The AI Buildout Is Constrained by Its Slowest Layer

An accelerator can be allocated before the power to run it exists.

That is the constraint shaping the AI buildout. Demand can move in quarters. The transformers, substations, transmission lines, cooling plants, permits, and construction crews underneath it often move on multi-year delivery schedules.

The result is not one shortage. It is a synchronization problem across industries that were never required to scale together at this speed.

The useful unit of progress is therefore not an ordered GPU or an announced megawatt. It is energized, cooled, networked, commissioned compute that can complete useful work.

TL;DR

• AI capacity depends on a chain of constrained systems; the slowest unfinished layer determines when the entire investment becomes usable.
• U.S. data-center electricity use reached 176 TWh in 2023 and is projected to reach 325–580 TWh in 2028, but new generation and grid connections cannot be added at semiconductor speed.
• Leading-edge logic, HBM, advanced packaging, optics, transformers, switchgear, cooling equipment, skilled labor, land, permits, and capital can each become the binding constraint.
• Heavy-duty gas turbines and their specialized blade-and-vane castings become additional bottlenecks when a campus depends on new onsite or utility generation.
• “Announced capacity” hides several unfinished states between a planned campus and commissioned compute.

Demand reached the physical infrastructure

The U.S. Department of Energy estimates that data centers consumed 176 TWh in 2023—4.4% of U.S. electricity consumption. Its 2028 projection ranges from 325 to 580 TWh, or 6.7% to 12% of national electricity use.

xychart-beta
    title "U.S. data-center electricity use: measured and projected"
    x-axis ["2023 actual", "2028 low", "2028 high"]
    y-axis "TWh" 0 --> 600
    bar [176, 325, 580]

The range is wide because the future workload mix, hardware efficiency, utilization, and construction pace remain uncertain. The direction is not. The International Energy Agency expects global data-center electricity consumption to more than double by 2030, with AI as the main driver of the increase.

The demand also lands unevenly. A national grid can have enough annual energy while a specific utility territory lacks a substation, transmission path, fault-current margin, or generation reserve for another large campus. PJM's large-load analysis associated roughly 30 GW of its projected 2024–2030 peak-load growth with data centers.

AI demand stopped being only a semiconductor forecast. It became a regional infrastructure forecast.

Capacity has six different meanings

A useful analytical model separates announced campuses, contracted power, installed servers, and available compute into six states.

This state model explains why the same project can be described as both “under construction” and “capacity constrained.” Each description can refer to a different boundary.

flowchart LR
    A["Announced campus"] --> B["Contracted land<br/>and power"]
    B --> C["Powered data hall"]
    C --> D["Energized rack"]
    D --> E["Commissioned cluster"]
    E --> F["Useful workload capacity"]

    S["Silicon + HBM<br/>+ packaging"] --> D
    N["Switches + optics<br/>+ storage"] --> E
    P["Generation + grid<br/>+ transformers"] --> C
    M["Cooling + water<br/>+ heat rejection"] --> D
    O["Software + operators<br/>+ runbooks"] --> F

The final state requires every incoming path. A missing transformer can strand a completed building. A missing optical tier can strand installed accelerators. A cooling-control problem can strand an energized rack.

The binding constraint migrates

No layer remains the permanent bottleneck.

Fixing one constraint exposes the next. More packaged accelerators increase demand for optical ports. Denser racks increase demand for liquid cooling and high-voltage equipment. More completed buildings increase demand for grid connections. More energized clusters increase the need for operators, scheduling discipline, and workloads that can use them.

This is why a single capacity number cannot describe the buildout. Capacity has a state, a boundary, a location, and a date.

The supplier map is concentrated by layer

The companies below are major suppliers to track, not a market-share ranking. The relevant boundary changes by row: a vendor can lead one component while depending on another supplier for the finished system.

This map also shows why “the power supplier” is too broad a category. A transformer manufacturer, turbine OEM, blade foundry, generator supplier, cooling vendor, utility, and construction contractor own different pieces of the same energization date.

