VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Super Intelligence | 0002 — Why Power Becomes the Limiting Resource

Three students work together at a classroom table, reviewing open textbooks and notes.

Super Intelligence master guide › Energy, work, learning and safety series › Article 0002

Power can become a limiting resource for AI when electricity supply, grid connections or a computing facility’s electrical systems cannot support the workload being planned. A more capable model and a larger collection of processors do not remove that constraint. The service must receive enough usable power at the right location, under conditions that allow it to operate reliably. That is why energy planning belongs inside the discussion of advanced computing.

The issue concerns current AI infrastructure as well as any hypothetical future superintelligence. These are different capability claims, but both would depend on physical systems. Understanding the electricity connection helps a reader distinguish an ambitious announcement from an operating service, a national total from local availability, and a maximum capacity figure from the energy actually delivered over time.

For students and parents, this is also a practical lesson in systems thinking. A chain performs according to the connections between its parts. Adding strength to one part may produce little benefit if another part remains the constraint. Electricity provides a concrete way to study that principle because its quantities, timing and delivery requirements can be examined separately.

Power, energy and capacity answer different questions

Power describes a rate of using or supplying energy. Energy describes the amount supplied or used over a period. A kilowatt measures power, while a kilowatt-hour measures energy. The US Energy Information Administration explains this distinction with the relationship between a device’s power and its operating time. Large computing announcements often use megawatts or gigawatts, while consumption over a month or year requires an energy unit. EIA: measuring electricity

A hypothetical device drawing two kilowatts continuously for three hours uses six kilowatt-hours. That calculation is straightforward because the power is assumed constant. If its power varies, the energy used depends on the pattern over time. One cannot infer the exact energy total from the largest momentary reading alone. This is why a peak figure and a consumption figure are not interchangeable.

Capacity adds another question: how much power could a system supply under specified conditions? It does not establish what the system actually supplies throughout the day. EIA distinguishes generation from capacity and notes that generating units generally do not operate at full capacity every hour. That distinction is essential when assessing whether a new source can support a sustained computing load. EIA: generation and capacity

Before accepting a large number, therefore, ask what it describes. Is it planned demand, equipment rating, approved connection capacity, maximum generation or average consumption? The answer changes its meaning. A sound comparison uses matching units, matching periods and clearly stated boundaries. These questions often resolve confusion before any sophisticated forecasting is needed.

Electricity must be generated and delivered

Generation and delivery are distinct requirements. A region might have sources capable of producing substantial electricity while a particular site lacks a connection that can carry its requested load. Conversely, a site might have physical connection capacity while the wider system faces supply constraints at important times. Both conditions matter to an operating service.

The grid connects producers and consumers through power lines, substations and transformers. EIA’s explanation of electricity delivery describes how these components move electricity and change voltage along the route. The relevant connection is therefore a system of equipment, rather than an abstract right to use a national supply. EIA: delivery to consumers

For a computing project, that system may need examination before the building is ready. The required electrical route must fit the proposed load and the conditions under which it operates. This introduces dependencies on planning, equipment, construction and coordination. A finished building cannot make an incomplete connection available simply because processors have arrived inside it.

The implication is a useful distinction between location and readiness. A site can be attractive for land, access or other reasons and still require substantial work before it can receive the intended power. Evaluating readiness means examining the actual route and commitments, not merely the presence of electricity infrastructure somewhere nearby.

A local bottleneck can matter more than a national total

National electricity totals are useful for understanding scale, but they do not describe every site’s circumstances. Computing facilities concentrate demand in particular places. The International Energy Agency identifies that geographic concentration as one reason integrating data centres into grids can be challenging. A modest share of a large national system can still be a substantial addition to a local network. IEA: energy demand from AI

Consider a hypothetical country with many regions. Some have available generation but limited transmission to the location where a facility is proposed. Others have strong connections but substantial existing demand. Comparing only the national annual total conceals those differences. A project needs a local assessment of supply, delivery and its operating pattern.

This helps explain why simple international comparisons can mislead. One economy’s larger electricity generation does not by itself prove that it can provide more useful AI services in every setting. Hardware supply, efficiency, network arrangements, workforce capability and adoption also matter. Electricity is a strategic input, but it is not a complete measure of computing success.

