Artificial intelligence is becoming part of the working infrastructure of science. Machine-learning models now assist with protein-structure analysis, materials screening, microscopy reconstruction, climate emulation, surrogate modeling, inverse design, literature synthesis, experimental planning, and the control of automated laboratories. In many of these applications, AI does not merely accelerate an existing calculation. It changes the structure of the research workflow by making it practical to evaluate far more hypotheses, simulate more configurations, process larger datasets, or operate instruments in closed-loop campaigns.
Need help ? connect
The scientific benefits can be substantial, but the resource cost is often poorly represented. A research group may record accelerator hours, cloud expenditure, or the nominal thermal design power of a GPU while omitting host processors, memory, storage, networking, idle capacity, cooling, power conversion losses, and the water consumed both at the data centre and in electricity generation. The resulting accounting gap is especially important in AI-driven scientific research because models are frequently embedded inside repeated optimization, uncertainty-quantification, screening, or experimental-control loops. A single training run may be visible; the thousands of preliminary, failed, replicated, and downstream runs surrounding it are not.
The issue is not that every AI-assisted analysis is environmentally excessive. Small models executed on efficient hardware can consume little energy, and a well-designed surrogate may avoid far more expensive simulations or physical experiments. The issue is that scientific value and environmental cost are rarely measured within the same system boundary. Without that measurement, researchers cannot distinguish a genuinely efficient AI workflow from one that merely shifts cost into infrastructure that remains outside the project ledger.
AI Compute Is Physical Infrastructure
An AI model is an abstract mathematical object, but executing it is a physical process. Transistors switch, memory is refreshed, data move through interconnects, storage systems serve training samples, voltage regulators dissipate heat, and cooling equipment removes that heat from the facility. Electricity is converted into computation and ultimately into thermal energy. Water may then be used to reject that heat, while additional water may be consumed at the power plants supplying electricity.
This physical chain matters because the energy reported by an accelerator is only one component of total demand. For a workload running over the interval from $t_0$ to $t_1$, the IT energy can be represented as
$
E_{\mathrm{IT}} = \int_{t_0}^{t_1} P_{\mathrm{system}}(t)\,dt,
$
where $P_{\mathrm{system}}(t)$ is the instantaneous power of the complete computing system, including accelerators, CPUs, memory, local storage, and relevant network equipment. Measuring only GPU power replaces $P_{\mathrm{system}}$ with a partial term and therefore produces a lower bound rather than a complete result.
Facility electricity is commonly related to IT electricity through power usage effectiveness, or PUE:
$
E_{\mathrm{facility}} = \mathrm{PUE}\,E_{\mathrm{IT}}.
$
A PUE of 1 is the idealized lower bound in which every unit of facility electricity reaches IT equipment. Operating facilities incur additional demand from cooling, power distribution, lighting, controls, and other infrastructure, so measured PUE is normally above 1. PUE is useful, but it is not an environmental impact metric by itself. It says nothing about the carbon intensity of the electricity, the local scarcity of water, the cooling technology, or the scientific utility produced by the workload.
The scale of this infrastructure is increasing rapidly. The International Energy Agency’s Energy and AI analysis projects that global data-centre electricity consumption could reach roughly 945 TWh by 2030 in its base case, with AI as a major source of growth. At the national level, the 2024 United States Data Center Energy Usage Report from Lawrence Berkeley National Laboratory estimated 176 TWh of US data-centre electricity use in 2023 and developed 2028 scenarios ranging from approximately 325 to 580 TWh. These figures describe the data-centre sector rather than scientific AI alone, but they establish the infrastructure context in which AI-intensive research now operates.
Why Scientific AI Has Distinctive Resource Multipliers
Public discussion often reduces AI resource use to either the training of a frontier model or the cost of one inference request. Neither unit describes most scientific workflows. Research computation is organized around campaigns, not isolated calls.
