Within AI Power Demand
How Much Electricity Does Running AI Actually Use?
Training AI models requires powerful computing clusters, while billions of everyday AI requests can create substantial ongoing electricity demand.
Published
On this page
- Training versus inference: where the electricity goes
- How AI model size, reasoning and repeated requests change energy consumption
- Why more efficient AI chips may not reduce total electricity demand
Page outline Jump by section
Updated
Introduction
AI electricity demand is driven by two different activities: training models and using them. Training requires large clusters of specialised chips working intensively for weeks or months. Inference —the process of generating an answer, image or prediction from a trained model —usually consumes far less electricity per task, but can happen billions of times.
The key distinction is between the electricity needed to build an AI model and the ongoing electricity needed to operate it at scale. Neither can be judged in isolation. A model that is expensive to train may become relatively efficient over millions of uses, while a cheap individual AI request can contribute to substantial electricity demand when repeated across a global service.
Research published in 2026 suggests that an ordinary text request to an efficiently operated AI system can consume less than one watt-hour. More demanding reasoning tasks can use over ten times as much. Meanwhile, the International Energy Agency (IEA) that electricity consumption from AI-focused data centres will triple between 2025 and 2030. ScienceDirect
For readers assessing AI’s energy costs, three comparisons matter most:
- Training versus inference: Training creates large, concentrated electricity loads; inference creates continuing demand that grows with adoption.
- Simple versus complex requests: Short text responses can be relatively efficient, while long reasoning tasks and automated workflows require substantially more computation.
- Efficiency versus total consumption: Better chips can reduce electricity per calculation without reducing total demand if companies perform many more calculations.
Understanding these differences is essential for judging whether AI electricity forecasts are realistic, which infrastructure investments are necessary and where the resulting costs fall.
Training versus inference: where does the electricity go?
Training and inference use similar underlying hardware, but their electricity demand follows different patterns.
| Factor | AI training | AI inference |
| Main purpose | Develop or improve a model | Produce outputs using a trained model |
| Typical workload | Repeated calculations across large datasets | Processing prompts and generating responses |
| Electricity pattern | Intensive, often prolonged computing runs | Continuous, variable demand from users and applications |
| Main driver | Training computation, chip numbers and run duration | Request volume, model choice and computation per request |
| Main infrastructure challenge | Supplying enough power to large computing clusters | Serving large volumes of requests efficiently and reliably |
| Main uncertainty | Future training scale and frequency | Future adoption and complexity of AI use |
These differences matter because a data centre built primarily for training can face different power requirements from one designed to serve millions of users throughout the day.
Training: concentrated electricity consumption
Training a large AI model involves processing enormous quantities of data and repeatedly adjusting billions or trillions of numerical values known as parameters. These parameters determine how the model processes information and generates outputs.
The calculations are carried out by specialised processors, including graphics processing units (GPUs) and tensor processing units (TPUs). Large training systems connect thousands of these chips through high-speed networks so that they can work together.
The electricity required depends on several interacting factors: the number and type of chips, their power consumption, the duration of the training run, how effectively the hardware is utilised and the electricity needed for supporting equipment.
A simple calculation illustrates the scale.
Illustrative training workload
Average computing load
Duration
Electricity consumed by the computing load
50 MW × 24 hours × 60 days = 72,000 MWh. This is an illustrative calculation, not a measurement of a particular model. Cooling and other data centre overheads would increase the facility total.
A gigawatt-hour (GWh) is one million kilowatt-hours. The example shows why even a single major training run can require substantial electricity, despite occupying a relatively short period compared with a data centre’s operating life.
Research organisation Epoch AI estimated that training Google’s Gemini Ultra model required approximately 35 megawatts of electrical power. Its 2024 analysis also found that electricity accounted for approximately 2–6% of the estimated development costs of selected frontier models, while computing hardware and research staff accounted for much larger shares. Epoch AI
This reveals an important distinction between financial and physical costs. Electricity may represent a relatively small share of an AI company’s overall model-development expenditure, yet securing tens or hundreds of megawatts of power can still determine whether a training project is feasible.
Training also extends beyond the final, most visible run. Developers conduct experiments, test different designs, prepare datasets, fine-tune models and sometimes repeat unsuccessful runs. Measuring only the final training operation can therefore understate the electricity used throughout development.