A more informative comparison asks how reliably each system can support its intended workloads and what expansion is realistically available. This turns competition from a single-number contest into a question about coordinated capability. It also creates room for different strategies: some places may expand supply, others may improve efficiency or select workloads that fit the resources available.

The computing workload determines the power requirement

An AI facility does not have one unchanging demand simply because it contains AI hardware. Its operating pattern depends on the work performed, equipment used and service commitments. Developing models, running scheduled jobs and responding to interactive requests create different planning questions. The facility needs to understand its workload before making credible statements about power needs.

A hypothetical overnight processing job may tolerate a delay, while a live assistance service may need responses throughout the day. That difference could affect when computing is scheduled and how much capacity must be available at short notice. Flexibility should be evaluated for the actual application. It cannot be assumed from the general fact that work is digital.

The workload also includes supporting systems. Power used by processors is not necessarily the total power used by the facility. Cooling, networking and electrical equipment can contribute to the full requirement. A proposal should therefore state whether a number concerns computing equipment alone or the whole site. Confusing those boundaries can make a supply plan appear more complete than it is.

This is another instance of the AI computing stack operating as a connected system. The application defines an experience, the software schedules work, the hardware performs it and the physical installation supports it. The power requirement emerges from the arrangement, rather than from a model’s name.

Power requirements also depend on matching processors and memory to the workload; equipment that stalls on another bottleneck can consume resources without producing equivalent useful work.

Reliability concerns what happens under difficult conditions

A system that works on an ordinary day may still face problems during equipment failure, unusual demand or other disturbances. Reliability planning therefore asks how the supply and load behave under defined conditions. It is more demanding than checking whether an average quantity appears sufficient.

The North American Electric Reliability Corporation’s work on emerging large loads identifies challenges associated with uncertain load forecasts and resource adequacy. Projects may change timing, size or location, while electricity infrastructure planning must still prepare for plausible demand. This makes the quality of project information important to the wider system. NERC: characteristics and risks of emerging large loads

A useful hypothetical test asks what happens if one part of the planned supply is unavailable. Can work be delayed? Can the facility reduce demand without losing essential functions? Is alternative supply genuinely available under the same conditions? The answers determine what a reliability claim means. Merely listing backup equipment does not explain whether the operating plan is adequate.

Users encounter reliability through availability and continuity. A service may recover from a disruption, but a task could still be interrupted or a result delayed. The appropriate standard depends on the consequences. A casual creative exercise and an important operational process may require different arrangements. Good planning connects the technical response to the actual user need.

The same distinction between capacity and dependable service helps when comparing orbital and terrestrial electricity budgets; generation alone does not establish continuous usable computing.

Worked example: capacity is not continuous output

Imagine a hypothetical facility planning to use an average of ten megawatts throughout a day. Its daily energy requirement, under that simplified constant-load assumption, is 240 megawatt-hours. Now suppose a proposed generation source has a maximum rating of ten megawatts but produces that maximum for only part of the day. The two ten-megawatt labels do not establish a complete supply match.

To see the issue, assume for illustration that the source produces ten megawatts for six hours and nothing during the remaining eighteen. It provides sixty megawatt-hours over the day. The facility’s assumed requirement remains 240. These numbers are invented for the example; the lesson is arithmetic. Equal maximum power ratings can accompany very different energy totals and operating patterns.

One response could involve additional sources. Another could involve storage, changes to the load or a combination. Each option would need its own analysis of capacity, energy, timing and losses. The example does not prove that one technology is unsuitable. It demonstrates why matching supply and demand requires a time-based description.

Now change the workload. Suppose three quarters of the computing can be delayed and performed during selected hours, while the remaining quarter is a continuous service. That creates a different planning problem, but it does not make the continuous requirement disappear. The facility must identify which jobs are flexible, how far they can move and what obligations remain.

A sound comparison would therefore show the demand pattern and the proposed supply pattern together. It would also test the conditions under which the match fails. This is more informative than comparing two nameplate ratings or multiplying a maximum by every hour of the year without examining whether that assumption is justified.

Worked example: a building waits for its connection

Consider a hypothetical computing project with three milestones: building completion, equipment installation and electrical connection readiness. The building finishes first and the processors arrive shortly afterwards, but an essential connection component is delayed. The facility is physically present yet cannot operate at its intended load.

