How Superintelligence Works | Energy, Water and Environmental Load — What AI Resource Numbers Mean

1. The environmental accounting boundary comes first · 2. Power and energy are not synonyms · 3. Define the three important scopes · 4. The fictional Cypress annual packet · 5. Calculate PUE accurately · 6. Calculate WUE with the right denominator · 7. Allocate annual facility energy to tasks · 8. Allocate annual direct site water

9. A second configuration changes two resources · 10. Report absolute changes alongside ratios · 11. Direct site water is not the whole water story · 12. Local water stress matters · 13. Cooling systems move heat somewhere · 14. Carbon estimates require an emissions factor · 15. Time of use changes the inference · 16. Hardware manufacturing is outside operational PUE

17. Efficiency per task versus absolute consumption · 18. Forecasts are scenarios, not future observations · 19. Independent transfer: the museum compute room · 20. Full transfer answers · 21. A repeatable environmental-reporting checklist · 22. Sources and next learning route

Beyond the ratios: measurement, trade-offs and trustworthy decisions

Meter readings and calculated estimates have different evidential status

A facility operator can measure electricity at defined meters, but attributing shared electricity to one model or task requires a rule. When the total is known but the split is not, an analyst should describe the allocation method instead of claiming direct per-task metering. A modelled number and an observed number can appear equally precise in a table yet support different conclusions.

For Cypress, the annual electricity figure is stipulated as a fictional measured-looking input and the per-task number is an arithmetic allocation. No one observed the resource consumption of any individual prompt. The distinction should remain visible whenever a calculator turns aggregate data into a per-user estimate.

A denominator of successful work matters

If a facility completed one million attempted jobs but only 900,000 met the user’s accuracy or verification requirement, dividing total electricity by one million would describe energy per attempt rather than energy per acceptable completion. Those are distinct performance measures.

Assuming the same fictional 60,000 kWh facility energy, the average across 900,000 accepted tasks is 60,000/900,000 ≈ 0.0667 kWh, or 66.7 Wh, higher than the earlier 60-Wh per-attempt figure. This is an illustrative change of denominator, not evidence that failures consumed a fixed share of energy.

An idle server still uses resources

A powered service may consume electricity when no student is actively using it. Capacity kept ready for peak demand can support reliability while increasing the facility’s base load. A per-task average depends partly on how much useful work shares those background resources.

If task volume falls but standby consumption remains, the allocated resource use per completed job may rise even without changes to the model. A claimed saving from shorter output alone can therefore miss the larger operating conditions. Measure idle and load-sensitive components where practical.

Different tasks do not have equal computation profiles

A short text classification, a long-context science explanation, a video-generation request and a multi-step agent search can have very different compute and memory needs. A facility with several task types cannot assume one average prompt is a physically meaningful unit without defining the mix.

A credible workload analysis groups tasks by relevant characteristics such as input and output length, media, tools and completion criteria. Weighting by observed use is better than treating every user message as identical. The task mix also changes over time, so published numbers need a review date.

The role of hourly electricity supply

Annual facility kWh is an essential quantity, but the electricity mix and the grid’s operating conditions vary with time and location. Electricity demanded during a capacity-constrained hour may have different implications from the same energy used when supply is abundant.

An annual emissions factor is useful for particular accounting purposes but should not be described as a measured causal impact for every request. More detailed grid-effect questions require appropriate time-resolved power and generation information. Do not substitute a global average for a local analysis.

Cooling and water trade-offs are physical, not rhetorical

Evaporative cooling can reject heat through the evaporation of water; other cooling designs can use different equipment and electricity. A lower PUE under certain conditions may be accompanied by higher operational site-water use. Conversely, water-saving cooling may increase some electrical loads.

The Cypress comparison illustrates the possibility without claiming one named cooling technology always causes that result. Actual systems may combine approaches, operate under changing weather and rely on different water qualities or treatment. Engineering decisions require local measurements.

Withdrawal, consumption and return

A water intake meter records water entering a boundary. Some may be returned, discharged or reused under defined conditions. Water consumption, water withdrawal and water quality are different dimensions. A simple statement that a facility ‘uses’ litres may obscure which definition the source follows.

The WUE calculation should use the appropriate operational site-water definition and a compatible IT-energy denominator. Analysts should also identify source, discharge and reuse arrangements where material. A litre withdrawn and returned is not always equivalent in ecological consequence to a litre evaporated.