A materials-discovery project may pretrain or adapt a graph neural network, generate millions of candidate structures, evaluate uncertainty, send selected candidates to density functional theory calculations, retrain the model, and repeat the cycle. A microscopy pipeline may reconstruct large image volumes, segment them with an ensemble, quantify structures, and rerun the analysis under several calibration assumptions. An autonomous laboratory may invoke a model after every batch of measurements and use the predictions to choose the next experiment. In each case, model execution is multiplied by outer loops that are scientifically necessary but easy to omit from environmental accounting.
Hyperparameter and Architecture Search
The published model is usually the survivor of a larger search process. Learning rates, data splits, augmentation policies, architectures, loss functions, regularization terms, random seeds, and precision modes may all be varied. Automated hyperparameter optimization can make this exploration systematic, but it can also convert a single training cost into tens or hundreds of training-equivalent runs. Early stopping reduces some waste, yet unsuccessful trials still consume resources before termination.
The relevant unit is therefore not the energy of the final checkpoint. It is the energy of the complete model-development campaign, including failed trials and diagnostic runs. Excluding them creates the computational equivalent of reporting only successful experiments while ignoring the material consumed by unsuccessful ones.
Ensembles, Uncertainty, and Reproducibility
Scientific claims often require more than point prediction. Researchers use ensembles, cross-validation, bootstrap resampling, posterior sampling, repeated seeds, ablation studies, and out-of-distribution tests to estimate uncertainty and establish robustness. These practices improve scientific quality, but they multiply inference or training demand.
The multiplier is not necessarily a defect. A ten-member ensemble that produces calibrated uncertainty may be more valuable than a single overconfident model. The correct question is whether the additional energy produces decision-relevant information. Energy per prediction is less informative than energy per calibrated prediction, energy per accepted candidate, or energy per unit reduction in uncertainty.
AI Coupled to Simulation and Experiment
AI is increasingly placed inside workflows that also contain conventional high-performance computing and laboratory equipment. A neural surrogate may reduce the number of finite-element or computational-fluid-dynamics solves, but training data may first require thousands of such solves. A model that schedules experiments may reduce reagent use while increasing computation. A foundation model for molecular science may accelerate candidate generation yet transfer validation work to molecular dynamics, quantum chemistry, or wet-lab assays.
Assessing only the AI component can therefore misrepresent the net effect. The comparison should be made between complete research pathways that deliver an equivalent scientific outcome. In some cases AI lowers total energy and water demand; in others it increases throughput so strongly that aggregate resource use rises even though each prediction becomes cheaper.
Data Preparation, Storage, and Movement
Large scientific datasets are expensive to move and maintain. High-resolution imaging, genomics, particle physics, remote sensing, and numerical simulation can generate data volumes that stress storage systems and networks. Training pipelines may repeatedly decode, shuffle, augment, and transfer the same samples. Poor data locality leaves accelerators idle while still drawing substantial power.
These effects are rarely visible in model-level estimates based on floating-point operations. FLOPs describe arithmetic work, not the full energy of memory access, communication, storage, checkpointing, or underutilized hardware. For modern AI systems, data movement and memory behavior can be as important as nominal compute.
The Water Footprint: Direct, Indirect, and Local
Water accounting is more difficult than electricity accounting because “water use” can refer to different physical quantities. Water withdrawal is the volume removed from a source, even if much of it is later returned. Water consumption is the portion not returned to the immediate watershed, commonly because it evaporates or is incorporated into another stream. The distinction is essential when comparing cooling systems and electricity-generation technologies.
Direct Water Use at the Data Centre
Many data centres use evaporative cooling because evaporating water can reject heat with lower electrical demand than fully mechanical dry cooling under suitable conditions. The trade-off is direct water consumption. Cooling towers also require blowdown to control dissolved solids, and water may be used for humidification or other facility processes.
Water usage effectiveness, standardized in ISO/IEC 30134-9:2022, expresses data-centre water consumption relative to IT energy. A simplified site metric is
$
\mathrm{WUE}_{\mathrm{site}} = \frac{W_{\mathrm{direct}}}{E_{\mathrm{IT}}},
$
where $W_{\mathrm{direct}}$ is direct water consumption, usually in litres, and $E_{\mathrm{IT}}$ is IT energy in kilowatt-hours. The resulting unit is typically L/kWh.