The chip supply chain is a coordinated stack

The Semiconductor Industry Association's 2026 AI supply-chain report estimates that one AI server rack contains more than 4,500 packaged chips assembled from approximately 20,000 individual dies. It also estimates that semiconductors account for more than 95% of the rack's content value.

The separate AI supply-chain article covers logic, HBM, packaging, substrates, assembly, and test. The constraint-system point is narrower: server delivery depends on the last required component, not the first completed one. More logic wafers do not create deployable accelerators if HBM or advanced packaging remains behind.

The rack changed the electrical and cooling contract

The NVIDIA GB200 NVL72 moved the top-end scale-up domain into a rack with 72 GPUs and approximately 120 kW of documented rack consumption. ASHRAE's AI data-center framework discusses rack densities moving toward several hundred kilowatts, with megawatt-class designs on the roadmap.

The existing rack deployment article owns the power and cooling mechanics. The scheduling consequence is enough here: a data hall can have room for more racks while lacking the electrical distribution, cooling distribution units, heat rejection, or utility allocation to operate them.

Networking can strand installed accelerators

The interconnect article owns the fabric mechanics. The deployment constraint includes switch silicon, network adapters, optical transceivers, fiber, topology, firmware, and collective-communication software arriving as one qualified system.

Corning's NVIDIA infrastructure guide estimates that an AI data center can require more than ten times the fiber of a traditional data center. A cluster can therefore boot while remaining unable to scale: missing links, incorrect topology, or scheduler placement can leave installed accelerators waiting for data or synchronization.

Transformers turn power into a delivery schedule

A utility commitment is not a cable waiting to be connected. Power must travel through generation, transmission, substations, transformers, switchgear, protection systems, UPS equipment, generators, busways, power shelves, and point-of-load conversion.

The Department of Energy's supply-chain analysis reports that distribution-transformer lead times increased from three to six months in 2019 to 12–30 months in 2023. A separate Department of Energy resilience report says 36-month lead times for large power transformers are commonly quoted, with maximum waits reaching 60 months.

Large power transformers are not interchangeable catalog parts. Voltage, power rating, impedance, insulation, cooling, fault behavior, noise, footprint, utility standards, and transport limits shape each installation.

The transformer therefore does more than step voltage down. It converts an AI construction plan into a manufacturing and logistics schedule.

Grid connection is a separate constraint from generation

A region can add generation and still lack the transmission or substation capacity to deliver it to a campus. A project can secure a grid connection and still face generation-adequacy questions during peak demand.

JLL's 2026 data-center outlook reports average grid-connection waits above four years in major markets. The same report describes operators pursuing onsite generation and “bring your own power” arrangements to bridge or avoid grid delays.

Onsite power changes the dependency map; it does not remove it. Gas turbines, fuel cells, batteries, reciprocating engines, and microgrids introduce fuel supply, emissions permits, maintenance, noise, redundancy, independent restart, and operation during a grid outage. A behind-the-meter system still needs an operating contract for the transition between local generation and the grid.

This is where nominal megawatts become misleading. The questions are when the power is available, under which contingencies, at what duration, with which emissions and fuel constraints, and whether cooling remains available during the same event.

Generation equipment has its own manufacturing bottleneck

A campus that depends on new gas generation enters another concentrated supply chain.

GE Vernova's second-quarter 2026 results say the company expects at least 125 GW of gas equipment under contract by the end of 2026. Its planned annual gas-turbine output rises from 20 GW in the third quarter of 2026 to 24 GW in 2028 and 30 GW in 2030.

Mitsubishi Power says its ten-year forecast for turbines above 100 MW nearly doubled within one year, stretching component suppliers and skilled labor. Siemens Energy describes a data-center power connection projected to take five years, including two years of equipment lead time.

Those are vendor disclosures, not a matched market-share study. They still establish the scheduling problem: turbine output is expanding, but new manufacturing capacity arrives more slowly than proposed generation demand.

