Date: 6 October 2026 Publication Type: Research Project: MaMeeFarm™
Abstract
Artificial intelligence is commonly discussed through models, chips, software, and computational performance.
But operational AI capacity depends on a much larger physical system.
Compute hardware must be installed in servers. Servers must be integrated into racks. Racks require electricity. Electricity must reach the site through transformers, substations, cables, switchgear, and grid connections. High-density computing produces heat that must be removed. Buildings must be constructed. Land must be available. Cooling systems, backup power, networking, and supporting infrastructure must be installed and commissioned.
The growth of AI therefore depends not only on the supply of advanced semiconductors, but on the coordination of multiple physical systems whose development cycles can be substantially slower than technology demand cycles.
Current evidence illustrates this shift. The International Energy Agency estimates that data centres consumed about 415 TWh of electricity globally in 2024 and projects around 945 TWh by 2030 in its Base Case. It also estimates that grid constraints could delay around 20 percent of global data-centre capacity planned for construction by 2030 unless connection risks are addressed.
This research examines the physical infrastructure required to transform computing hardware into operational AI capacity.
The central finding is that AI scaling is increasingly an infrastructure coordination problem.
The presence of chips does not by itself create usable compute.
Operational capacity exists only when computing, electrical, thermal, construction, network, site, and supporting industrial infrastructure become available together.
1. Research Question
What physical infrastructure is required to transform AI compute hardware into operational AI capacity, and how do constraints in electricity, grid connection, cooling, land, construction, networking, and supporting equipment shape the speed and geography of AI deployment?
2. Artificial Intelligence Has a Physical Layer
AI appears digital because its visible outputs are digital.
Models generate text.
Systems process images.
Applications deliver software services.
But every computational operation occurs on physical hardware.
That hardware occupies physical space.
It consumes electricity.
It produces heat.
It depends on networking.
It requires maintenance.
It operates inside a facility.
The physical system behind AI therefore extends far beyond the semiconductor itself.
A simplified physical dependency can be expressed as:
semiconductor → server → rack → data centre → electrical connection → energy supply → cooling → network → operational compute
This is not a universal deployment sequence.
It illustrates the physical dependencies that must ultimately coexist for computational hardware to become usable capacity.
A major constraint in any critical part of that system can restrict operational compute.
3. Chip Capacity Is Not Operational AI Capacity
A large inventory of processors can represent potential computing capacity.
It does not automatically represent operational capacity.
Processors still require:
- compatible servers;
- memory;
- networking;
- rack infrastructure;
- power distribution;
- cooling;
- facility space; and
- electricity.
This creates several distinct physical states:
- available chips;
- installed hardware;
- powered hardware; and
- operational compute.
These should not be treated as equivalent.
AI infrastructure becomes productive only when surrounding systems allow the hardware to operate.
4. Compute Demand and Deployable Capacity Are Different
Demand for AI computation can increase rapidly.
Physical infrastructure responds more slowly.
Companies can decide to train larger models or deploy more AI services quickly.
But a large data centre requires:
- site preparation;
- permits;
- electrical engineering;
- construction;
- utility coordination;
- cooling installation;
- network connectivity; and
- commissioning.
This creates a timing mismatch between digital demand and physical deployment.
The IEA notes that a data centre can become operational in approximately two to three years, while broader energy infrastructure can require substantially longer periods to plan and build.
AI deployment therefore depends partly on whether supporting infrastructure can arrive when computing demand requires it.
5. Electricity Is a Core Production Input
Computing hardware cannot operate without continuous electricity.
Electricity therefore belongs inside the physical production boundary of AI infrastructure.
It is not merely an external utility expense.
AI servers consume electricity directly.
Additional electricity is required for:
- cooling;
- power conversion;
- networking;
- storage;
- pumps;
- fans;
- lighting;
- backup systems; and
- other facility functions.
The IEA estimates that global data-centre electricity consumption was approximately 415 TWh in 2024, around 1.5 percent of global electricity consumption.
Its Base Case projects approximately 945 TWh by 2030.
6. AI Is an Important but Not Exclusive Driver of Data-Centre Demand
Data-centre electricity should not automatically be classified as AI electricity.
