This publication distinguishes peer-reviewed research, preprints, engineering hypotheses and pending experimental measurements. References
IDEnza Research · Computational Sustainability Series · 2025

The Hidden Cost
of Artificial Intelligence

The energy, water and computational cost of intelligence — and the engineering challenge of doing more with less.

A scientific publication on the physical footprint of artificial intelligence, the cost of a single inference, and the architectural question that follows: how much of that computation is actually necessary?

Peer-reviewed Preprint Engineering hypothesis Pending measurement
01 — Introduction

The Invisible Cost of Intelligence

A request to an artificial intelligence system feels weightless. It is not.

Every prompt resolves into a measurable chain of physical events: silicon switching, memory moving, fans turning, water evaporating, generation capacity being scheduled. None of this appears on the screen.

Individually, the cost is small. In aggregate — hundreds of millions of requests per day, across millions of users — it becomes infrastructure: power purchase agreements, cooling plants, reservoirs, transmission lines.

This publication has a narrow and deliberately conservative purpose. It does not claim that artificial intelligence is unsustainable, nor that any specific architecture is environmentally superior. It presents what independent research has measured and projected, and it presents — transparently — an engineering hypothesis about unnecessary computation.

The question is not whether artificial intelligence consumes resources. It does.

The question is how much of that consumption is structurally necessary — and how much is the consequence of architectures that compute again what has already been computed.

Answering it honestly requires two things that are often missing from the public debate: a rigorous account of the scientific evidence, and a willingness to publish measurements that do not yet exist.

415 TWh · 2024Electricity consumed by data centres worldwide — approximately 1.5% of global electricity demand.
731–1,125 million m³ / yearProjected annual water footprint of AI servers in the United States, 2024–2030.
24–44 Mt CO₂e / yearAdditional annual emissions projected for US AI servers over the same period.
>65× energy intensity spreadReported difference in inference energy intensity between LLM systems under comparable conditions.
Reading note

Each figure on this page is labelled by its evidentiary status: peer-reviewed measurement, preprint estimate, institutional projection, or pending experimental validation. Scientific projections and IDEnza experimental data are never presented on the same axis.

02 — Energy

The Energy Scale of Digital Infrastructure

Before discussing artificial intelligence specifically, it is necessary to establish the scale of the infrastructure that hosts it.

The International Energy Agency estimated that data centres consumed approximately 415 TWh of electricity in 2024 — roughly 1.5% of global electricity consumption.

In its base-case projection, the IEA indicates approximately 945 TWh by 2030, driven by the expansion of digital services and accelerated computing capacity.

These figures describe data centres as a whole. They are not an attribution of consumption to artificial intelligence alone, and they are presented here exclusively as context for the physical scale of the digital infrastructure on which AI systems run.

Figure 01 Projection · Institutional

Global data centre electricity consumption — 2024 observed, 2030 base-case projection

Source: International Energy Agency, Energy and AI. Unit: TWh. Region: global. 2024 value is an estimate; 2030 value is a base-case projection. Scope: all data centre workloads — not exclusively artificial intelligence.
iea.org/reports/energy-and-ai
Scope note

A meaningful share of data centre electricity is consumed by non-AI workloads — storage, networking, databases, streaming. Attributing the full 415 TWh to artificial intelligence would be inaccurate. The relevant reading is that AI capacity is being added on top of an already substantial and growing digital load.

03 — Water

Direct and Indirect Water Consumption

Water enters the system twice: on-site, for cooling, and upstream, in the generation of electricity.

The Nature Sustainability study of US AI servers projects an annual water footprint between 731 and 1,125 million cubic metres across the 2024–2030 period.

Approximately 71% of that footprint is indirect — water consumed in the generation of the electricity the servers draw from the grid. The remaining 29% is direct: on-site cooling and facility operation.

This distinction matters for engineering. A strategy that reduces the volume of computation required does not act on cooling equipment — it acts, potentially, on the upstream electricity demand. The two are connected, but the connection is neither linear nor guaranteed.

Figure 02 Projection · Peer-reviewed

Projected annual water footprint of US AI servers, 2024–2030

Source: Xiao et al., Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA, Nature Sustainability (2025). Unit: million m³ per year. Region: United States. Type: projection range.
Figure 03 Estimate · Peer-reviewed

Composition of the projected water footprint — indirect versus direct consumption

Source: Xiao et al., Nature Sustainability (2025). Indirect consumption corresponds to water used in electricity generation; direct consumption corresponds to on-site cooling and facility operation.
Important distinction

The study reports that optimising server and infrastructure efficiency could reduce energy consumption by up to 12% and the water footprint by up to 32% under a best-practice industrial scenario. Combined best practices, siting and grid decarbonisation could reduce emissions by up to 73% and the water footprint by up to 86% relative to the reference scenario.

These are scenario results from combined mitigation pathways. They are not reductions produced by software, by a protocol, or by any single AI architecture. They are not attributed to IDEnza.

04 — Inference

The Cost of a Single Request

Aggregate infrastructure statistics conceal an important variable: the efficiency of the operation itself.