Epoch AI’s 2025 analysis suggested that training, experiments and inference each accounted for substantial shares of AI computing power demand, with no single activity necessarily dominating across the entire industry. These are estimates rather than a comprehensive measurement of every AI operator. epochai.substack.com
Inference: smaller tasks, repeated continuously
Inference begins when someone uses a trained model. Asking a chatbot a question, generating an image, translating a document or requesting a prediction all require the system to perform new calculations.
Unlike training, which adjusts a model’s parameters, inference generally uses parameters that have already been learned.
A single inference request may be relatively inexpensive in electricity terms. But a successful AI service can receive requests continuously from individuals, businesses and automated software.
The basic relationship is straightforward:
Total inference electricity = electricity per request × number of requests.
For example, if an AI service consumed an average of 0.3 watt-hours (Wh) per request and processed one billion requests daily, it would use approximately 300 megawatt-hours (MWh) per day, or 109.5 GWh annually.
That calculation is illustrative. Actual electricity use depends on the mixture of requests, hardware and supporting infrastructure.
Inference can eventually consume more electricity than the original training run because the model may remain in use for years. However, this is not inevitable: a specialised model with relatively few users may never accumulate enough inference demand to exceed its development energy.
The 2026 Stanford Institute for Human-Centered Artificial Intelligence AI Index identifies inference as a growing share of AI’s energy footprint and notes that widely deployed models can accumulate inference energy exceeding their one-off training consumption within months. AI INDEX REPORT 2026
The implication is that reducing training energy alone cannot guarantee that an AI service has a small long-term electricity footprint.
How much electricity does an AI request actually use?
There is no universal electricity figure for an AI query.
A request to summarise a short paragraph is fundamentally different from one that processes a large document, performs extensive reasoning or generates video. Even apparently identical requests can consume different amounts of electricity depending on the model, hardware, operating conditions and number of users sharing the system.
Nevertheless, recent research provides useful reference points.
What recent measurements show
In August 2025, Google published an assessment of the electricity consumed by its Gemini AI assistant. It reported that a median Gemini text prompt used approximately 0.24 Wh, including the active AI processors, supporting computing equipment, idle capacity and data centre overheads.
The breakdown was revealing:
| Electricity component | Energy per median prompt |
| Active AI accelerators | 0.14 Wh |
| Supporting processors and memory | 0.06 Wh |
| Allocated idle equipment | 0.02 Wh |
| Data centre overhead | 0.02 Wh |
| Total | 0.24 Wh |
Source: Google’s 2025 technical paper, Measuring the environmental impact of delivering AI at Google Scale[services.google.com]services.google.comServices Measuring the environmental impact of delivering AI at Google ScaleServices Measuring the environmental impact of delivering AI at Google Scale. Google Services
The same research found that a narrower measurement method produced a figure of just 0.10 Wh. The difference illustrates how easily electricity estimates can become misleading when studies count different parts of the computing system.
Google’s measurement is valuable because it reflects production infrastructure rather than a laboratory-only test. However, it remains a company-reported result for a particular service, period and request distribution. It should not be treated as a universal figure for AI.
A separate study published in the journal Joule in April 2026 estimated a median of 0.31 Wh per ordinary text query for large models operating under realistic, optimised conditions. Its estimated middle range was 0.16–0.60 Wh.
The researchers also found that some widely circulated estimates overstated ordinary inference electricity consumption by factors of four to twenty because they did not adequately represent efficient production systems. ScienceDirect
These results support a more precise conclusion than either claiming that every AI query is environmentally enormous or that AI electricity consumption is negligible.
Individual ordinary text requests can be comparatively efficient. Their collective electricity demand can still be substantial, particularly as the complexity and frequency of requests increase.
What happens at a billion requests a day?
The following figures illustrate how total electricity consumption changes with request volume and energy intensity. They assume the same average electricity per request throughout the year.
Illustrative inference electricity calculator
Requests per day
100 million
1 billion
10 billion
Average electricity per request
0.3 Wh
0.1 Wh
5 Wh
Estimated annual electricity
0.3 GWh per day
Illustrative arithmetic only, not a forecast of any provider’s electricity use. Assumes 365 days of demand and an all-in average electricity figure per request.