Data centres also support:
- cloud services;
- storage;
- conventional computing;
- enterprise systems;
- networking; and
- other digital services.
The IEA distinguishes accelerated servers, largely associated with AI adoption, from conventional servers in its modelling.
In the Base Case, electricity consumption from accelerated servers grows much faster than conventional server electricity consumption and accounts for almost half of the net increase in global data-centre electricity consumption to 2030.
This distinction is important.
Data-centre demand
and
AI-specific demand
overlap, but they are not identical categories.
7. A Small Global Share Can Produce a Large Local Constraint
At the global level, data-centre electricity demand remains a minority share of total electricity consumption.
But data-centre load is geographically concentrated.
A large cluster can create substantial local demand even when the global share appears modest.
This distinction matters because electricity systems are not one unified global network.
Constraints occur in specific:
- regions;
- substations;
- transmission zones;
- utility territories; and
- generation markets.
A relatively small global demand category can therefore become a major local infrastructure problem.
8. The United States Illustrates Concentrated Demand Growth
Lawrence Berkeley National Laboratory’s 2025 Update estimates that U.S. data centres could account for about 11.8 percent of national electricity consumption by 2030, with scenarios ranging from approximately 9.5 to 15.3 percent.
The report uses a bottom-up model incorporating expected equipment shipments, device electricity use, cooling performance, facility types, and locations.
The range is important.
It represents scenarios rather than a guaranteed outcome.
But it demonstrates why national and local planning increasingly need to consider data-centre growth as a physical electricity-demand problem.
9. Grid Connection Can Become the Binding Constraint
A completed data-centre building does not create operational capacity if sufficient electrical connection is unavailable.
The facility may require:
- new transmission;
- substation expansion;
- transformers;
- switchgear;
- distribution upgrades; or
- additional generation.
These components can create a bottleneck independently of chip supply.
The IEA estimates that around 20 percent of global data-centre capacity planned for construction by 2030 could face grid-connection delays if current risks are not addressed.
This creates a fundamental distinction:
constructed data-centre capacity
does not necessarily equal
energized data-centre capacity.
10. Grid Infrastructure Moves More Slowly Than Technology Cycles
Digital technology can change rapidly.
Electricity infrastructure operates on different planning cycles.
The IEA reports that building new transmission infrastructure in advanced economies can take approximately four to eight years.
It also reports that waiting times for critical grid components such as transformers and cables have roughly doubled in recent years.
The result is a coordination problem.
Computing demand can arrive before electrical infrastructure.
11. Lead-Time Misalignment Can Become the Constraint
AI infrastructure contains components with different development cycles.
Semiconductor generations can change quickly.
Servers can be ordered and installed comparatively quickly.
Buildings take longer.
Grid upgrades can take longer still.
Generation projects may require several years.
Transformers, turbines, switchgear, cables, and other equipment may have separate manufacturing queues.
The limiting factor may therefore not be the absolute scarcity of one component.
It may be the inability to align delivery schedules across several systems.
This can be understood as lead-time misalignment.
12. Component Capacity and System Capacity Are Different
Suppose a region has:
- available GPUs;
- available capital;
- available land; and
- a completed data-centre shell.
But its grid connection is delayed.
Operational AI capacity remains constrained.
Alternatively, electricity may be available while cooling equipment is delayed.
Or the building may be ready while networking equipment is not.
This creates another important distinction.
Component capacity means an individual resource exists.
Coordinated system capacity means critical resources exist together in an operable configuration.
AI deployment depends on the second.
13. Transformers and Substations Belong Inside the Effective AI Supply Chain
Transformers are not typically described as AI components.
But they can determine whether AI hardware receives electricity.
The same applies to:
- substations;
- switchgear;
- high-voltage cables;
- circuit breakers;
- power distribution units; and
- other electrical equipment.
If these components become constrained, AI deployment can be delayed even while semiconductor availability improves.
This expands the effective AI supply chain beyond the traditional technology sector.
14. Infrastructure Constraints Can Propagate Back Into the Chip Supply Chain
Physical infrastructure shortages do not necessarily remain isolated within electricity systems.
If power constraints delay data-centre deployment, customers may delay the installation of supporting components.