This example illustrates a coordination risk. If publicity focuses on the building’s progress, observers may assume that capacity is nearly available. If the project record shows the dependencies clearly, the remaining issue becomes easier to understand. The relevant milestone is the readiness of the complete operating system, not merely the most visible construction activity.

The project team could respond by phasing the installation, changing the start date or operating a smaller initial workload if the available connection and approvals permit it. Each option changes costs and service commitments. None should be represented as equivalent to the original plan unless it actually provides the same result.

An evaluator should ask whether the connection work has an identified scope, schedule and responsible parties. Which elements are secured, which remain conditional and what is the consequence of delay? These questions are useful without requiring a general reader to understand the engineering design. They examine whether the claim of readiness has a concrete foundation.

For learners, the analogy extends beyond data centres. A school event can have a venue and programme but still lack transport. A manufacturing line can have machines but lack a necessary input. The visible assets matter, yet performance depends on the unfinished connection. Recognising that pattern is the broader value of studying power bottlenecks.

Worked example: scheduling can create flexibility

Imagine a service offering two kinds of work. One is live question answering, where users expect a prompt response. The other is generating draft summaries that staff will review the following morning. Treating both as equally urgent would ignore a potentially useful difference in their deadlines.

The organisation could examine whether the scheduled summaries can run at selected times without affecting their usefulness. It would need to consider the amount of work, deadlines, available computing and any additional cost of moving jobs. A workload is flexible only when its timing can change while its required outcome is preserved.

The live service may offer less flexibility. Reducing its capacity during a busy period could increase waiting time or cause failed requests. If the service promises a particular response standard, that commitment constrains the operating choices. Flexibility cannot be created simply by describing an application as optional after users have come to depend on it.

A realistic trial would move a defined portion of the scheduled work and measure completion, energy use and operational difficulty. If deadlines are missed or review becomes harder, the change may not be worthwhile. If the same acceptable work is delivered under better operating conditions, the trial would provide evidence of useful flexibility.

The broader principle is to match the timing of computing with the timing of the service’s value. This can contribute to energy planning, but it is an application decision as well as an infrastructure decision. It requires cooperation between the people who understand the tasks and the people who manage the systems.

Efficiency and supply expansion are complementary

A power constraint can be addressed by changing how much energy is available, how efficiently work is performed or how the work is organised. These approaches need not compete. A facility may benefit from better algorithms while also requiring a stronger connection. The appropriate combination depends on the scale and nature of the constraint.

Efficiency should be measured against an acceptable result. Reducing the length of an answer might save resources, but it is not useful if it removes information essential to the task. Using a smaller system might be appropriate for routine classification, while a harder task requires something else. The aim is to reduce unnecessary effort while preserving the required quality.

Expansion should likewise be justified through an operating plan. Additional power can create opportunities, but those opportunities need hardware, applications and demand. A supply announcement does not establish the amount of useful work ultimately produced. The companion article on the economics of intelligence per watt examines that connection between physical inputs and economic results.

A useful comparison includes several scenarios. What happens if demand grows slowly? What if efficiency improves faster than expected? What if a connection is delayed? Scenarios do not predict the future exactly. They reveal whether a plan depends on one optimistic combination of assumptions, and where flexibility might protect it against change.

Policy matters through concrete mechanisms

Supportive energy policy can influence the conditions under which infrastructure is planned and built, but a broad political label cannot establish the effect of a specific policy. A durable assessment asks which mechanism changes: approval processes, investment conditions, workforce development, grid planning or another identifiable part of the system.

Suppose a hypothetical policy is intended to shorten the review of connection projects. Its success would depend on whether the change actually reduces unnecessary delay while preserving the relevant technical and public requirements. If the main constraint is equipment availability, changing the review process alone may have little effect. The diagnosis should come before the claim of policy success.

Similarly, national computing strategy needs more than energy. Chip production, memory supply, algorithms, skills and adoption affect the outcome. Abundant electricity can be valuable without being sufficient. A policy comparison should therefore connect inputs to operating capacity and then to useful services, rather than treating each announcement as proof of the final result.