A low site WUE is not automatically superior. A facility may reduce direct water use by relying more heavily on compressor-based cooling, thereby increasing electricity consumption. If the local grid is carbon-intensive or water-intensive, the apparent water saving can shift impacts from the data centre to power generation. Cooling decisions must therefore be evaluated jointly through energy, carbon, and water metrics.
Indirect Water Use in Electricity Generation
Electricity can carry a substantial water footprint. Thermal power stations may withdraw and consume water for cooling; hydropower reservoirs can be associated with evaporative losses; and the relevant intensity varies with generation technology, climate, plant design, and accounting method. The indirect water attributable to a workload can be estimated as
$
W_{\mathrm{indirect}} = I_{W,\mathrm{grid}}\,E_{\mathrm{facility}},
$
where $I_{W,\mathrm{grid}}$ is the water-consumption intensity of delivered electricity in L/kWh. Total operational water can then be approximated as
$
W_{\mathrm{operational}} = W_{\mathrm{direct}} + W_{\mathrm{indirect}}.
$
The LBNL report estimated that US data centres directly consumed about 66 billion litres of water in 2023, while their indirect water footprint from electricity generation was nearly 800 billion litres under the report’s regional-grid methodology. The result should not be transferred uncritically to an individual project, because grid mixes, facility designs, contractual electricity arrangements, and local conditions differ. It does, however, show why reporting only cooling-tower water can omit the larger part of operational water demand.
Water Impact Is Geographically Specific
A litre of water consumed in a water-abundant basin is not equivalent to a litre consumed during drought in a highly stressed basin. Annual averages can also conceal seasonal stress. The same facility may impose limited pressure during a cool, wet month and become problematic during a hot period when cooling demand and municipal water demand peak together.
This spatial and temporal variability is central to research-computing decisions. The paper “Making AI Less Thirsty” formalized the distinction between on-site and off-site water use and emphasized that the water efficiency of AI workloads varies by location and time. Tools such as the World Resources Institute Aqueduct Water Risk Atlas can provide basin-level context, although such screening tools do not replace facility-specific data.
Scheduling computation to a lower-water period or region may help, but carbon and water objectives can conflict. A region with low-carbon electricity may use water-intensive generation or cooling; a region with low water consumption may rely on fossil generation. Multi-objective scheduling is therefore more defensible than optimizing a single environmental metric.
A Research-Grade Accounting Boundary
The most consequential methodological decision is not the equation used to calculate energy. It is the boundary around the system being measured. A narrow boundary can produce precise numbers that answer the wrong question.
For AI-driven scientific research, a useful boundary should include the complete computational campaign required to support a claim. That normally means data preprocessing, exploratory runs, training, tuning, validation, repeated seeds, inference, post-processing, and the share of storage and networking that is materially attributable to the project. When the AI workflow replaces or triggers simulations and experiments, the comparison boundary should include those coupled activities as well.
A practical inventory separates three layers:
| Layer | Typical quantities | Why it matters |
|---|---|---|
| Workload | Runtime, samples, tokens, simulations, GPU/CPU utilization | Connects resources to scientific activity |
| IT system | GPU, CPU, memory, storage, and network energy | Captures more than accelerator telemetry |
| Facility and supply | PUE, cooling water, grid carbon, grid water intensity | Converts IT activity into environmental impact |
Embodied impacts from manufacturing servers, accelerators, buildings, and cooling equipment form a fourth lifecycle layer. They are important, particularly when specialized hardware is replaced rapidly, but attribution is difficult because equipment supports many workloads over its service life. A transparent study should either allocate embodied impacts using a stated method—such as occupied device time over expected lifetime—or explicitly declare that they are outside the boundary. Silence is not a boundary definition.