Reuters reported on 5 October 2026 that Morgan Stanley viewed Nvidia and Broadcom as relatively protected from current power-related deployment delays while warning that suppliers of memory and optical components could be more exposed if delayed AI deployments reduce near-term demand.
This illustrates a plausible reverse transmission:
power constraint → deployment delay → equipment installation delay → effects elsewhere in the semiconductor supply chain
Infrastructure can therefore influence technology demand upstream.
15. Cooling Is a Second Major Physical Constraint
Electricity entering computing equipment eventually becomes heat.
High-density AI systems therefore require substantial heat removal.
A basic physical relationship is:
electricity → computation → heat → heat removal
If cooling capacity is insufficient, available electrical and computing capacity cannot necessarily be fully used.
Thermal infrastructure is therefore part of usable computing capacity.
16. Rack Density Changes Cooling Requirements
AI hardware increasingly concentrates large amounts of computing power inside dense rack configurations.
Higher density can reduce the physical area required per unit of compute while increasing the amount of heat that must be removed from a small space.
Advanced cooling development increasingly reflects this engineering challenge.
As rack power density rises, thermal management becomes a more important constraint on how much installed hardware can operate reliably within a given facility.
17. Electrical Capacity and Thermal Capacity Must Expand Together
A site may have enough electricity to power a high-density AI cluster.
But if its cooling system cannot reject the resulting heat, full operation may still be impossible.
The opposite is also possible.
Cooling infrastructure may exist while electrical supply is insufficient.
Operational capacity therefore depends on both.
A simplified relationship is:
usable compute ≤ available electrical capacity
and
usable compute ≤ available thermal-management capacity
Whichever becomes more restrictive can become the bottleneck.
18. Cooling Requirement Is Universal but Water Consumption Is Not
All computing systems must remove heat.
But cooling technologies differ.
Some systems use evaporative cooling.
Others use:
- closed-loop systems;
- liquid cooling;
- air cooling;
- dry coolers; or
- hybrid designs.
Advanced cooling approaches can reduce energy consumption and, depending on system design, reduce or eliminate direct water consumption for cooling.
It is therefore inaccurate to treat all data centres as having identical water intensity.
A safer distinction is:
cooling requirement is universal; water requirement depends on system design and local conditions.
19. Site Selection Is Becoming an Energy Decision
Traditional site selection might emphasize:
- land;
- taxes;
- network connectivity;
- labor; and
- proximity to users.
AI infrastructure adds another major factor:
available power.
The IEA identifies locating data centres in areas with greater power and grid availability as one option for reducing connection delays.
Electricity availability is therefore moving upstream into investment and location decisions.
20. AI Facilities Are Beginning to Co-Locate With Energy Infrastructure
A contemporary example appeared in Japan in October 2026.
JERA, Dell Technologies, and RHAELM announced plans for a 400 MW hyperscale AI data-centre project in Chiba.
The planned site is near JERA’s thermal power facilities, and JERA is expected to provide land and long-term power capacity.
The project is planned to begin phased operation in 2028 and reach full capacity in 2029.
This structure is significant.
Power is not being treated merely as an input to acquire after the computing facility is built.
Energy access is being incorporated into project development from the beginning.
21. AI Infrastructure Can Reshape the Geography of Computing
If electricity availability becomes a major constraint, AI infrastructure may increasingly move toward places with:
- large power systems;
- available grid capacity;
- generation assets;
- suitable land;
- cooling options; and
- network access.
This can change the geography of the digital economy.
Compute location may become increasingly influenced by physical infrastructure rather than only by software ecosystems or corporate locations.
22. Behind-the-Meter Power Creates an Alternative Architecture
When grid connection becomes slow, developers may seek power outside the conventional grid-connection process.
Possible approaches include:
- onsite generation;
- dedicated generation;
- energy storage; or
- private electrical infrastructure.
These systems can reduce dependence on immediate grid expansion.
But they also change the relationship between electricity producer and electricity consumer.
A data centre may become more closely integrated with generation infrastructure.
23. The Data Centre Can Become an Energy Asset
Large data centres may increasingly contain:
- generation;
- battery storage;
- backup power;
- flexible computing loads; and
- sophisticated energy-management systems.