This comparison makes clear why the composition of AI traffic matters. A service with fewer but more demanding requests may consume as much electricity as one processing a much larger number of short requests.
It also explains why forecasts based only on user numbers can be unreliable. The amount of computation performed on behalf of each user is equally important.
How model size, reasoning and repeated requests change electricity use
The electricity required for inference depends not just on how often people use AI, but on what they ask it to do.
Three factors are particularly important: model architecture, the amount of information processed and the amount of work performed before producing a response.
Larger models do not always mean proportionally higher electricity use
AI models differ in the number of parameters they contain. These are the numerical values learned during training.
A larger model may require more memory and computation, but parameter count alone is an unreliable guide to electricity consumption.
Some models use a mixture-of-experts architecture, in which only selected parts of a larger model are activated for a particular task. Others use techniques that reduce the amount of memory or computation required.
The relationship between model size and electricity use also depends on how efficiently a model runs on particular hardware.
Research published in the 2026 AAAI conference proceedings introduced TokenPowerBench, a method for measuring the power consumption of large language model inference. Its experiments examined how model choice, request length, hardware configuration, parallel processing and numerical precision influence electricity use. AAAI Publications
The practical lesson is that the largest available model is not necessarily the most electricity-efficient option for a particular task.
For straightforward classification, translation or summarisation, a smaller specialised model may provide an adequate result while requiring less computation. For complex scientific or technical work, a more demanding model may be justified if it produces materially better results or avoids repeated unsuccessful attempts.
The meaningful comparison is therefore electricity consumed per useful result, not simply electricity consumed per model invocation.
Longer reasoning can multiply electricity consumption
Modern AI systems increasingly perform additional calculations before producing their final answer.
This is sometimes called test-time computation or reasoning. Rather than generating a short response immediately, the system may work through multiple intermediate steps, evaluate alternatives or use tools.
These operations can substantially increase the number of tokens processed. A token is a small unit of text that an AI model reads or generates.
The April 2026 Joule study estimated that a conventional query producing around 300 output tokens required a median of approximately 0.31 Wh under its modelled conditions. A longer reasoning task producing around 5,000 output tokens required approximately 3.91 Wh —about thirteen times as much. ScienceDirect
Estimated electricity per query under comparable modelled conditions
Joule 2026 study median estimates for standard and long reasoning queries, not universal measurements.
0 Wh 2 Wh 4 Wh 6 Wh Standard text Long reasoning
- Standard text: Electricity per query 0.3 Wh
- Long reasoning: Electricity per query 3.9 Wh
Source: Oviedo and colleagues, Joule (2026). Estimates depend on the assumed models, hardware and output lengths.
The additional electricity may be worthwhile when the task demands careful analysis. However, applying extensive reasoning to every simple request can increase electricity consumption without providing a proportionate benefit.
This creates a design choice for AI providers: when should a system use its most capable, energy-intensive mode, and when should it select a faster, lighter approach?
Automated AI agents introduce another source of demand
An AI agent is software that uses an AI model to carry out a sequence of actions, sometimes without requiring a person to approve every step.
For example, a request to prepare a business report might trigger searches, document analysis, calculations, repeated model calls and revisions.
What appears to the user as one task may involve dozens or hundreds of individual inference operations.
The electricity footprint depends on the number of calls, the complexity of each call and any additional tools or computing resources used.
This distinction matters because the number of visible user requests may become a poor measure of actual AI activity.
The IEA’s April 2026 assessment identifies reasoning, automated agent tasks and video generation as increasingly important sources of electricity demand. It warns that some advanced applications can consume hundreds or thousands of times more electricity per task than simple text generation. IEA
A shift towards these workloads could therefore increase data centre electricity consumption even without a comparable increase in the number of human users.
Why more efficient AI chips may not reduce total electricity demand
AI hardware and software are becoming substantially more energy-efficient. That is a genuine technical achievement, but it does not automatically mean that the industry will use less electricity overall.
The distinction is between energy intensity —electricity required per unit of work —and total demand —electricity consumed across all work performed.
Efficiency gains can be overtaken by greater usage
Consider a hypothetical AI service that halves its electricity consumption per request while tripling the number of requests it processes.