The 2025 benchmark study How Hungry is AI? measured energy, water and carbon associated with the inference phase of thirty language models, comparing short and long requests under controlled conditions.

Its central finding is not a single number but a range: under certain comparison conditions, the difference in energy intensity between the most and least efficient systems exceeded 65 times.

The study also illustrates how small per-request costs compound. A short request estimated at approximately 0.42 Wh, multiplied by 700 million requests per day, produces an annual energy impact comparable to the consumption of roughly 35,000 US households.

The engineering implication is straightforward: at this scale, the difference between a necessary inference and an unnecessary one is not a rounding error.

Figure 04 Schematic · Preprint

Reported spread in inference energy intensity between LLM systems

Source: Jegham et al., How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference, arXiv:2505.09598 (2025). Unit: relative energy intensity. Type: preprint estimate. This figure is a schematic representation of a reported ratio — it is not a ranking, and it does not identify individual models.
Preprint caveat

This work is a preprint. Its estimates derive from public data, environmental multipliers and inferences about infrastructure. They are not direct per-server measurements. The values are reproduced here with that qualification, and they should be read as an indication of magnitude and variance — not as a measurement of any specific provider, model or deployment.

05 — Evidence base

The Scientific Record

Three independent sources underpin this publication. Each is identified by type, and each is linked to its original record.

Peer-reviewed

Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA

Tianqi Xiao et al.
Nature Sustainability · 10 Nov 2025
DOI 10.1038/s41893-025-01681-y

  • Annual water footprint projected at 731–1,125 million m³.
  • Additional annual emissions projected at 24–44 Mt CO₂e.
  • ≈71% of the water footprint is indirect (electricity generation); ≈29% is direct.
  • Best-practice efficiency could cut energy up to 12% and water up to 32%.
  • Combined mitigation scenarios: up to 73% emissions and 86% water reduction.
nature.com/articles/s41893-025-01681-y
Preprint

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference

Nidhal Jegham et al.
arXiv:2505.09598 · 2025
Preprint — not peer-reviewed

  • Benchmarks energy, water and carbon during inference across 30 models.
  • Reports efficiency differences exceeding 65× under certain comparison conditions.
  • A short request at ≈0.42 Wh, multiplied by 700M daily requests, is estimated to equal the annual consumption of ≈35,000 US households.
  • Compares short and long requests, and model-level intensity differences.
arxiv.org/abs/2505.09598
Institutional assessment

Energy and AI

International Energy Agency
Report · 2025
Contextual data — all data centres

  • Data centres consumed ≈415 TWh of electricity in 2024.
  • Equivalent to approximately 1.5% of global electricity consumption.
  • Base-case projection of ≈945 TWh by 2030.
  • Scope covers all data centre workloads — not artificial intelligence exclusively.
iea.org/reports/energy-and-ai
06 — Computational cost

The Anatomy of a Request

A single prompt is not a single operation. It is a pipeline, and each stage carries a physical cost.

01Input processing
02Tokenisation
03Inference execution
04CPU / GPU utilisation
05Memory & storage
06Data transfer
07Cooling
08Electricity & external infrastructure

Training

The development and adjustment of a model. It can require very large volumes of computation concentrated in a defined period, and it determines the capability of the model that will later be served.

Inference

The execution of a trained model to produce a response, an analysis or a result. Inference cost scales with usage, not with development, and is therefore the dominant cost of deployed systems.

Reprocessing

The execution of operations that could, in principle, have been avoided by retrieving information that had already been processed. This is the category that architectural design can most directly influence.

Dependency note

Total computational cost depends on the model, the hardware, the volume of input data, the number of operations performed and the infrastructure on which they run. No single metric — tokens, latency or watt-hours — describes it in isolation. This is why the benchmark described later in this publication reports all three.

07 — Architecture

A Different Approach to Computational Efficiency

Reducing unnecessary computation through intelligent semantic processing.

The IDEnza Generative Protocol (IGP) is a computational architecture oriented toward intention, semantic retrieval, contextual continuity and the selective use of external artificial intelligence resources.

It is not a conventional language model. It does not attempt to replace the large external models on which it depends; it governs when, with what context, and whether those models are invoked.

The architectural premise is simple to state and difficult to engineer: in conventional systems, a request can trigger a new inference even when the information required to answer it has already been processed and is available. An architecture with retrieval and reuse mechanisms can, under certain conditions, avoid some of those operations.

IGP investigates and implements this principle. Its objective is not to eliminate external artificial intelligence, but to use it more selectively.

Path A — Conventional request handling
Request
Context assembly
Full inference
Response

Every request consumes a new inference, regardless of whether the answer already exists in prior processing.

Path B — Retrieval-mediated request handling
Request
Semantic retrieval
Resolution check
Selective inference
Response

Where existing knowledge resolves the request, external inference may be avoided. Where it does not, the request is forwarded with a more selective context.