This creates the possibility that data centres function not only as electricity consumers but, under some configurations, as controllable resources within the energy system.
Operational flexibility can potentially shift some computational demand through time.
However, the extent of that flexibility depends on workload, service requirements, reliability needs, contractual obligations, and technical design.
Not all AI computing tasks are equally flexible.
24. Backup Power Is Part of Operational Capacity
AI services often require high availability.
A brief electricity interruption can affect:
- training jobs;
- inference services;
- network operations; and
- stored data.
Data centres therefore use backup systems such as:
- batteries;
- uninterruptible power supplies; and
- backup generators.
These systems do not normally provide the primary energy supply.
They provide resilience.
Operational AI capacity therefore depends not only on normal electricity supply but also on the ability to maintain service when normal supply is interrupted.
25. Reliability Matters Alongside Electricity Quantity
A data centre requires more than sufficient annual electricity.
It requires electricity with appropriate:
- availability;
- voltage;
- frequency;
- quality; and
- reliability.
A region may have enough total generation over a year while still facing local conditions unsuitable for high-density computing if delivery reliability is insufficient.
Energy quantity alone therefore does not define usable AI infrastructure.
26. Land Is More Than Physical Space
Data centres require land for buildings.
But site requirements can extend beyond the building footprint.
Sites may also require space for:
- substations;
- backup generation;
- cooling equipment;
- storage;
- security;
- network connections; and
- future expansion.
Nearby grid infrastructure and generation can change the economic value of land.
Land with strong electrical connectivity may therefore have a different infrastructure value from land with similar physical characteristics but weak grid access.
27. Network Infrastructure Remains Essential
AI computation requires movement of data.
Data-centre clusters depend on:
- fiber;
- switching;
- optical equipment;
- internal networking; and
- external network connections.
Advanced AI systems also require very high bandwidth between processors and memory and across computing nodes.
Network infrastructure therefore interacts with computational capacity.
A cluster of powerful processors cannot deliver full distributed performance if data movement becomes the limiting constraint.
28. Memory and Interconnection Are Part of the Physical Compute System
AI hardware performance depends not only on arithmetic capability.
It also depends on moving data between:
- memory;
- processors;
- storage; and
- other compute nodes.
This is why high-bandwidth memory and high-speed interconnection have become increasingly important components of AI systems.
The physical architecture of computing therefore includes both processing and data movement.
29. Construction Capacity Can Become a Bottleneck
AI deployment requires physical construction.
Projects depend on:
- engineers;
- electricians;
- construction firms;
- equipment manufacturers;
- commissioning teams; and
- specialized contractors.
Rapid simultaneous development can create shortages even when financing is available.
Infrastructure growth therefore depends on human and industrial capacity as well as technology.
30. Announced Capacity Is Not Operational Capacity
Data-centre developers frequently announce future projects in megawatts or gigawatts.
These announcements indicate development intention.
They do not establish completed capacity.
A project may still require:
- financing;
- permitting;
- site acquisition;
- equipment procurement;
- grid connection;
- construction;
- cooling installation;
- network integration;
- testing; and
- commissioning.
Research should therefore distinguish:
- announced capacity;
- under-construction capacity;
- energized capacity; and
- operational capacity.
Combining them can substantially overstate near-term usable infrastructure.
31. Power Availability Can Shape Project Timing
The planned 400 MW Chiba project illustrates this distinction.
Its announced capacity is 400 MW.
But the project is expected to enter operation in phases beginning in 2028 before reaching full capacity in 2029.
The full announced number should therefore not be interpreted as capacity available in October 2026.
This temporal distinction is essential when analysing rapid AI infrastructure growth.
32. Financing Is Part of the Infrastructure Chain
Large AI facilities require substantial capital before producing computing services.
Investment is needed for:
- land;
- buildings;
- servers;
- power systems;
- cooling;
- network infrastructure; and
- grid upgrades.
The physical system therefore exists within a financing structure.
A technically feasible project can still be delayed or cancelled if financing conditions change.
Announced infrastructure represents a pipeline of potential capacity, not guaranteed future supply.
