CompanyGraph
Electrical power availability gates facility siting more than any other factor, cooling requirements constrain density and geography, and fiber connectivity determines which locations can serve latency-sensitive workloads.
How power availability, semiconductor dependency, and thermal density determine where compute infrastructure exists and how fast it can grow.
Introduction
A supply chain is the sequence of transformations that turns raw materials into usable products and services. The data center supply chain converts copper, silicon, steel, concrete, water, and electricity into compute capacity — the ability to process, store, and transmit data. The physical output is a building full of servers connected to power and network infrastructure. The functional output is the computation that those servers perform.
This supply chain is unusual because the facility itself is not the product. A data center is a container for compute, and compute is only useful when three conditions are met simultaneously: processors are available to install, electricity is available to run them, and cooling is available to remove the heat they generate. A facility with servers but insufficient power is a warehouse. A facility with power but no processors is an empty room. A facility with both but inadequate cooling throttles its own output. All three constraints must be satisfied together, and each draws on a different upstream supply chain with its own bottlenecks and timelines.
The AI demand surge that began in 2023 compressed what had been a steady, predictable growth curve into a step function. Data center capacity that was planned over five-to-ten-year horizons is now being demanded in two-to-three-year windows. This compression exposes the physical constraints that steady growth had obscured — constraints that cannot be resolved by spending more money, because they are rooted in the physics of power delivery, semiconductor fabrication, and thermodynamics.
Root Constraints
Why does securing electrical power gate where a data center can be built?
A single hyperscale data center can consume 100 to 300 megawatts of electrical power — equivalent to a small city. A campus of multiple facilities at a single site can exceed one gigawatt. This is not a transient peak load. Data centers consume power continuously, 24 hours a day, 365 days a year, with load factors above 90 percent. The electrical grid was not designed for concentrated loads of this magnitude at single points of connection.
Securing electrical capacity is now the binding constraint on data center deployment in most regions. Grid interconnection — the process of connecting a new large load to the transmission system — requires studies, upgrades, and regulatory approvals that stretch three to five years in many parts of the United States and Europe. In Northern Virginia, the largest data center market in the world, Dominion Energy has publicly stated that new large-load interconnections face multi-year queues. The same pattern is emerging in Dublin, Amsterdam, Singapore, and other established data center markets where available grid capacity has been absorbed.
The constraint is not generation capacity in the abstract. It is the specific ability to deliver power at the required voltage and reliability to a specific location. A region may have adequate total generation but lack the transmission and substation infrastructure to deliver hundreds of megawatts to a single point. Building that infrastructure — new substations, upgraded transmission lines, additional generation — involves the same permitting and construction timelines that constrain the electricity grid supply chain broadly.
This has produced a structural shift in how data centers are sited. Historically, proximity to population centers and fiber connectivity drove location decisions. Increasingly, power availability dominates. Hyperscalers are siting facilities near power plants — including nuclear stations — because the electrical capacity exists there even when it does not exist in traditional data center markets. Microsoft’s agreement to purchase power from the restarted Three Mile Island Unit 1 reactor, and Amazon’s acquisition of a data center campus adjacent to a Pennsylvania nuclear plant, are structural responses to this constraint, not branding exercises.
The power constraint also interacts with cost structures. Electricity is the largest ongoing operating expense for a data center, typically 40 to 60 percent of total operating cost. As AI workloads increase power consumption per rack from 10-20 kilowatts to 40-100 kilowatts or more, the economics tilt further toward power as the dominant cost and constraint. Operators who secured favorable power contracts and grid connections before the current demand surge hold structural advantages that new entrants cannot replicate quickly.
Why does a data center inherit the semiconductor supply chain's constraints?
The compute hardware inside a data center — CPUs, GPUs, TPUs, custom ASICs, networking chips, memory — is produced by the semiconductor supply chain. The data center supply chain inherits every constraint that the semiconductor supply chain carries: geographic concentration of fabrication in Taiwan and South Korea, single-source dependencies for lithography equipment, and lead times measured in quarters to years for advanced processors.
AI and machine learning workloads have concentrated demand on a narrow category of processors. Nvidia’s GPU architectures dominate AI training and inference, with TSMC as the sole fabrication source for leading-edge designs. Google’s TPUs, Amazon’s Trainium and Inferentia chips, and Microsoft’s Maia accelerators are all fabricated at TSMC. This means the data center industry’s ability to deploy AI compute is gated by TSMC’s advanced node capacity — specifically, the allocation of wafer starts on the most advanced process nodes.
