling systems. Announcements matter less than delivered modules that pass acceptance testing and enter service.

Backlogs, production expansions, and delivery intervals will indicate whether vendors can scale alongside hyperscaler demand. Longer queues would weaken the labor-relief thesis.

Stable delivery with repeatable quality would strengthen it. That result would show that manufacturing capacity can absorb work previously assigned to local construction crews.

The second signal is measurable progress from enclosure to energized compute at rapid-build campuses. Meta's New Albany structures provide a visible test.

Observers should separate structural completion from operational capacity. The useful milestone is not another standing shell, but systems installed, commissioned, and serving computing workloads.

Short intervals between enclosure and operation would support the simplified-shell strategy. Persistent delays would show that power, equipment, or commissioning still controls the schedule.

The third signal is greater disclosure around vendor scope and performance baselines. Customers need schedule comparisons that identify design start, site readiness, installation, and commissioning.

Specific definitions would strengthen the semianalysis wild argument by turning modular from a marketing category into a measurable construction method.

Vague claims would weaken confidence, especially when suppliers compare products covering different portions of a facility.

Readers should also watch who captures integration responsibility. Contract announcements involving complete power, cooling, and white-space packages will reveal where buyers place execution risk.

That decision affects far more than construction companies. Developers, cloud customers, and enterprise technology buyers all depend on when new computing capacity becomes usable.

Delayed infrastructure can limit access to accelerators, extend cloud capacity shortages, and complicate AI deployment plans. Faster delivery can change when teams receive the resources they reserved.

Knowledge workers may experience this indirectly through product availability and service reliability. Developers will see it through cluster access, regional capacity, and deployment schedules.

Tracking such projects requires joining construction updates, vendor claims, permits, and technical documents across many sources. A searchable knowledge base can help teams preserve that evidence and revisit earlier assumptions.

The practical question is no longer whether modular construction exists. It is which parts produce repeatable gains without shifting risk into factories, logistics, or rigid designs.

Watch completed, energized capacity rather than announced gigawatts. Compare identical project stages, and treat every speed claim as incomplete until its scope is clear.

The LEGO datacenter is becoming real, but the bricks remain highly specialized. Over the next quarter, follow factory deliveries, operational milestones, and contract scope. Those signals will show whether modularization has solved the labor problem or only moved its boundaries.