The artificial intelligence industry has spent enormous amounts of capital concentrating computational power into increasingly large facilities. Hyperscale data centers, enormous GPU clusters, dedicated power generation, and expanding transmission infrastructure have become central to the ambitions of technology companies and the governments seeking to establish leadership in AI. The scale of these investments reflects a reasonable expectation that machine intelligence will become an increasingly important input into economic production.
Yet as we discussed in Analysis 09, The Watt Is the Constraint, the expansion of computational capacity eventually encounters the physical limitations of electricity generation and delivery. Capital can finance a data center, but the availability of electricity determines when that investment becomes productive. The location of that electricity, the infrastructure required to deliver it, and the economics of using it are becoming increasingly important considerations in the development of AI.
Once we begin treating machine intelligence as an economic resource, we have to consider where that intelligence should physically reside. The answer becomes particularly interesting when AI moves beyond generating information and begins participating directly in the operation of farms, factories, mines, utilities, transportation networks, and other parts of the physical economy. Intelligence has a geography, and the economics of that geography will help determine how AI infrastructure develops.
The Physical World Operates on Its Own Schedule
A modern agricultural operation may have soil-moisture sensors, irrigation equipment, refrigeration facilities, solar generation, batteries, cameras, and machinery collecting operational data throughout the day. Increasingly, those systems can be coordinated using machine learning, predictive analytics, and automated controls.
Imagine that a soil sensor detects declining moisture while temperature readings and weather information indicate conditions likely to place crops under stress. A local computing system can evaluate those readings, determine whether irrigation is appropriate, and instruct the equipment to respond. Meanwhile, another system might recognize that refrigeration temperatures are rising beyond an acceptable range, or that a water pump is exhibiting patterns associated with an impending mechanical failure.
If the farm loses its internet connection, the crops continue growing, temperatures continue changing, machinery continues operating, and refrigeration equipment continues consuming electricity. The physical processes governing the farm remain active regardless of whether a distant data center is reachable. An agricultural operation designed around continuous cloud connectivity may find itself unnecessarily vulnerable under those circumstances. Although some decisions can wait for connectivity to return, others require immediate action, particularly where equipment protection, crop preservation, energy management, and operational safety are concerned.
This is one of the clearest arguments for edge computing: placing computational capability near the equipment and processes that generate the data and require a response. The implications extend beyond agriculture because as intelligence becomes integrated into physical operations, the location of computation becomes part of the operational design.
From Centralized Compute to Distributed Intelligence
Cloud computing has been enormously successful because centralization offers significant economic and technical advantages. Large facilities can concentrate specialized hardware, networking, storage, engineering expertise, and supporting infrastructure. Computational resources can be allocated across customers and workloads, while software and services can be delivered without every organization maintaining its own extensive computing environment.
AI benefits substantially from those characteristics because training large foundation models requires concentrations of computational resources that would be impractical for most individual enterprises to reproduce. Large-scale inference, scientific modeling, and other computationally intensive applications also benefit from centralized infrastructure.
However, the advantages of centralization are not uniform across every workload: a camera inspecting produce on a processing line may generate continuous streams of visual data, while the useful output of its analysis might be a simple determination that a particular item should be rejected. An industrial sensor may produce thousands of measurements while the operational decision involves adjusting a valve or identifying an abnormal vibration. A refrigeration controller may need to evaluate several inputs and respond within seconds.
Sending every observation to a distant computing facility can introduce unnecessary bandwidth consumption, network dependencies, and recurring processing costs. For workloads with sufficiently predictable requirements, it may be more economical and operationally reliable to perform the necessary computation locally. This distinction becomes especially important as AI models become more efficient and capable of operating on smaller hardware. Specialized models, compressed neural networks, and dedicated inference accelerators are expanding the range of tasks that can be performed on devices consuming a fraction of the electricity required by large centralized AI installations.
The result is an increasingly practical separation between where intelligence is developed and where it is deployed. A sophisticated agricultural model might be trained using enormous datasets and substantial computational resources in a centralized facility. A smaller, specialized version could subsequently operate on a computer at an individual farm, evaluating local conditions without continuously communicating with the infrastructure responsible for its original development. The same relationship can exist between a major industrial AI platform and the equipment operating on a factory floor, because the infrastructure that produces an AI model does not necessarily need to be the infrastructure that executes it.
An Intelligent Hierarchy
I think the most useful way to understand this development is as an emerging hierarchy of intelligence. At one end are hyperscale facilities providing the computational concentration necessary for frontier-model training, large-scale inference, and workloads requiring extensive resources. Below that are regional and enterprise computing environments that can support specialized models, shared applications, data processing, and organizational coordination.