Its total electricity consumption rises by 50%, despite each request becoming twice as efficient.
This is an example of the rebound effect: when a service becomes cheaper or more efficient, increased usage can offset some or all of the savings.
The effect is particularly relevant to AI because efficiency improvements can make new applications economically viable. Cheaper inference may encourage businesses to embed AI into more products, automate additional processes or run models continuously.
The IEA’s 2026 analysis reports that electricity use per individual AI task has fallen exceptionally quickly in recent years, while the overall electricity consumption of AI-focused data centres increased by approximately 50% during 2025. IEA
These findings are not contradictory. They describe different measures: electricity per task and total electricity consumed.
Better hardware can still make a substantial difference
Efficiency improvements are not meaningless simply because total demand continues to rise.
Modern AI systems can reduce electricity consumption through several mechanisms:
- More efficient accelerators: Chips perform more useful calculations per watt.
- Reduced numerical precision: Models use less detailed numerical representations where this does not materially affect output quality.
- Batching: Multiple requests share computing resources, reducing wasted capacity.
- Model compression: Smaller or simplified models reproduce useful capabilities with fewer calculations.
- Intelligent routing: Straightforward tasks use smaller models, reserving larger models for difficult work.
- Better utilisation: Software reduces the time expensive hardware remains powered but underused.
The 2026 Joule study identified combinations of hardware, model and serving improvements that could potentially reduce electricity per query by factors of eight to twenty relative to its baseline. These are estimated opportunities under specified assumptions, not guaranteed reductions across the entire AI industry. ScienceDirect
Google separately reported a thirty-threefold reduction in electricity consumption for its median Gemini text prompt over a twelve-month period. That company-specific finding demonstrates how rapidly efficiency can improve, although it does not establish how much electricity Google’s total AI operations saved or consumed over the same period. Hugging Face
The unresolved question is whether efficiency gains will outpace growth in total computation.
Efficiency also involves performance trade-offs
Operating an AI model at maximum speed is not necessarily the same as operating it at minimum electricity consumption.
Processing several requests together can improve hardware utilisation, but waiting to assemble a batch may increase response times. Using a smaller model can save electricity, but may reduce accuracy on difficult tasks. Limiting reasoning can shorten a response, but may also make complex work less reliable.
Data centre operators must balance electricity consumption against response speed, quality, reliability and the cost of computing hardware.
This is why a useful energy comparison should consider both the electricity consumed and the quality of the result. A system that uses less electricity per attempt but repeatedly fails may not be more efficient overall.
Why AI workloads create particular problems for electricity grids
Electricity grids must supply enough power when it is needed, not merely enough energy over the course of a year.
This makes the distinction between power and electricity consumption important.
Power, measured in kilowatts or megawatts, describes the rate at which electricity is being used. Energy, measured in kilowatt-hours or megawatt-hours, describes the amount consumed over time.
A data centre may have manageable annual electricity consumption but still require an unusually large grid connection because of its maximum power demand.
Training clusters can create sudden power changes
AI training involves periods of intensive calculation interspersed with communication between processors and other activities.
Because many chips operate together, their electricity consumption can rise and fall in a synchronised way.
The IEA’s 2026 report explains that AI workloads can produce power fluctuations over periods ranging from fractions of a second to several minutes. These changes can create additional challenges for electrical equipment and local grid stability. IEA Blob Storage
The problem is not simply that a training cluster consumes a large amount of electricity. It is that its demand can change quickly while remaining concentrated at a single location.
Operators may use batteries, electrical equipment and software controls to smooth these fluctuations. Some approaches involve maintaining more consistent power consumption even when useful computing activity temporarily falls, which can reduce the efficiency of electricity use.
These operational characteristics influence the infrastructure required to support AI computing.
Inference creates a different planning challenge
Inference demand is shaped by how people and businesses use AI services.
Requests may rise during working hours, follow regional patterns or surge when a popular service experiences unusually high demand.
Providers must maintain sufficient computing capacity to deliver acceptable response times, even when demand is unpredictable.
Keeping spare capacity available improves reliability but can mean that some equipment consumes electricity while performing little useful work.
The challenge becomes greater when providers offer demanding applications that require many processors to work together or produce long responses.