The dependency runs deeper than processors alone. High-bandwidth memory (HBM), required for AI accelerators, is produced by only three companies: SK Hynix, Samsung, and Micron. HBM production involves stacking memory dies and connecting them with through-silicon vias — a process with lower yields than standard memory fabrication. HBM supply has been a binding constraint on GPU production since 2023, and capacity additions take 12 to 18 months to bring online.
Networking equipment within data centers — switches, optical transceivers, cables — depends on its own set of specialized semiconductors and components. As data center networks scale to handle AI training workloads that move terabytes between servers, networking hardware has become another potential bottleneck. Optical transceiver production, particularly for 800G and 1.6T speeds, depends on specialized semiconductor lasers and photonic integrated circuits with limited supply bases.
Why is removing heat a hard physical limit on a data center?
Every watt of electricity consumed by a server becomes heat that must be removed from the facility. This is not an engineering approximation — it is the first law of thermodynamics. A 100-megawatt data center produces 100 megawatts of heat, continuously. Removing that heat is a physical requirement, not an operational preference. If cooling fails, servers throttle performance within seconds and shut down within minutes to prevent hardware damage.
Traditional data center cooling uses air — cold air is pushed through server racks, absorbs heat, and is returned to cooling systems (typically chillers that reject heat to the outside air or to water). This approach works for power densities up to roughly 15-20 kilowatts per rack. AI accelerator racks operate at 40 to 100 kilowatts or more, with next-generation systems targeting 120+ kilowatts per rack. At these densities, air cooling becomes physically inadequate — the volume of air required to remove the heat cannot be moved through the rack quickly enough.
This has driven adoption of liquid cooling. Direct-to-chip liquid cooling circulates fluid through cold plates attached directly to processors, removing heat at the source with far greater efficiency than air. Immersion cooling submerges entire servers in dielectric fluid. Both approaches can handle power densities that air cooling cannot, but they require different facility designs, different plumbing infrastructure, and different operational expertise than the air-cooled data centers the industry has built for three decades.
The transition to liquid cooling is not a simple upgrade. It requires redesigning server hardware, modifying or replacing rack infrastructure, installing liquid distribution systems, and training operations staff. Facilities designed for air cooling cannot be easily retrofitted for liquid cooling at scale. This means that much of the existing data center inventory — optimized for air-cooled workloads at 10-20 kilowatts per rack — cannot efficiently host the highest-density AI workloads without significant modification.
Water consumption is a related constraint. Evaporative cooling, which many data centers use to reject heat to the atmosphere, consumes significant quantities of water. A large data center can consume one to five million gallons of water per day, depending on climate and cooling design. In water-stressed regions, this consumption faces growing regulatory scrutiny and competition with agricultural, municipal, and industrial users. Some operators have shifted to closed-loop cooling systems that do not consume water, but these systems are less energy-efficient, increasing the power required for cooling and tightening the power constraint further.
The Physical Chain
What raw materials does the data center supply chain begin with?
The data center supply chain begins at mines and refineries. Copper — for electrical wiring, busbars, heat exchangers, and cabling — is consumed in enormous quantities. A single large data center can use thousands of tons of copper. The copper supply chain has its own constraints: mining permits take years, new mines take five to ten years to develop, and global copper demand is growing across multiple sectors simultaneously (electrification, EVs, grid expansion, construction).
Steel and concrete form the structural envelope. While these materials are generally available, the scale and pace of data center construction have created localized supply pressures in markets with heavy building activity. Diesel generators for backup power require engines and fuel systems. Uninterruptible power supply (UPS) systems require batteries — increasingly lithium-ion, connecting the data center supply chain to the EV battery supply chain.
Rare earth elements appear in the permanent magnets used in cooling fans, hard drives, and some power equipment. Fiber optic cables require ultra-pure glass and specialized coatings. Each of these material inputs connects the data center supply chain to other supply chains, inheriting their constraints.
Where are data center servers manufactured and assembled?
Server manufacturing assembles processors, memory, storage, and networking components onto motherboards and into chassis. This assembly is concentrated in a small number of contract manufacturers — primarily Foxconn, Quanta, Wistron, and Inventec — operating factories in Taiwan and increasingly in Mexico and other locations. Server design is often customized for hyperscale operators, with proprietary specifications that limit manufacturing flexibility.