Public architectural principles

01 Intelligent retrievalRecovery of information already available within the system.
02 Fewer unnecessary callsReduction of external invocations that would not add information.
03 Reuse of processed knowledgePreviously computed results remain available rather than being recomputed.
04 Retrieval before generationContextual recovery is attempted before a new generation is requested.
05 Less redundant processingIdentical or equivalent operations are not executed repeatedly.
06 Appropriate escalationRequests are routed to external processing when that is genuinely required.
07 Semantic continuityContext is preserved across interactions rather than rebuilt from zero.
08 Token efficiencyContext sent to external models is selective rather than exhaustive.
Disclosure boundary

This publication describes the principles of the IDEnza Generative Protocol and nothing beyond them. Source code, proprietary structure, internal architecture, configurations, data structures, storage mechanisms and industrial-secret details are not disclosed. The principles above are published so that the resulting measurements can be independently reproduced.

08 — From computation to environment

From Computational Efficiency to Environmental Efficiency

The link between an architectural decision and a water footprint is a chain, not a leap. Each link must be measured independently.

Table 01 — Identified problem and the corresponding architectural principle under investigation
Identified problem Potential contribution of the IGP architecture
Repeated inferenceRetrieval of existing knowledge
Unnecessary external callsTriage and retrieval before generation
Context reprocessingContextual continuity across interactions
Unnecessary token consumptionMore selective context transmission
Overuse of complex modelsRouting appropriate to the actual requirement
Redundant computational operationsIntelligent reuse of prior results
Fewer external callsmay mean fewer inferences executed
Fewer tokens processedmay mean less computational work
Less computational workmay reduce energy demand
Lower energy demandmay reduce indirect water consumption from generation
Epistemic status

Every link in this chain is conditional. It depends on the infrastructure, the model, the energy mix of the grid, the utilisation of the hardware and the operating conditions of the specific deployment. No proportional environmental reduction is claimed here. The chain describes a mechanism that must be measured, not a result that has been obtained.

09 — Methodology

IDEnza Computational Efficiency Benchmark

A controlled comparison between direct use of an external model and the same model used through the IGP architecture.

The benchmark defines two arms over an identical workload:

Arm A — Baseline. The external model is used directly, without the IGP architecture.

Arm B — IGP-mediated. The same external model is used through the IGP architecture.

Both arms share the same request set, the same input content, the same external model and, wherever possible, the same processing conditions. Repetitions are executed to characterise variance. Measured values and estimates are recorded separately.

The current specification compares direct calls to the Groq API against calls routed through the integrated IGP Worker. The laboratory records observed latency, HTTP status, provider-reported token counts, declared response origin and external call counts, with JSON and CSV export.

No efficiency result is published yet. The initiative has not produced validated comparative measurements. This page will populate automatically when real data becomes available.

Table 02 — Required indicators for the comparative benchmark
Indicator Unit Measurement basis
External callscountMeasured
Input tokenstokensProvider-reported
Output tokenstokensProvider-reported
Total tokenstokensDerived
Tokens avoidedtokensDerived
Retrieval rate%Measured
Resolution without external inference%Measured
LatencymillisecondsMeasured
Electricity consumptionWhMeasured or explicitly modelled
Water footprintlitresMeasured or explicitly modelled
Operating costBRLProvider pricing + infrastructure

Reduction formulas

Token reduction Δtokens = [ 1 − ( TIGP / Tbase ) ] × 100
Call reduction Δcalls = [ 1 − ( CIGP / Cbase ) ] × 100
Energy reduction Δenergy = [ 1 − ( EIGP / Ebase ) ] × 100
Water reduction Δwater = [ 1 − ( WIGP / Wbase ) ] × 100
Applicability condition

The energy and water formulas may only be reported when comparable measurements exist, or when estimates are explicitly identified as such and accompanied by a documented methodology. Financial cost must account for actual provider pricing, owned infrastructure, storage and processing. Token and call reductions alone do not constitute an environmental result.

10 — Experimental indicators

Results

Experimental measurement — pending validation.

External calls—count
Total tokens—tokens
Tokens avoided—tokens
Resolution without external inference—%
Latency—milliseconds
Electricity consumption—Wh
Water footprint—litres
Operating cost—BRL
Status of the experimental programme

The fields above are intentionally empty. The IGP benchmark specification exists, the measurement harness has been defined, and the data collection protocol has been described — but no validated comparative results have been produced. No percentage of savings is displayed, simulated or implied.

When real measurements are released, this section will render them from a versioned data file, with the measurement date, sample size and confidence interval reported alongside every figure.

The future of artificial intelligence is not only about building more powerful systems. It is also about building systems that use intelligence, energy and computational resources more responsibly.

More intelligence. Less unnecessary computation.

11 — References

Scientific References

  1. Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA Xiao, T. et al. — Nature Sustainability, 10 November 2025. https://doi.org/10.1038/s41893-025-01681-y
  2. How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference Jegham, N. et al. — arXiv:2505.09598, 2025. Preprint, not peer-reviewed. https://arxiv.org/abs/2505.09598
  3. Energy and AI International Energy Agency — Report, 2025. https://www.iea.org/reports/energy-and-ai