Closer to physical operations are local servers, industrial computers, and edge devices capable of interpreting sensor data, executing models, and supporting automated decisions. At the lowest levels, embedded controllers and specialized processors handle immediate equipment functions, sometimes with little need for sophisticated machine learning at all.
These layers have different economic and operational characteristics. A workload that requires substantial computational resources but has little sensitivity to response time may be well suited to a centralized facility. A task involving immediate equipment control, sensitive operational data, or unreliable network connectivity may be better placed locally. The objective is to allocate computation according to the requirements of the work being performed, taking into account electrical capacity, computational intensity, latency, bandwidth, reliability, data governance, equipment cost, maintenance requirements, and the consequences of a system becoming unavailable.
The distinction between intelligence and control remains important because industrial automation has relied on local controllers for decades to provide deterministic operation and safety. Machine learning can add predictive and adaptive capabilities without replacing the established control systems responsible for keeping equipment within safe operating limits. A well-designed hierarchy can therefore combine conventional automation, local AI inference, enterprise applications, and centralized models, with each performing the functions for which it is best suited. The future of AI infrastructure is likely to involve a distribution of computational responsibility across several layers, rather than a contest in which either cloud or edge must prevail.
Agriculture Offers an Important Development Lesson
Agriculture provides an especially useful example because it combines biological processes, mechanical equipment, environmental uncertainty, and substantial geographic dispersion. In many African agricultural markets, unreliable connectivity, limited electrical infrastructure, high bandwidth costs, and the expense of sophisticated imported equipment can complicate the adoption of conventional precision-agriculture systems. Those conditions create an opportunity to approach the architecture differently.
Rather than designing around the assumption that every farm will maintain continuous access to distant cloud services, systems can be developed around local capability, intermittent synchronization, and distributed energy resources. A farm equipped with modest computing hardware, appropriate sensors, and reliable local controls may be able to perform useful monitoring and automation without the expense of maintaining constant high-bandwidth connectivity. Solar generation and battery storage can support selected computing and control functions, while local data collection allows the system to maintain operational records during network outages. When connectivity becomes available, the farm can synchronize information, receive software or model updates, and access computational services that exceed its local capabilities.
The approach has practical limits: local hardware requires maintenance, equipment must be protected against environmental conditions, and even relatively small computing systems need dependable electricity. The economics of deployment also depend on farm size, crop value, equipment utilization, and the measurable benefits of automation. Nevertheless, this approach creates room for agricultural technology designed around actual operating conditions rather than assumptions imported from more extensively connected markets.
That distinction also creates a broader economic-development opportunity, since local computing infrastructure can support regional technical services, equipment integration, maintenance, agricultural data systems, and specialized software development. The economic activity surrounding intelligent agriculture can extend beyond purchasing imported sensors and subscribing to foreign cloud platforms.
The ability to retain and govern operational data locally also deserves consideration because farm-specific records concerning soil conditions, crop performance, equipment behavior, and energy consumption may become valuable inputs for specialized models and agricultural services. Ownership, contractual access, security, and the ability to transfer that data between systems will influence who ultimately captures the economic value. Agriculture therefore offers a useful demonstration of how distributed intelligence can be adapted to infrastructure constraints while creating opportunities for local technical capability.
The Same Architecture Extends Across the Physical Economy
The underlying principles apply across a much larger range of industries. A manufacturing facility may use local inference to identify production defects, detect abnormal equipment behavior, or adjust operational parameters. A mine may require computational systems capable of processing machinery telemetry and environmental data in locations where connectivity is unreliable. Logistics facilities can use local machine vision to coordinate inventory movement, inspect shipments, and monitor equipment.
Within utilities, electrical distribution networks, substations, energy-storage facilities, and distributed generation assets increasingly produce operational data that can support forecasting, diagnostics, and system optimization. Some of these functions benefit from regional coordination, while others must remain available at the equipment level. Hospitals and healthcare networks have their own requirements involving privacy, continuity of operations, medical-device integration, and the handling of sensitive information. Local inference can be useful where it improves responsiveness or reduces unnecessary movement of data, provided the systems meet the appropriate clinical, security, and regulatory requirements.
Municipal infrastructure offers particularly interesting possibilities because water-treatment facilities, pumping stations, traffic systems, public buildings, fleet operations, and other municipal assets generate information that could support more efficient operations. Local computing can help identify equipment anomalies, optimize energy consumption, and provide operational visibility without requiring every system to depend on continuous communication with an external service. For smaller municipalities, the opportunity may be especially significant because sophisticated operational capabilities can increasingly be deployed through relatively modest computing installations. The challenge becomes one of integration, procurement, maintenance, cybersecurity, and organizational competence rather than simply acquiring access to the largest available AI model.