Water quality and local infrastructure

Cooling systems may require water treatment and maintenance. Different water sources can have different qualities, costs and local environmental implications. Reuse may help under suitable designs but brings its own infrastructure and operational requirements.

An article should not infer safe or unsafe local water practices from a global average. Responsible site assessment needs documentation of intake sources, applicable standards, discharge conditions and local stakeholders. Those are engineering and governance questions beyond the Cypress arithmetic.

Source water sits outside the narrow WUE boundary

Water consumed to generate the electricity supplying a facility can occur far from the server room. A site may report low direct water consumption while drawing electricity from a system with upstream water use. A full comparison needs energy-supply information and a consistent accounting method.

Do not add a hypothetical upstream litres figure to a measured site WUE without identifying its source and period. A combined life-cycle or source-water indicator may be valuable, but it is not the same as the conventional direct facility WUE ratio.

Embodied equipment impacts depend on lifetime

Manufacturing accelerators, servers, networking equipment and cooling infrastructure requires materials and energy. Operational gains from a more efficient system may be relevant, but frequently replacing equipment can add manufacturing impacts.

A lifecycle comparison must define the studied equipment, useful service life, reuse or recycling assumptions and system boundary. The article cannot certify whole-life superiority from annual PUE, WUE or generation speed alone. These metrics are inputs to a broader analysis, not substitutes for it.

Reliability reserves are part of a resource decision

Facilities may keep redundant equipment, backup systems or spare capacity to reduce outage risk. Those resources can increase operating and embodied demands, but an interrupted service can also waste computation and prevent a user from receiving any useful result.

Resource efficiency therefore needs a service-level definition: energy per completed checked output, relevant recovery needs and legitimate safety requirements. It would be poor policy to remove all redundancy simply to improve a displayed intensity number.

Absolute demand grows with new uses

Efficiency improvements may reduce the resources needed for a comparable task, while more users and new, more intensive applications increase total demand. The two trends can occur together. A provider cannot promise total resource reductions from a per-task improvement alone.

Compare absolute annual totals, workload volume and task mix. The earlier fictional threefold-volume calculation is intentionally separate from any empirical claim about why users adopt AI. A causal explanation for increased demand requires evidence about behaviour and markets.

A local grid connection is not guaranteed by a global forecast

Data centres are connected to particular power systems and may face local connection capacity, permitting, tariff and equipment constraints. An international forecast for total demand cannot establish whether a specific proposed facility has secured a connection or will operate at its planned load.

The IEA’s October 2026 Powering Artificial Intelligence report discusses grid integration, affordability and sustainability. It should be interpreted as sector analysis rather than a certification of an individual development.

Energy and emissions are connected through assumptions

The Cypress example’s hypothetical constant emissions factor is a teaching device. Real emissions reporting may use location-based grid factors, market-based procurement information or other prescribed methods depending on its purpose. These methods need clear documentation.

Do not compare two emissions numbers computed under different boundaries as if they were directly equivalent. Likewise, purchasing a certificate is not the same observation as proving the exact generator whose electricity powered a given model request.

Seasonal reporting can reveal hidden contrasts

An annual intensity ratio may smooth seasonal peaks in cooling demand or water use. Two sites can have similar annual WUE while one faces significant stress during a hot dry period. The appropriate analysis should inspect when and where resource consumption occurs.

More granular measurements can improve understanding, but they should retain the definitions of the annual KPI and avoid relabelling a single day as a full annual metric. Report both the annual result and relevant seasonal observations when available.

Task correctness belongs alongside efficiency

A model that returns an incorrect answer rapidly may look efficient per generated token while delivering little value to learners. A longer verified explanation may require more computation and still be appropriate for its educational purpose.

For an honest comparison, define the quality threshold and report resource use per task that reaches it, where measurement permits. Avoid ranking a short guess against a complete checked response as though both were identical products. The purpose of the tool matters.

Careful use of environmental claims in education

Students may see dramatic single-number claims about AI’s water or electricity use. A useful classroom exercise is to ask what was counted, over which period, and whether the claim describes a modelled average, an actual meter reading, a marginal effect or a future projection.

An answer that lacks these fields should be treated as incomplete rather than necessarily true or false. Teaching scientific source criticism is more durable than memorising one context-free figure that may change with hardware and technology.

A variation with half the number of tasks

Return to Cypress’s original 60,000-kWh annual facility consumption but assume just 500,000 equivalent checked tasks are delivered, with energy otherwise unchanged. Average allocated electricity rises to 60,000/500,000 = 0.12 kWh, or 120 Wh per task.