Power distribution equipment — transformers, switchgear, power distribution units — has its own manufacturing supply chain. Large power transformers are built by a small number of manufacturers globally, with lead times that have stretched to two years or more as demand from both data centers and grid expansion competes for production capacity. This is a quiet bottleneck: a data center cannot operate without power distribution equipment, and the equipment cannot be produced faster than transformer manufacturing capacity allows.
Cooling equipment manufacturing — chillers, cooling towers, liquid cooling distribution units, cold plates — is scaling to meet new demand but faces the typical ramp-up timeline of industrial manufacturing. Liquid cooling in particular is transitioning from niche to mainstream, and the manufacturing base is still developing.
How long does it take to construct and commission a data center?
Data center construction involves site preparation, structural building, electrical infrastructure installation, cooling system installation, and IT equipment deployment. Construction timelines for a new facility typically run 18 to 36 months from ground-breaking to first server deployment. Modular and prefabricated approaches can compress this to 12 to 18 months for standardized designs.
But construction is rarely the longest timeline. Power interconnection, as noted, often takes longer. Permitting — environmental review, zoning approval, building permits — adds months to years depending on jurisdiction. In some European markets, moratoriums on new data center construction have been imposed or considered due to power grid strain, creating regulatory constraints on top of physical ones.
Commissioning — the process of testing and validating all systems before production workloads begin — takes weeks to months. Power systems must be tested under load. Cooling systems must demonstrate they can handle design conditions. Network connectivity must be proven end-to-end. A data center that is physically complete but not commissioned cannot serve customers.
Structural Patterns
- Geographic Clustering Around Infrastructure — Data centers cluster where power, fiber, and favorable conditions coexist. Northern Virginia hosts the densest concentration of data centers in the world because it sits at the intersection of major fiber routes (connecting to transatlantic cables), historically abundant and affordable power, and proximity to federal government customers. New clusters are forming around power sources — nuclear plants, hydroelectric dams, natural gas pipelines — rather than around population centers.
- Hyperscaler Vertical Integration — The largest operators — Amazon Web Services, Google, Microsoft Azure, and increasingly Meta and Oracle — are integrating vertically in response to supply chain constraints. They design custom chips, negotiate directly with semiconductor foundries, develop proprietary cooling systems, sign power purchase agreements directly with generators, and in some cases build their own electrical substations. This vertical integration is a structural response to supply chain constraints, not a strategic preference. When the market cannot deliver inputs at the required scale and pace, large operators build their own supply.
- The AI Demand Step Function — Pre-2023, data center demand grew at a steady and predictable rate driven by cloud computing, streaming, and enterprise digitization. The AI training and inference surge compressed years of anticipated demand growth into months. This step-function increase in demand hit a supply chain with multi-year response times, creating structural shortages across power, semiconductors, and cooling equipment simultaneously.
- Power as the New Real Estate — In the traditional data center market, location value was determined by connectivity and proximity. In the current market, the most valuable asset a data center operator can hold is secured electrical capacity — a grid interconnection agreement with confirmed megawatts. Operators with existing power capacity can deploy new compute faster than operators who must wait in interconnection queues, regardless of how much capital the latter can deploy.
- Construction Timelines Shorter Than Supply Timelines — The physical building can be erected faster than power can be interconnected, faster than GPU allocations can be secured, and faster than liquid cooling infrastructure can be manufactured at scale. The building is no longer the critical path. This inverts the traditional real estate development model where construction is the longest and most capital-intensive phase.
Flows and Visibility
Capital flows in the data center supply chain have reached extraordinary scale. Hyperscalers are collectively spending over $200 billion annually on capital expenditure, a significant fraction directed at data center infrastructure. This capital intensity creates barriers to entry — not because the technology is secret but because the financial commitment required to secure power, land, and equipment at scale excludes most participants.
Information flows are asymmetric. Hyperscalers have detailed visibility into their own demand forecasts, chip allocation schedules, and power procurement pipelines. Colocation operators and smaller cloud providers have less visibility and less negotiating power with upstream suppliers. The semiconductor companies — Nvidia, AMD, Intel — have visibility into their order books but limited visibility into how their customers’ demand forecasts will evolve. This asymmetry means that demand signals propagate imperfectly through the chain, creating the conditions for both shortages and eventual oversupply.