Select a Functional Unit That Represents Scientific Value
Environmental results become interpretable only when normalized to a meaningful output. “One run” is rarely adequate because runs differ in dataset size, accuracy, and scientific purpose. Better functional units include:
- kilowatt-hours per validated sample or reconstructed volume;
- litres of water per accepted design candidate;
- joules per inference at a specified accuracy, latency, and batch size;
- energy per statistically independent simulation sample;
- energy per percentage-point reduction in predictive error;
- total campaign energy per published figure, benchmark result, or research claim.
No single unit fits every discipline. The requirement is equivalence: competing methods must be compared at comparable scientific quality. A smaller model that uses less energy but fails the required uncertainty or accuracy threshold is not an efficient substitute. Conversely, a larger model should not be credited merely for marginal benchmark improvement if that improvement does not change the scientific decision.
Report Both Absolute and Intensity Metrics
Efficiency ratios can improve while total demand grows. A laboratory may reduce energy per candidate by 60% but screen ten times more candidates, increasing aggregate energy. This is a form of rebound: lower unit cost expands use.
A credible report should therefore include both total resource use and normalized intensity. Total kilowatt-hours reveal infrastructure demand; kilowatt-hours per useful result reveal process efficiency. The same principle applies to water and carbon.
Measuring an AI-Driven Research Campaign
The measurement process should be designed before large runs begin. Retrospective estimation is possible, but it often loses information about failed jobs, utilization, location, and configuration.
Instrument the Whole Node Where Possible
External power distribution unit measurements provide the strongest operational evidence because they capture the complete server. When that is unavailable, platform telemetry can be combined across components. GPU telemetry from tools such as NVIDIA Data Center GPU Manager can be integrated over time, while CPU energy interfaces, job scheduler records, and memory estimates fill other parts of the system boundary.
Software tools such as CodeCarbon and the methodology behind Green Algorithms make resource estimation accessible, but estimates should not be treated as direct measurements. Power models, assumed utilization, facility factors, and grid-intensity databases introduce uncertainty. Where decisions depend on small differences between methods, external metering or calibrated telemetry is preferable.
At minimum, every run record should capture hardware type, device count, runtime, average and peak utilization, precision mode, batch size, model and dataset version, facility region, and whether energy represents the accelerator, node, or facility. The same metadata that supports reproducibility supports environmental auditing.
Preserve Failed and Abandoned Runs
Failed runs are part of the resource cost of discovery. Job logs should distinguish infrastructure failure, numerical instability, coding errors, unsuccessful hypotheses, and deliberately terminated trials. This classification is useful because mitigation differs. Better testing can reduce coding failures; checkpointing can reduce restart cost; scheduler changes can reduce pre-emption losses; and sequential experimental design can reduce scientifically uninformative trials.
Excluding failed runs also creates perverse incentives. A team with weak engineering discipline can appear as efficient as a team that invested in reliability, provided both report only their final successful job.
Convert Energy to Carbon and Water Transparently
Operational greenhouse-gas emissions can be estimated as
$
C_{\mathrm{operational}} = I_{C,\mathrm{grid}}\,E_{\mathrm{facility}},
$
where $I_{C,\mathrm{grid}}$ is the grid carbon intensity in kg CO$_2$e/kWh. Researchers should state whether the factor is annual-average, marginal, hourly, location-based, or based on contractual procurement. These choices answer different questions and can produce materially different results.
Water estimates require similar disclosure. State whether direct water is measured or inferred from WUE, whether indirect water is included, whether the metric refers to withdrawal or consumption, and whether water stress is incorporated. A single unqualified “litres used” value is not sufficiently precise for technical comparison.
Quantify Uncertainty Rather Than Hiding It
Environmental accounting is rarely exact. PUE may be a facility average rather than a workload-specific value. Grid intensity varies within an hour. Cloud providers may not expose host-level power. Storage and network energy may be estimated through allocation factors. Water intensity datasets may use different treatment of hydropower or cooling technologies.