This approach keeps the discussion evidence-based and useful across changing administrations or institutions. Ask what changed, where it changed and what results followed. A policy deserves credit to the extent that the mechanism and outcome can be demonstrated. Its name or stated intention is not a substitute for that examination.

Power planning is also a community question

Large electricity demand enters an existing local system. Other users already rely on that system, and new infrastructure may require land, construction and public decisions. A credible computing plan should therefore explain how it fits the place where it will operate, including who bears costs and who receives benefits.

The details vary by location, so broad promises cannot resolve them. A new connection might include improvements useful to other users, or it might primarily serve the facility. A project might support local employment while also creating pressure on other services. These possibilities should be assessed through the actual proposal and operating conditions.

Community participation can improve the information available to planners. Residents and local institutions may identify concerns that a computing forecast does not capture. Those concerns can then be examined, answered or addressed through design changes. Engagement is most useful when it occurs early enough to influence decisions, rather than after every significant choice has been fixed.

The connection here is direct: power delivery is both an engineering arrangement and a local development decision. Its success includes reliable operation and a clear account of how the project interacts with its surroundings.

The practical consequence is who carries grid costs and local burdens; the distribution of connection costs and service pressures belongs alongside the capacity plan.

Storage has a power limit and an energy limit

Storage is easier to understand when two capacities are separated. One describes how much power the system can supply at a moment. The other describes how much usable energy is available over time. A storage proposal can have a substantial energy total while being unable to meet a higher instantaneous demand, or a high power capability while supporting that demand only briefly.

For a simplified hypothetical example, suppose a storage system can deliver at most five megawatts and contains twenty megawatt-hours of usable energy. At a constant five-megawatt output, that energy corresponds to four hours before accounting for any further conditions. The arithmetic does not establish that the system can supply ten megawatts for two hours, because the assumed power limit is five megawatts.

Now imagine a facility needing two megawatts for a support function during a temporary interruption. The same illustrative energy total corresponds to ten hours at that constant output. That might be relevant to the task, but the facility would still need to assess whether the equipment and operating arrangement permit the required service. Matching the arithmetic is necessary for the example; it is not a full design approval.

Storage also needs a charging plan. It shifts energy from one time to another rather than creating a continuing source of energy by itself. If the proposed supply cannot provide enough energy to charge it while also serving the load, a large storage installation does not resolve the missing energy. The complete system must account for supply, consumption, charging and losses.

This distinction helps a nontechnical reader ask useful questions. What power can be delivered? For how long under the relevant load? When and how is the system recharged? What happens if the interruption lasts longer? An answer that gives only one impressive capacity figure leaves part of the operating problem unexplained.

Different parts of a project run on different clocks

A computing project combines activities that can have different schedules: preparing a site, constructing a building, obtaining equipment, making a connection and establishing the operating service. A plausible schedule for one activity does not establish the schedule for all the others. The earliest completed part may wait for the latest essential dependency.

Consider an invented project with a building scheduled for one year and a necessary electrical component scheduled for eighteen months. If both begin together and no alternative arrangement is available, the building’s completion does not determine the operating date. The project needs to recognise the dependency in its plan rather than treating the six-month difference as an unexpected surprise.

The same reasoning applies to workforce and software readiness. A facility can have power but lack the trained team or application needed to use it as intended. This does not make power planning less important. It means that useful capacity is an intersection of requirements. The project becomes operational only when the necessary conditions coincide.

Good planning makes those clocks visible. A milestone should identify what it enables and what it still depends on. A connection agreement may enable a construction step without guaranteeing the final service date. A hardware purchase may secure equipment without securing the infrastructure required to operate it. Clear language protects readers from inferring more than a milestone actually establishes.

Phasing can sometimes help. An initial service might use a smaller completed portion while later expansion continues. That can create value sooner if the operating arrangement genuinely supports it. But the phase should be evaluated on its own requirements and results. It should not be counted as full completion of a larger plan merely because some useful work has begun.

A power plan depends on the commitments behind a deliverable supply; a demand forecast must become an agreement that can support construction and operation.

How to compare two infrastructure plans fairly

Suppose two hypothetical regions offer to host a computing facility. One advertises cheaper land and a larger national electricity supply. The other offers a clearer local connection route and a smaller but better-defined initial operating plan. Choosing between them requires more than ranking their most attractive numbers.