Material flows are slow relative to demand changes. A decision to build a new data center campus today initiates supply chain processes that unfold over two to five years: power procurement, land acquisition, permitting, construction, equipment procurement, and commissioning. If demand shifts during that period — as it can when technology adoption curves are uncertain — the committed capital cannot be easily redirected.
What the AI Surge Has Revealed
The acceleration of AI workloads beginning in 2023 functioned as a stress test for the data center supply chain. Several structural features became visible under this stress that were obscured during steady-state growth.
First, the coupling between the data center and electricity grid supply chains became apparent. Data center operators had historically been large but manageable loads on the grid. The step-function increase in demand made data centers the dominant source of new electrical load growth in many regions, fundamentally changing the relationship between data center operators and utilities. Grid planners who had forecasted steady load growth found their projections obsolete within a single planning cycle.
Second, the semiconductor supply chain’s concentration became a direct constraint on data center deployment. When Nvidia GPUs became the bottleneck, data center operators discovered that their deployment timelines were governed not by their own construction capabilities but by TSMC’s fabrication schedule and Nvidia’s allocation decisions. Operators who had invested in custom silicon — Google with TPUs, Amazon with Trainium — found themselves with alternative paths, though still dependent on TSMC fabrication.
Third, the cooling transition from air to liquid exposed how much of the existing data center inventory was designed for a workload profile that AI has rendered obsolete for the highest-performance applications. Facilities optimized for conventional cloud workloads at 10-15 kilowatts per rack face structural limitations when asked to host AI training clusters at 60-100 kilowatts per rack. This is not a matter of incremental upgrade — it represents a different class of physical infrastructure.
Fourth, the geographic assumptions underlying data center siting are shifting. Fiber connectivity, which was the primary siting constraint for decades, is being supplemented or overridden by power availability. A location with abundant power but limited fiber can add fiber connectivity in months. A location with abundant fiber but limited power cannot add electrical capacity for years. The relative timeline of these two infrastructure types has inverted their importance in siting decisions.
What This Reveals About Industrial Structure
- Supply chains inherit their upstream constraints — The data center supply chain does not exist independently. It inherits constraints from the semiconductor, electricity, copper, water, and construction supply chains. A bottleneck in any upstream chain propagates forward. This interconnection means that the data center supply chain cannot be analyzed in isolation.
- Physical constraints do not respond to price signals on relevant timescales — Spending more money does not make the grid interconnection process faster, does not add TSMC fab capacity this quarter, and does not change the thermodynamic properties of cooling systems. Capital can fund future capacity but cannot compress the physical timelines for deploying it. The gap between when money is committed and when capacity materializes is structural.
- Step-function demand changes expose constraints that steady growth conceals — A supply chain calibrated for 10-15 percent annual growth cannot absorb a sudden doubling of demand. The AI surge demonstrated that the same system that appeared adequate under steady growth can become severely constrained under step-function demand — not because it was poorly designed but because it was designed for a different demand profile.
- Vertical integration is a constraint response — When market mechanisms cannot deliver inputs at the required scale and pace, large operators integrate vertically. Hyperscalers designing their own chips, building their own substations, and signing nuclear power purchase agreements are not pursuing vertical integration as a strategy. They are pursuing it because the alternative — waiting for the market to deliver — imposes timelines incompatible with their demand.
- The binding constraint rotates over time — At any given moment, one constraint dominates. But as investment and innovation relax that constraint, the next-tightest constraint becomes binding. This rotation means that solving any single bottleneck — more GPUs, more power, better cooling — does not resolve the supply chain challenge. It moves the bottleneck to the next constraint in the chain.
Epistemic Limits
This analysis describes structural relationships and constraints as they are observable in the current system. It does not forecast how quickly any particular constraint will relax, whether the AI demand trajectory will continue at its current rate, or which specific investments will prove well-timed. The interactions between constraints are complex enough that predicting second-order effects — for example, whether power scarcity will redirect AI investment to different geographies or to more power-efficient architectures — exceeds what structural observation alone can establish. What the analysis can claim is that the constraints described are physical and real, that they interact in observable ways, and that their timelines are measured in years, not months.
Inside CompanyGraph
View this perspective inside CompanyGraph.