A useful result can therefore be reported as a range or accompanied by sensitivity analysis. For example, a project may calculate a low case using measured IT energy and the provider’s best PUE, and a high case that adds plausible storage, network, and facility overhead. The goal is not to manufacture precision but to identify which assumptions dominate the result.
Reducing Energy and Water Without Weakening the Science
The strongest interventions change the research design rather than merely improving the efficiency of a fixed workload. The GREENER principles for environmentally sustainable computational science emphasize transparency, efficient algorithms, appropriate hardware, and reproducible reporting. In AI-driven science, these principles can be translated into decisions at the model, workflow, facility, and institutional levels.
Use the Smallest Model That Meets the Scientific Requirement
Model selection should be constrained by the required accuracy, calibration, robustness, context length, or physical consistency—not by the largest model available. A compact domain model can outperform a general foundation model on a narrow scientific task while using less memory and energy. Distillation, parameter-efficient adaptation, retrieval, and hybrid physics-ML architectures can reduce the need for full-model training.
The comparison must include preprocessing and support infrastructure. A small model paired with an expensive retrieval pipeline may not be smaller at the system level. Nevertheless, right-sizing remains one of the most direct ways to reduce both training and inference demand.
Improve Experimental Design Around Training
Hyperparameter searches should use informed priors, multi-fidelity methods, successive halving, early stopping, and small-scale pilot experiments. Before launching a large campaign, teams should verify data pipelines, numerical stability, checkpoint recovery, and metric implementation on reduced configurations.
A useful practice is to assign an explicit compute budget to the research question. The budget can be revised when preliminary evidence justifies expansion, but it prevents open-ended search from becoming the default. The decision criterion should be expected information gain per unit of compute rather than the number of configurations that can be afforded.
Increase Hardware Utilization
Accelerators that wait for data or synchronization consume energy without producing proportional scientific work. Profiling should identify input bottlenecks, communication overhead, load imbalance, excessive checkpoint frequency, and unsuitable batch sizes. Mixed precision, kernel fusion, efficient attention implementations, compilation, gradient accumulation, and data caching can improve throughput when they preserve numerical validity.
Benchmark results are configuration-specific. The ML.ENERGY initiative and MLCommons benchmarks illustrate the importance of measuring energy under defined workloads and quality constraints rather than relying on processor specifications. Thermal design power is a capacity rating, not a measurement of job energy.
Reduce Data Movement and Retention
Data locality can reduce both runtime and network activity. Scientific datasets should be staged near compute, represented in formats suited to sequential or parallel access, and cached when reuse is likely. Repeatedly converting or decompressing data during every epoch may be more expensive than a one-time transformation.
Retention policies also matter. Not every checkpoint, intermediate tensor, augmentation cache, or failed-run artifact has permanent scientific value. Teams can define tiers for archival, reproducibility, and temporary scratch data. This reduces storage growth without compromising the evidence needed to reproduce published results.
Choose Location and Time with Joint Carbon-Water Criteria
Flexible workloads can sometimes be moved to regions or periods with cleaner electricity, lower water stress, or favorable ambient conditions. Training, large inference campaigns, and simulation-data generation are often more schedulable than interactive services.
The optimization should remain multi-objective. Moving work solely to minimize carbon may increase water stress; minimizing direct cooling water may increase electricity use. Researchers should also consider data-governance constraints, latency, hardware availability, and whether transferring large datasets offsets the benefit.
Evaluate Avoided Work
AI can create environmental benefit when it replaces more resource-intensive activity. A surrogate model may avoid repeated multiphysics simulations. Active learning may reduce the number of laboratory experiments. Automated quality control may prevent failed manufacturing batches or unnecessary instrument time.
These avoided impacts should be estimated cautiously and only against a credible baseline. The baseline must deliver the same scientific function, and claimed savings should not assume that every AI prediction would otherwise require the most expensive alternative. Net impact is the difference between complete pathways, not the resource use of the AI component alone.