33. AI Scaling Can Create Infrastructure Competition
Grid capacity used by one large new load may not be simultaneously available to another.
The IEA warns that rapid data-centre load growth may create trade-offs with priorities such as electrification, manufacturing expansion, or affordability if energy-system investment does not keep pace.
This creates a broader systems question.
AI infrastructure does not develop outside the rest of the economy.
It competes for:
- electricity;
- equipment;
- construction labor;
- land;
- grid capacity; and
- capital.
34. Infrastructure Competition Can Become a Coordination Problem
Suppose an electricity system is simultaneously trying to support:
- new factories;
- electric vehicles;
- building electrification;
- data centres; and
- population growth.
Each demand category may be viable individually.
Together they may exceed local infrastructure expansion rates.
The constraint may therefore emerge from simultaneous growth rather than from AI alone.
Research should preserve this distinction when interpreting local power shortages.
35. Efficiency Can Change the Infrastructure Requirement
AI electricity demand is not determined only by the quantity of computing.
Hardware efficiency can improve.
Software can become more efficient.
Cooling can improve.
Utilization can change.
Model architecture can change.
The IEA therefore uses scenarios rather than one deterministic forecast because future data-centre demand remains uncertain.
Infrastructure research should preserve this uncertainty.
Current projections are planning scenarios, not guaranteed outcomes.
36. More Efficient Chips Do Not Automatically Reduce Total Electricity Demand
Higher computing efficiency can reduce energy required per computational operation.
But lower cost per unit of computation can also support larger workloads.
Total electricity use therefore depends on both:
efficiency per unit
and
total quantity of computation demanded.
Rapid AI adoption can increase aggregate electricity demand even while hardware becomes more efficient.
Efficiency and total demand should therefore be analysed separately.
37. Physical Constraints Can Influence AI Economics
Infrastructure delays affect more than engineering.
A GPU that cannot be energized does not generate computing revenue.
A data centre waiting for connection represents capital that cannot yet operate fully.
Higher power costs can increase computing cost.
Cooling infrastructure adds capital and operating expense.
Grid upgrades can affect project economics.
Infrastructure therefore influences the economic productivity and timing of AI investment.
38. Infrastructure Can Determine Which Investment Becomes Productive First
Two companies may purchase similar computing equipment.
One has immediate access to power and cooling.
The other waits for grid connection.
Their hardware capability may be similar.
Their usable capacity is not.
The first converts capital into productive compute earlier.
Physical infrastructure therefore influences the timing of return on AI investment.
39. AI Infrastructure Is a Multi-Industry Production System
The physical system supporting AI connects sectors that are often analysed separately:
- semiconductors;
- electrical equipment;
- construction;
- utilities;
- energy generation;
- cooling;
- telecommunications;
- real estate;
- industrial manufacturing; and
- finance.
AI growth therefore creates demand across a distributed industrial network.
This makes AI infrastructure fundamentally different from a purely software scaling problem.
40. The Effective AI Supply Chain Extends Beyond Technology Companies
The operational AI supply chain may include:
- chip foundries;
- memory producers;
- server manufacturers;
- optical equipment suppliers;
- transformer manufacturers;
- utilities;
- power developers;
- cooling companies;
- engineering firms;
- construction contractors;
- fiber providers; and
- landowners.
A constraint in any critical segment can slow deployment.
Supply-chain research should therefore follow physical dependency rather than industry labels alone.
41. Physical Bottlenecks Can Move
During one phase of expansion, GPUs may be scarce.
Later, memory may become the constraint.
Then transformers.
Then power.
Then cooling.
Then construction labor.
The limiting resource can migrate as investment responds to earlier shortages.
This resembles other infrastructure systems.
Solving one bottleneck can reveal the next.
42. AI Deployment Is Increasingly a Synchronization Problem
A successful large-scale deployment may require several systems to become ready within the same period:
- compute equipment;
- facility construction;
- electrical connection;
- cooling;
- networking; and
- staffing.
If one arrives substantially later than the others, capital can remain underutilized.
The problem is therefore not only building more of each component.
It is making them available together.
This leads to a central research proposition:
AI scaling increasingly depends on coordinated infrastructure timing.