For grid planners, the important figures are therefore not only annual electricity consumption but also peak demand, the rate at which demand changes and how much consumption can be shifted or controlled.
How much of future data centre electricity demand will come from AI?
AI is a major driver of new data centre investment, but it is not responsible for all data centre electricity consumption.
Traditional cloud services, online storage, business computing and other digital activities continue to consume substantial amounts of electricity.
Separating these activities is difficult because the same facilities may operate several types of workload.
The IEA’s April 2026 assessment provides one of the clearest global comparisons.
Projected global data centre electricity consumption
IEA 2026 central outlook for all data centres, not exclusively AI.
0 TWh 250 TWh 500 TWh 750 TWh 1K TWh 2025 2030
- 2025: Electricity consumption 485 TWh
- 2030: Electricity consumption 950 TWh
IEA, Key Questions on Energy and AI (2026). The 2030 figure is a projection, not an observed outcome. IEA
The IEA estimates that global data centres consumed approximately 485 TWh in 2025 and around 950 TWh in 2030.
Within that total, electricity consumption from AI-focused data centres is projected to more than triple, reaching approximately 465 TWh by 2030. IEA Blob Storage
These AI-focused facilities are identified by their extensive use of specialised computing accelerators. Their electricity consumption is a useful indicator of AI-related demand, but it is not a perfect measurement of electricity used exclusively for AI tasks.
The IEA’s earlier 2025 analysis projected annual electricity consumption growth of approximately 30% for accelerated servers, compared with 9% for conventional servers. Accelerated servers were expected to account for nearly half the increase in global data centre electricity consumption between 2024 and 2030. IEA
The scale of future demand depends on several uncertain developments: how many AI chips are installed, how efficiently they operate, how intensively they are used and how rapidly new AI applications spread.
It also depends on whether announced data centre are completed. Grid connection delays, shortages of equipment, financing conditions and uncertain commercial returns could slow expansion.
These uncertainties explain why electricity forecasts should be treated as scenarios rather than fixed predictions.
Who pays when training and inference demand more electricity?
The immediate electricity bill is generally paid by the company operating the data centre or purchasing computing services. But the economic consequences of increased electricity demand can extend beyond the AI provider.
Training and inference affect costs in different ways.
A large training cluster may require access to substantial electrical capacity before it can operate. That can involve major upfront investment in grid connections and supporting equipment.
Inference facilities face continuing operating costs that depend on the volume and complexity of requests. Providers must decide whether to recover these costs through subscriptions, usage charges, advertising or other commercial arrangements.
For electricity networks, the central question is whether infrastructure is being expanded to meet demand that will actually materialise.
If large investments are made to accommodate anticipated AI loads, the allocation of those costs matters. Depending on local regulation and commercial agreements, some costs may be borne by data centre operators, while others may be distributed among electricity customers.
The IEA’s 2026 assessment warns that poorly aligned investment in data centres and electricity infrastructure could increase electricity prices in some locations. It also highlights the need for better demand forecasts and closer cooperation between technology companies and electricity system operators. IEA
There is a related commercial risk. If demand for AI services falls short of expectations, expensive computing and electricity infrastructure may be underused. If demand grows faster than expected, companies may face capacity shortages and higher operating costs.
Neither outcome is inevitable, but both are reasons to distinguish genuine electricity consumption from speculative future capacity requirements.
What should be measured to understand AI’s real electricity cost?
A single headline figure, such as electricity per AI query or the number of chips in a data centre, is insufficient to describe AI electricity demand.
A more useful assessment needs to distinguish four measurements.
Electricity per task indicates how efficiently a system performs a particular activity. It is most useful when comparing similar workloads under comparable conditions.
Total electricity consumption shows the cumulative demand created by training, inference and supporting operations. It is the measure most relevant to annual electricity generation requirements.
Peak power demand shows how much electricity infrastructure must be capable of supplying at a particular moment. This matters for grid connections, transformers and electrical reliability.
Electricity per useful outcome connects energy consumption to what the system actually achieves. It helps distinguish necessary computation from repeated or unnecessarily intensive processing.
These measurements should be accompanied by information about model type, task length, hardware, utilisation and the boundaries of the calculation.