If site water stays at 24,000 litres, the corresponding allocation is 0.048 litre or 48 millilitres. No change in PUE or WUE follows automatically because the IT and facility annual readings were held fixed. This test shows how task volume and facility ratios describe different dimensions.

A variation with a changed IT load

Keep the original facility total at 60,000 kWh but let IT electricity rise to 48,000 kWh in a new fictional case. PUE would become 60,000/48,000 = 1.25. If operational site water remained 24,000 litres, WUE would be 24,000/48,000 = 0.5 L/kWh.

Those ratio improvements occur without any stated reduction in absolute facility electricity or direct site water. A dashboard displaying only lower PUE and WUE could therefore exaggerate the environmental outcome. The absolute quantities are essential.

A learner’s final evidence checklist

Record the task and quality criterion, number of completed jobs, operating period, facility total and IT electricity, site water, locality and relevant source assumptions. Compute ratios with consistent units and compare actual totals beside intensities.

If an upstream or lifecycle category is unmeasured, identify that gap. If a claimed environmental saving depends on a modelled scenario, state its assumptions. If two designs trade electricity against local water stress, describe both effects rather than inventing a universal winner.

Applied environmental accounting: matched boundaries, trade-offs and independently solved cases

Matched boundaries before ratios

An annual environmental indicator is meaningful only when its numerator and denominator cover the same facility and time. If IT energy is measured over one year but water is counted over a shorter period, dividing them can create a plausible number that describes no coherent operating interval. A change in meter scope can also produce an apparent improvement without any physical change.

For the Cypress packet, the IT energy, total facility energy and direct site water all refer to the same fictional year. That is why its PUE and WUE calculations can be performed directly. Before applying the method to a real facility, a reader would need metering records, timestamps, boundaries and appropriate checks for missing observations.

What a counter reading can and cannot tell you

A cumulative electricity meter reports energy over its observation interval. Two readings allow the operator to calculate a difference, but that total does not identify which piece of hardware used it unless separate submeters provide the division. A water meter can similarly report site volume without revealing the water’s purpose or ultimate fate.

An analyst should not assign all unexplained electrical load to cooling, or all water volume to evaporation, without the necessary component observations. Distinguish source measurements from later calculated allocations. The conversion formula may be exact while the attribution remains unknown.

A ten-hour energy example

Suppose an invented facility holds a constant 120-kW total draw for ten hours. It uses 120×10 = 1,200 kWh over that span. If its IT draw is 80 kW over the same period, IT energy is 800 kWh and the implied overhead energy is 400 kWh. The interval-specific PUE is 1,200/800 = 1.5 under these simplified matched boundaries.

If the operator also reports 300 litres of site water in that ten-hour interval, 300/800 = 0.375 L per kWh of IT energy is an interval-specific ratio. It should not automatically be called annual WUE, whose standard site-based definition uses annual site water and annual IT energy. Temporal labels prevent a short demonstration from masquerading as a yearlong measurement.

Demand peaks and energy totals answer different questions

Two facilities can consume the same daily energy while imposing different maximum instantaneous electrical demand. A steady 100-kW load for twenty-four hours uses 2,400 kWh. A 200-kW load for twelve hours followed by twelve hours with no load also uses 2,400 kWh, but its peak power is twice as large.

Grid operators, facility engineers and environmental analysts may ask different questions. Daily kilowatt-hours, peak kilowatts and rapid power changes all matter under different conditions. A single site-energy figure cannot prove how the surrounding electricity network responds during the busiest hour.

Local water withdrawal versus water consumption

Water withdrawal refers to water taken from a source, while consumption typically concerns water not returned to the same usable source and timeframe. Discharge may have quantity, temperature or quality characteristics. Their precise definitions vary by accounting method, so the report must say what the meter and metric represent.

A cooling system can circulate the same water many times before replacement or evaporation. Counting every internal circulation as fresh withdrawal would create a misleading total. Equally, claiming all withdrawn water is permanently consumed may be false. Source-specific measurements and a clear water balance are needed before drawing environmental conclusions.

Why site WUE is not full water footprint

The US Department of Energy explains site-based water usage effectiveness as annual on-site water use in litres divided by annual IT energy in kilowatt-hours. This ratio is valuable for tracking a site’s defined direct water demand. It does not by itself include all water associated with electricity generation or semiconductor manufacturing.

Comparing two water figures requires checking whether both use site-based or broader source-based definitions. The numerator may differ substantially when upstream systems are included. A lower on-site water number can coexist with unknown or higher upstream effects; neither outcome follows without the appropriate additional records.