43. Physical Domains Supporting Operational AI
The physical system behind operational AI can be examined across several interacting domains.
Semiconductor Infrastructure
GPUs, accelerators, CPUs, memory, and networking silicon provide computational capability.
Server and Rack Infrastructure
Servers, storage, rack systems, power distribution, and high-speed networking organize computing hardware into usable systems.
Data-Centre Infrastructure
Buildings, electrical rooms, backup systems, physical security, and facility management provide the operating environment.
Thermal Infrastructure
Liquid cooling, air cooling, pumps, heat exchangers, chillers, and heat-rejection systems remove the heat generated by computation.
Electrical Connection Infrastructure
Substations, transformers, switchgear, cables, and utility interconnections connect facilities to electrical supply.
Energy Supply Infrastructure
Generation, storage, onsite power, backup generation, and grid resources provide the energy required for operation.
Site Infrastructure
Land, permitting, local cooling conditions, transport access, and network connectivity affect where and how facilities can be developed.
Industrial Delivery Infrastructure
Construction, engineering, equipment manufacturing, skilled labor, commissioning, and maintenance determine whether planned infrastructure can be physically delivered and operated.
These domains are descriptive rather than sequential.
A facility may develop several of them simultaneously, and their relative importance varies by project and location.
Operational compute emerges from their interaction.
44. Research Considerations for Physical AI Infrastructure
Research into AI infrastructure involves several forms of evidence that should remain distinct.
Computing demand and operational capacity are different quantities.
Demand for additional computation does not establish that sufficient deployable infrastructure exists.
AI demand and total data-centre demand overlap but are not identical.
Data-centre statistics include conventional computing, cloud services, storage, networking, enterprise systems, and other workloads in addition to AI.
Hardware availability is only one part of deployment.
Processors, memory, servers, networking equipment, and racks must ultimately operate within facilities that can power, cool, connect, and maintain them.
Project status affects the meaning of announced capacity.
Announced, permitted, financed, under-construction, energized, and operational capacity represent different states. They should not automatically be aggregated as immediately usable compute.
Electrical connection can become independent of electricity generation.
A region can possess substantial generation while a particular site remains constrained by transmission, substations, transformers, distribution infrastructure, or connection queues.
Thermal capacity can constrain electrically available compute.
A site capable of receiving additional electricity may still be unable to operate additional high-density hardware if cooling and heat-rejection systems are insufficient.
Location changes infrastructure feasibility.
Power availability, grid conditions, land, permitting, network access, cooling options, and local industrial capacity vary geographically.
Supporting supply chains can become binding constraints.
Transformers, switchgear, cables, turbines, cooling equipment, optical systems, construction capacity, and specialized labor can affect deployment independently of semiconductor availability.
Lead times differ across infrastructure systems.
Technology equipment, buildings, grid upgrades, generation projects, and industrial equipment do not necessarily arrive on comparable schedules.
The binding constraint can migrate.
Resolving a semiconductor shortage may reveal a power constraint. Resolving power may reveal cooling, construction, networking, or another constraint.
Operational capacity is the strongest physical endpoint.
Installed hardware represents productive AI capacity only when it can be powered, cooled, connected, and operated under the required conditions.
These considerations do not constitute a universal method for evaluating AI infrastructure.
They identify common distinctions needed when interpreting the physical capacity behind AI deployment.
45. What Current Evidence Can Establish
Current evidence supports several findings.
Global data-centre electricity demand is increasing rapidly, with AI-related accelerated computing an important driver.
Grid constraints can delay data-centre development even when demand and investment remain strong.
Transmission infrastructure and critical grid equipment can have materially longer lead times than data-centre construction.
Advanced AI hardware is creating increasingly demanding thermal-management requirements.
Power access is becoming part of data-centre location and project architecture, as illustrated by projects designed around proximity to generation assets.
Power-related deployment delays can also transmit back into parts of the semiconductor supply chain.
These observations support the conclusion that AI deployment increasingly depends on infrastructure outside the semiconductor sector.