Across these sectors, computation is closely tied to physical activity: when an AI system is participating in a process that consumes energy, moves equipment, preserves inventory, controls infrastructure, or influences operational safety, its architecture must account for the conditions under which that process actually operates.
Power Constraints Change the Economics of Placement
The electricity problem examined in Analysis 09 adds another consideration to these placement decisions. If large concentrations of computational capacity are increasingly constrained by the availability and cost of power, there is an economic incentive to examine whether every workload belongs in those concentrations.
Centralized computing can be exceptionally efficient because large facilities benefit from high hardware utilization, sophisticated cooling systems, specialized accelerators, and economies of scale that small installations may struggle to match. Moving computation to the edge does not automatically reduce total energy consumption or total cost. Local hardware can be underutilized, and deploying thousands of individual devices creates its own capital and maintenance burden. Some workloads are cheaper and more energy-efficient when consolidated in a well-operated data center.
Yet centralization also carries costs, as network transmission, repeated movement of data, communication delays, cloud-service charges, and the operational consequences of connectivity failures all contribute to the economics of a computing architecture. The relevant comparison therefore extends beyond the electricity consumed by an individual processor to encompass the total cost of performing the computation, delivering the result, maintaining availability, and supporting the physical operation that depends upon it.
A small edge computer operating continuously may be economically justified if it prevents expensive refrigeration losses, reduces machinery downtime, or eliminates substantial recurring data-transfer requirements. A different workload may be considerably cheaper when performed in a centralized facility serving thousands of customers.
This is where productive intelligence per watt becomes a useful concept, provided we evaluate the productivity of the entire system rather than treating processor efficiency as the only measure. An AI system that consumes very little electricity but produces no meaningful operational improvement is hardly an efficient economic investment. A more energy-intensive system that prevents substantial material losses or increases industrial throughput may create considerably greater value. The economics of machine intelligence ultimately depend upon what the computation enables.
Geography Will Influence Competitive Advantage
As this hierarchy develops, the geography of AI infrastructure may become more varied than the current concentration of hyperscale investment suggests. Regions with abundant electricity, strong transmission infrastructure, and favorable industrial conditions will continue attracting large computing facilities. Those advantages are significant, particularly for workloads requiring massive computational concentration.
Other regions may develop advantages through the deployment of intelligence closer to productive activity. Agricultural regions can support specialized computing and data infrastructure serving farms and processing facilities. Manufacturing centers can integrate local inference into industrial equipment and production systems. Mining regions, ports, transportation corridors, and energy-producing areas may develop their own concentrations of specialized operational compute. These installations need not resemble conventional data centers, since many will consist of relatively small computing systems integrated directly into existing facilities and machinery.
That distinction has implications for economic development because communities seeking to participate in the AI economy may have opportunities beyond competing for hyperscale campuses. The integration of machine intelligence into existing productive sectors can create demand for technical services, software, infrastructure modernization, equipment maintenance, and specialized workforce capabilities.
Evaluating AI investment therefore requires attention to both the presence of large computing facilities and the ability of a regional economy to apply computation productively across its farms, factories, utilities, transportation systems, and public institutions. Both can contribute to economic competitiveness, but they represent different forms of capability: the first concentrates computational resources, while the second distributes their productive application.
From the Geography of Power to the Geography of Intelligence
Throughout this series, we have examined the relationship between electrical capacity, computing infrastructure, and the economic potential of machine intelligence. Electricity determines the physical resources available to support computation, while hardware, software, and infrastructure determine how efficiently those resources can be converted into computational capability; models and applications then determine how that capability becomes useful intelligence.
The next consideration is where those capabilities should operate. Some intelligence will continue to be produced and delivered from enormous centralized facilities, while other intelligence will reside in regional infrastructure, enterprise systems, local servers, and individual machines. Much of the physical economy will depend on combinations of these arrangements, with workloads distributed according to their technical requirements and economic value.
The geography of electricity will influence where large concentrations of computation can be built, while the geography of economic activity will influence where intelligence needs to be deployed. Those two geographies will overlap in some places and diverge considerably in others, and infrastructure planning will have to account for both as AI becomes more deeply integrated into production.
We have spent much of the current AI expansion asking how much computation we can finance, build, and power. Those questions remain essential, particularly as the industry encounters the electrical constraints discussed in the preceding analysis. Alongside them, we must now ask where intelligence should physically live, because the answer will help determine how the next generation of intelligent infrastructure is built, who can participate in its development, and how broadly the economic benefits of machine intelligence can be distributed.