If you're working on related challenges in this area and would find guidance helpful, feel free to reach out: CONTACT US.
Institutional Responsibilities
Individual researchers can optimize code, but many determinants of energy and water use are controlled by institutions, funding agencies, cloud providers, and data-centre operators. Researchers often cannot see facility PUE, cooling technology, direct water use, or hourly grid factors. Procurement decisions may prioritize purchase price and peak performance while omitting lifetime energy, utilization, repairability, and cooling requirements.
Research institutions can improve this situation by making energy telemetry available through schedulers, publishing facility-level PUE and WUE, exposing carbon and water intensity by region, and including environmental fields in job accounting. Shared infrastructure can improve utilization relative to fragmented laboratory servers, provided that queue policies and capacity planning prevent chronic overprovisioning.
Grant applications for compute-intensive projects can include a resource plan analogous to data-management and ethics plans. The purpose should not be to penalize computationally demanding science. It should be to require a defensible relationship between requested compute and the scientific design. Reviewers can ask whether model scale is justified, whether pilot evidence exists, whether repeated runs are statistically necessary, and whether lower-cost alternatives were evaluated.
Publishers and conferences can also normalize reporting. A compact computational-resources statement could disclose hardware, total campaign energy or accelerator hours, measurement boundary, facility assumptions, and major efficiency interventions. Such reporting would improve reproducibility while building the evidence base needed to compare methods across disciplines.
Common Accounting Errors
Several practices repeatedly produce misleading conclusions.
First, dividing a training estimate by the number of model users does not describe the marginal cost of inference, and reporting one inference does not describe the infrastructure required to serve a system at scale. Training and inference are different workload classes and should be accounted separately before being combined.
Second, using the GPU’s rated power multiplied by runtime can overestimate or underestimate actual accelerator energy and still omit the rest of the system. Measured power traces are better; complete-node measurements are better still.
Third, a low PUE does not guarantee low carbon or water impact. PUE measures facility overhead, not electricity provenance or water stress. Similarly, a low WUE does not show whether additional electricity has shifted water use to power generation.
Fourth, renewable-energy certificates or annual clean-energy matching do not by themselves establish that a workload was powered by carbon-free electricity at the time and location of execution. Contractual accounting can be relevant, but it should not replace a location-based result when the objective is to understand physical grid impact.
Fifth, cloud credits do not make compute environmentally free. They remove the project’s direct financial signal while leaving the physical resource demand intact. In fact, zero-price compute can encourage poorly bounded experimentation unless teams establish independent resource budgets.
Finally, per-query numbers are highly sensitive to model, prompt length, output length, batch size, hardware, utilization, and system boundary. They should not be generalized beyond the measured configuration.
Conclusion
AI-driven scientific research is creating new capabilities, but it is also embedding science more deeply in energy- and water-dependent infrastructure. The hidden cost does not arise only from large model training. It accumulates through data preparation, exploratory runs, hyperparameter search, repeated seeds, ensembles, uncertainty analysis, inference campaigns, storage, networking, facility overhead, cooling, and electricity generation.
A rigorous assessment begins by defining the scientific function and drawing a boundary around the complete research campaign. It then measures whole-system energy where possible, converts that energy to facility demand, distinguishes direct from indirect water consumption, accounts for geographic water stress, and reports uncertainty in the result. Absolute resource use should be presented alongside efficiency metrics so that rebound effects remain visible.
The objective is not to discourage computational science or to treat every kilowatt-hour as unjustified. High-value research may warrant substantial resources, just as major experimental facilities do. The objective is to make those resources explicit, comparable, and governable. Once energy and water are treated as research variables rather than invisible externalities, teams can ask a more useful question: not merely whether AI improves a scientific metric, but whether it produces enough reliable knowledge to justify the complete physical system required to obtain it.
Interested in collaborating on academic research ? feel free to get in touch 🙂.
Check out YouTube channel, published research
you can contact us (bkacademy.in@gmail.com)
Interested to Learn Engineering modelling Check our Courses 🙂