First compare the same service goal. If one proposal supports a different workload or availability standard, its lower cost may not represent a cheaper way to obtain the same result. Define the required output and conditions before comparing inputs. This keeps the assessment tied to purpose rather than to the easiest quantity to advertise.

Next compare uncertainty. Which costs are estimates, which schedules are commitments and which dependencies have not been resolved? A plan with a slightly higher stated cost could still be more workable if its conditions are clearer. Conversely, greater uncertainty could be acceptable for an exploratory project if the organisation can tolerate delay and change. The appropriate choice depends on the consequences.

Then compare the local relationship. How will connection and infrastructure costs be allocated? What benefit does the surrounding community receive? What information will be available once the facility operates? These questions connect the plan to the place that supports it. A national strategic argument does not answer them automatically.

Finally test several plausible futures using the same assumptions for both plans. Try lower demand, delayed equipment and a different efficiency trajectory. The exercise does not identify a universally winning location. It identifies which plan remains useful under the conditions the decision-maker considers important. That is a stronger basis for comparison than a contest between isolated maximum-capacity figures.

How to evaluate an AI power claim

Begin by identifying the quantity and boundary. Is the claim about a single facility, several sites or an entire economy? Is it planned capacity or measured use? Does it include supporting infrastructure? Does it refer to a moment, an average or a period? Many misleading comparisons arise from leaving one of these questions unanswered.

Next examine timing and evidence. An ambition, an approved connection and an operating supply are different stages. A credible account identifies which stage has been reached and which dependencies remain. It should also avoid counting the same potential project several times simply because several locations are under consideration.

Then connect the power to work. What workload will it support, what service will that workload provide and how will usefulness be evaluated? If the argument moves straight from electricity to large economic benefits, ask what assumptions link the two. Those assumptions may be plausible, but they still need to be visible.

Finally examine resilience and alternatives. What happens if demand changes, a component is delayed or part of the supply is unavailable? Can the plan adapt without losing its purpose? This is the standard of a useful assessment: enough detail to understand the claim, enough evidence to locate its current status and enough uncertainty to examine how it could fail.

These checks connect physical capacity with the wider questions of usefulness and public value explored in the Super Intelligence series hub. Electricity is an enabling input, while the ultimate assessment concerns the services people receive. Keeping those two levels connected helps readers take energy constraints seriously without treating every increase in power as proof of an equally large increase in useful intelligence.

Frequently asked questions about AI electricity demand

Why can electricity limit AI when software is improving?

Software improvements still need machines on which to run. If the available supply or connection cannot support the planned workload, additional hardware may be underused or unavailable. Better algorithms can help by reducing resources required for acceptable work, but they do not automatically solve every physical constraint. The useful question is which limit currently matters and which intervention would address it.

Does a gigawatt announcement mean that power is already available?

No. The number might describe a target, an equipment rating or an eventual connection. Its meaning depends on status and boundaries. Ask whether the relevant generation and delivery systems are operating, whether the figure describes continuous output and what conditions apply. A large number without those details establishes scale of ambition more readily than scale of completed service.

Can renewable generation support computing facilities?

It can contribute, but the complete arrangement must match the load’s timing and reliability requirements. Generation patterns, connections, other supply and storage may all matter. The right assessment examines a defined location and operating plan rather than treating a generation technology as automatically sufficient or insufficient. The arithmetic of capacity and energy is a starting point; the time-based match is the next step.

Will efficiency improvements reduce total electricity demand?

They can reduce energy required for a specified amount of acceptable work. Total demand also depends on how much work is requested. If lower cost or greater capability encourages more use, overall consumption could still grow. To evaluate a claim, separate efficiency per task from the total number and complexity of tasks. Both quantities belong in the account.

What should students learn from this topic?

Learn to distinguish power, energy and capacity, then practise tracing dependencies. Use a hypothetical supply-and-demand example to see why equal maximum ratings do not guarantee a match. Extend the reasoning to other systems: identify inputs, connections, timing and failure conditions. This makes electricity planning an accessible exercise in mathematical reasoning and in evaluating claims about complex technology.

Previous: 0001 — Computing Reinvented from the Ground Up · Next: 0003 — Chips, Memory and the Geography of Computing