The upstream constraint can sit inside the turbine. Hot-section blades and vanes use nickel-based superalloys, internal cooling passages, thermal coatings, and specialized casting processes. A U.S. Department of Energy ARPA-E project described lead times above one year for complex industrial-gas-turbine blade castings and targeted a reduction below three months through a new casting process.

A Wall Street Journal report syndicated by MSN, citing SemiAnalysis, identifies four companies that concentrate Western production: Howmet Aerospace, Precision Castparts, Consolidated Precision Products, and Doncasters. Each company's public product material confirms that it manufactures industrial-gas-turbine airfoils, blades, or vanes.

That evidence does not establish that only four suppliers exist worldwide, nor does it provide a combined utilization measure proving that every production line is fully booked. The defensible conclusion is narrower: Western production is concentrated, and qualified blade-and-vane casting capacity can govern turbine delivery even when the turbine OEM expands final assembly.

This constraint applies when new gas generation is on the project's critical path. Backup generator sets have a broader supplier base, while large continuously operating plants still need turbines or engines, generators, fuel delivery, emissions permits, transformers, switchgear, construction, controls, and commissioning.

Cooling moves the constraint into climate, water, and land

Liquid cooling moves the constraint rather than removing it. Climate, land, water availability, pumps, heat exchangers, and the selected heat-rejection system still determine how much rack capacity a site can operate.

The U.S. Government Accountability Office found that estimates of generative AI's energy and water use vary widely because the necessary data remains limited. A closed technology loop can circulate coolant without consuming its full flow volume, while the facility-side heat-rejection design may still consume water through evaporation and treatment. “Liquid cooled” therefore does not establish site water consumption or total cooling energy.

Construction is constrained by people, equipment, and permits

The physical build requires utility engineers, electrical contractors, pipefitters, controls specialists, network installers, commissioning teams, and operators trained on high-density liquid-cooled systems.

JLL estimates that average global construction cost reached $10.7 million per MW in 2025 and forecasts $11.3 million per MW in 2026, before tenant IT equipment. Its estimate for AI hardware and fit-out reaches as high as $25 million per MW.

Those numbers make sequence risk expensive. A late utility milestone can leave construction capital idle. A cooling redesign can delay rack acceptance. A new accelerator generation can change power and thermal requirements before the earlier facility design reaches service.

CBRE's 2025 market assessment also identifies permitting, zoning, and adequate power sourcing as contributors to longer construction timelines. The exact approval path varies by jurisdiction.

The buildout is constrained by social and regulatory capacity as well as engineering capacity.

What's missing

Public reporting remains strongest at the two ends of the chain: semiconductor shipment claims and campus megawatt announcements. The middle is harder to see.

The missing dataset includes contracted versus deliverable power, transformer and turbine delivery dates, qualified blade-casting capacity, rack-ready electrical capacity, commissioned liquid-cooling capacity, installed optical ports, cluster acceptance results, actual utilization, useful work completed, facility energy, water consumption, and failure recovery.

Without those boundaries, the market can count a planned megawatt and a productive megawatt as though they were the same asset.

So what

The AI buildout is an exercise in synchronized delivery.

The chip, rack, network, cooling plant, electrical path, grid connection, permits, software, and operating team form one capacity chain. The fastest layer cannot compensate for the slowest unfinished one.

That changes the planning question. The question is not, “How many accelerators can be ordered?” It is, “How much useful compute can be energized, commissioned, and kept productive under a shared physical envelope?”

The unresolved problem is who owns that full-chain promise. Chip companies publish system designs, utilities govern delivered power, equipment vendors qualify individual layers, and operators integrate the result.

The sources reviewed for this series do not expose a shared metric for converting an announced megawatt into sustained, useful AI work. They also do not show how long that conversion takes.

Part of the AI Compute Landscape and the Rack Infrastructure chapter. The physical transition begins with The AI Rack Became a Data Center Design Problem.