Heat and cooling must be discussed physically

Computing equipment consumes electricity and releases heat that must be managed through airflow, liquid loops, exchangers or other engineering arrangements. A cooling design does not eliminate heat; it transports and ultimately rejects or recovers it. Changes to cooling efficiency may alter electricity, water and maintenance demands together.

A facility considering evaporative cooling should examine local climate, water availability, temperatures, equipment density, operational reliability and the actual measurement boundary. No generic statement that air or liquid cooling is always greener survives all those context changes. The appropriate technology depends on the job and location.

A two-resource decision, not a single winner

Cypress’s revised design reduces annual total electricity from 60,000 to 52,000 kWh while increasing site water use from 24,000 to 40,000 litres. The reduction is 8,000 kWh and the water increase is 16,000 litres. Because the resources have different units, subtracting them from one another produces no meaningful net score.

Decision-makers need separate constraints and a reasoned account of the local trade-off. A school may care about grid demand, while a water-stressed community may prioritise scarce water. Combining unlike resources into an arbitrary one-number ranking without a transparent valuation would hide more than it explains.

Electricity intensity depends on timing and location

Equal kilowatt-hours do not always correspond to equal estimated emissions. The generation mix can vary by grid, hour and season. An annual-average emission factor may be appropriate for a defined inventory method while being insufficient to answer a marginal question about moving one workload from afternoon to night.

The Cypress 0.4-kg-CO₂e-per-kWh factor is an invented constant used only for arithmetic. It does not describe Singapore’s grid, a commercial provider or any real facility. To publish an actual emissions estimate, an analyst must obtain the relevant source and disclose the accounting approach.

Calculate emissions with an explicit fictional factor

For Cypress at 60,000 kWh, multiplying by its invented 0.4 kg CO₂e/kWh gives 24,000 kg CO₂e. At 52,000 kWh, the same assumed factor gives 20,800 kg CO₂e. The modelled difference is 3,200 kg CO₂e under fixed intensity. The units are kilograms because kilowatt-hours cancel.

If the electricity supply mix or accounting rules changed between scenarios, retaining the same factor could be unjustified. A result can be correct within a printed model while incomplete as a real-world environmental comparison. Record the sensitivity of the result to the missing factor, not simply its final decimal.

Annual allocations are not marginal effects

Dividing Cypress’s 60,000 kWh by one million tasks assigns an average of 0.06 kWh, or 60 Wh, to each task. That does not establish that one additional prompt increases electricity consumption by exactly 60 Wh. The facility may draw idle power while waiting and its efficiency may change with utilisation.

A marginal energy measurement would require a defined comparison of system operation with and without the extra task, holding relevant conditions appropriately controlled. It may produce a different number from the annual average allocation. The distinction is essential when readers ask what one extra AI query caused.

An idle-server thought experiment

Imagine a fictional server that draws 200 watts while idle and 250 watts while completing a short additional task. Under a simplified ten-second interval, full observed energy at 250 watts is 250×10/3600 ≈ 0.694 Wh. The idle baseline would have been 200×10/3600 ≈ 0.556 Wh.

The incremental energy over that interval is approximately 0.139 Wh. This calculation does not account for cooling changes, upstream networks or altered later workload scheduling. It is a classroom demonstration of marginal-versus-total accounting, not a measured energy figure for any actual AI model.

Task categories cannot always share one average

A short factual text response and a long video-generation job may consume very different computational resources. Treating both as one equivalent query can make a per-task mean hard to interpret. The IEA’s April 2026 energy-and-AI update identifies changing use-case mixes as an important uncertainty, alongside efficiency gains and adoption.

A defensible workload report distinguishes relevant task families, input sizes, output lengths and completed outcomes. Allocate shared overhead through an explicit method and explain how it was chosen. The goal is not a deceptively precise universal ‘AI query footprint’ but useful comparisons between genuinely comparable work.

Annual demand can rise while per-task efficiency improves

Cypress begins with one million tasks at 60 Wh average facility energy each, equalling 60 million Wh or 60,000 kWh. Suppose an imaginary efficiency improvement lowers the average to 30 Wh, but annual task volume triples to three million. New total demand is 90 million Wh, or 90,000 kWh.

Per-task intensity is halved while absolute annual energy grows by 30,000 kWh. This arithmetic pattern illustrates the difference between efficiency and total demand. It does not prove that an efficiency improvement caused the usage increase; behavioural and market effects would require their own evidence.