46. What Current Evidence Does Not Establish
The evidence does not establish that:
- all future data-centre electricity growth will come from AI;
- every announced project will be completed;
- every region will experience a power shortage;
- AI development is universally limited by electricity rather than chips;
- all data centres have the same cooling or water requirements; or
- current electricity-demand projections will occur exactly as forecast.
These are stronger claims than the available evidence supports.
Forecasts remain sensitive to:
- AI adoption;
- hardware efficiency;
- server utilization;
- facility design;
- economic conditions; and
- energy-sector investment.
47. Research Implications
AI should increasingly be analysed as both a digital technology and a physical infrastructure system.
Model capability matters.
Chip supply matters.
But neither becomes productive in isolation.
The expansion of AI also depends on whether physical systems can deliver:
- electricity;
- power conversion;
- thermal management;
- network capacity;
- land;
- buildings; and
- reliable operation.
This changes how AI capacity should be interpreted.
The useful question is not only:
How many chips exist?
It is also:
How much compute can be continuously powered, cooled, connected, and operated?
That is a different measure.
It moves the analytical boundary from semiconductor production toward infrastructure coordination.
48. Research Limitations
Future AI infrastructure demand remains uncertain.
Data-centre forecasts combine different types of computing activity.
AI workloads evolve rapidly.
Efficiency improvements may reduce energy use per computation.
Higher utilization or larger workloads may increase total demand.
Announced project pipelines may contain duplication or projects that never reach operation.
Electrical constraints are geographically specific.
Cooling technologies differ substantially across facilities.
Energy markets and regulatory conditions vary by jurisdiction.
Current projections should therefore be treated as scenarios or estimates within their stated assumptions rather than precise predictions of future AI infrastructure.
The research value lies primarily in identifying the physical dependencies that must be satisfied for deployment to occur.
49. Research Conclusion
Artificial intelligence is not produced by software alone.
It is produced through a physical system.
Semiconductors provide computational capability.
Servers organize that capability.
Racks concentrate it.
Buildings contain it.
Electrical infrastructure powers it.
Cooling infrastructure removes its heat.
Networks move its data.
Generation systems supply its energy.
Construction and industrial supply chains make the entire configuration physically possible.
These dependencies can be summarized conceptually as:
computing hardware → facility integration → power and thermal support → connected operational compute
The relationship is descriptive rather than a mandatory development sequence.
Every major dependency matters.
A shortage of processors can constrain AI.
So can a transformer.
So can a delayed substation.
So can insufficient cooling.
So can a transmission project that requires years longer than the data centre it is intended to supply.
This leads to the central research principle:
Computing hardware represents potential AI capacity. It becomes operational AI capacity only when the surrounding physical infrastructure is available at the same time.
The most important future bottleneck may therefore not always be semiconductor scarcity.
It may be the coordination of infrastructure systems with fundamentally different development speeds.
The IEA’s current outlook illustrates this tension.
Data centres can be developed within a few years, while transmission infrastructure can require substantially longer periods and critical electrical equipment can remain subject to long delivery queues.
Current projects are also beginning to integrate energy access directly into data-centre development and location.
The planned 400 MW Chiba AI facility, for example, connects data-centre development with long-term power capacity and land provided near the infrastructure of a major electricity producer. Its phased development also demonstrates why announced capacity should remain distinct from operational capacity.
The broader implication is that AI competition increasingly extends beyond models and chips.
It also involves access to and coordination among:
- energy;
- electrical infrastructure;
- thermal infrastructure;
- land;
- networking;
- industrial equipment;
- construction;
- capital; and
- time.
In that sense, the next stage of AI scaling is not only a computing problem.
It is an infrastructure problem.
References
International Energy Agency. (2025). Energy and AI. IEA, Paris.
International Energy Agency. Energy and AI: Energy Demand from AI.
International Energy Agency. Energy and AI: AI and Energy Security.
International Energy Agency. Energy and AI: Executive Summary.
Lawrence Berkeley National Laboratory. (2026). United States Data Center Energy Usage Report: 2025 Update.
Reuters. (2026, October 1). JERA teams up with Dell, RHAELM on Japan’s AI infrastructure, building data centre near Tokyo.
Reuters. (2026, October 5). Nvidia, Broadcom shielded as AI power crunch hits chip supply chain, says Morgan Stanley.