A forecast is a scenario, not a reading of 2030

The International Energy Agency’s April 2026 report, Key Questions on Energy and AI, projects in its central case that total data-centre electricity demand rises from around 485 TWh in 2025 to about 950 TWh in 2030. That is a modelled outlook for all data centres, not a measured 2030 outcome or a claim that all consumption is caused by AI.

Readers should retain the report date, scenario assumptions and uncertainty. Future electricity demand depends on AI use patterns, improvements in hardware and software, grid constraints, investment and other developments. A regional projection cannot establish the energy use of one fictional classroom computation.

When a local impact claim needs more evidence

A proposed computing facility may affect electrical infrastructure, water systems, local employment and land use. Statements about these consequences need project-specific plans, permits and relevant public records. A general global forecast does not prove that a particular site is approved, operating or imposing a measured local burden.

A careful community report separates developer proposals, official approvals, modelled scenarios and observations from active infrastructure. Useful questions include whose measurements were used, whether the relevant resources are locally scarce and which commitments or monitoring processes can be verified.

Hardware lifecycle is outside the operating meter

Producing accelerators, circuit boards, servers and cooling equipment involves material extraction, manufacturing and transport. Replacing equipment early can change lifecycle impacts even if the new operating configuration reduces electricity per task. Such impacts are not captured by annual facility PUE or site WUE.

Life-cycle comparisons require relevant manufacturing data, device lifetimes, reuse or recycling assumptions and a clearly defined functional unit. A claim of lower operating energy is valuable but narrower than a conclusion of lower total lifecycle impact. Keep the stages separate until a consistent assessment links them.

Repair a misleading sustainability headline

Consider the invented claim: ‘Cypress’s environmental impact falls because PUE improved from 1.5 to 1.3.’ The premise about PUE is true in the teaching packet, but the conclusion overreaches. Site water rose, electricity sources and hardware lifecycles were not described, and no single comprehensive impact metric was calculated.

A correct limited version says: ‘Under matched annual boundaries, Cypress’s revised fictional configuration uses 8,000 fewer kWh of total electricity and 16,000 more litres of direct site water than the baseline. Its PUE improves from 1.5 to 1.3 while WUE rises from 0.6 to 1.0 L/kWh.’ The revised sentence is less sweeping and much more useful.

Independent transfer: a smaller reading-hall service

Invent a Reading Hall service with annual IT electricity of 20,000 kWh, total facility electricity of 30,000 kWh and direct site water of 10,000 litres. It completes 500,000 equivalent checked jobs. Calculate PUE, WUE, total facility electricity per job and site water per job using the stated boundaries.

The answers are PUE 30,000/20,000 = 1.5; WUE 10,000/20,000 = 0.5 L/kWh; facility energy 30,000/500,000 = 0.06 kWh = 60 Wh per job; and site water 10,000/500,000 = 0.02 litres = 20 mL per job. None is a measured property of a real site.

A transfer that changes resource direction

Keep the Reading Hall workload and IT energy, but suppose an alternative fictional design uses 26,000 kWh of facility electricity and 18,000 litres of direct site water annually. Its PUE becomes 26,000/20,000 = 1.3 and WUE becomes 18,000/20,000 = 0.9 L/kWh.

Electricity decreases by 4,000 kWh while site water increases by 8,000 litres. Per-job allocations become 52 Wh and 36 mL. The learner’s job is to report both changes without declaring an overall environmental winner or inventing a water-scarcity valuation. The transfer follows the same method under changed values.

A practical environmental-data audit

Before calculating, name the task, facility, meter boundary, observation period, data quality and desired decision. Separate IT energy from total facility energy and operational site water from upstream and lifecycle effects. Compute PUE and WUE using matched denominators, then report absolute electricity and water totals as well as ratios.

Specify whether a per-task result is observed, allocated or modelled. Distinguish forecasts from historical readings and use the current date and source when facts may change. Check carbon factors and local water context independently. An honest limitation is better than a precise-sounding estimate based on imaginary inputs.

Scope notes and source owners

The US Department of Energy’s guidance defines site-based WUE using annual water and IT energy. The IEA’s April 2026 update provides the contemporary, scenario-based data-centre energy outlook. These are real reference documents; they do not endorse the fictional Cypress or Reading Hall values.

Connect the mechanism with Data Centres, The Data Centre Efficiency Paradox, The Cost of Intelligence, the Super Intelligence curriculum and the How X Works master library. Every new numerical case here is an invented teaching fixture.