Independent comparisons remain difficult because detailed production measurements are often commercially sensitive. Public benchmarks such as the[ML.ENERGY Leaderboard]ml.energyThe ML.ENER G Y LeaderboardThe ML.ENER G Y Leaderboard are improving visibility into the energy performance of different models and hardware configurations, but benchmark results do not automatically represent the complete electricity footprint of a commercial service. ml.energy
For governments and electricity system planners, disclosure of actual and expected power demand is particularly important. For businesses choosing AI services, electricity per successful task may be more informative than a generic estimate per prompt.
For the public, both perspectives are necessary. An AI system can be highly efficient per request while contributing to rising national electricity demand, just as a large training project can have a substantial immediate power requirement without accounting for most of the industry’s long-term energy consumption.
The decisive question is not whether training or inference is inherently the bigger problem. It is how much computation AI systems perform, how efficiently they perform it and how quickly total demand grows relative to those efficiency improvements. Those factors determine whether advances in AI computing ease pressure on electricity infrastructure or simply enable a larger wave of electricity consumption.
Amazon book picks
Further Reading
Books and field guides related to How Much Electricity Does Running AI Actually Use?. Use these as the next step if you want deeper reading beyond the article.
Mastering Sovereign Artificial Intelligence
Data, like oil or minerals, is a vital resource – and who owns it is critical, especially as it fuels the artificial intelligence (AI)...
eBay marketplace picks
Marketplace Samples
Live-tested eBay searches with available results related to this page.
Selected fromartificial intelligence poster oneBay.co.uk.
Current eBay listing
A.I. Artificial Intelligence Blu-ray with A3 Travel Poster new and sealed
Current eBay listing
AI ARTIFICIAL INTELLIGENCE . 2001 ORIGINAL MOVIE POSTER vintage 24 YEARS OLD
</article>
<article class="fr-book-card" data-ebay-listing-card data-ebay-card-kind="listing" data-ebay-card-position="4"data-offer-freshness="current" data-offer-observed-at="2026-10-26T20:24:24Z">
</article>
References
- Measuring the environmental impact of delivering AI at Google Scale ml.energy GY Leaderboard(https://ml.energy/leaderboard/)
Endnotes
1.
Source: sciencedirect.com
Link:https://www.sciencedirect.com/science/article/pii/S2542435126001145
2.
Source: iea.org
Link:https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary%C2%A0
3.
Source: epoch.ai
Title: AI How much does it cost to train frontier AI models?
Link:https://epoch.ai/publications/how-much-does-it-cost-to-train-frontier-ai-models
4.
Source: epochai.substack.com
Title: Projecting AI Training Power Demand
Link:https://epochai.substack.com/p/power-demands-of-frontier-ai-training
5.
Source: hai.stanford.edu
Title: ai index report 2026
Link:https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf
Source snippet
RESEARCH AND DEVELOPMENT April 13, 2026...
Published: April 13, 2026
6.
Source: services.google.com
Title: Services Measuring the environmental impact of delivering AI at Google Scale
Link:https://services.google.com/fh/files/misc/measuring_the_environmental_impact_of_delivering_ai_at_google_scale.pdf
7.
Source: ml.energy
Title: The ML.ENER G Y Leaderboard
Link:https://ml.energy/leaderboard/
Additional References
8.
Source: arxiv.org
Link:https://arxiv.org/abs/2609.33965
9.
Source: arxiv.t3mt.com
Link:https://arxiv.t3mt.com/html/2610.14389v1
10.
Source: microsoft.com
Title: www.microsoft.com Scaling AI with 8 to 20x energy efficiency
Link:https://www.microsoft.com/en-us/microsoft-cloud/blog/2026/06/15/scaling-ai-with-8-to-20x-energy-efficiency/
Source snippet
The Microsoft Cloud Blog June 15, 2026...
Published: June 15, 2026
11.
Source: zenodo.org
Title: Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute
Link:https://zenodo.org/records/20616458
Source snippet
Zenodo June 9, 2026...
Published: June 9, 2026
12.
Source: mailarchive.ietf.org
Title: [sustain] Energy use of AI inference, efficiency pathways, and test-time scaling
Link:https://mailarchive.ietf.org/arch/msg/sustain/zg9MfUwL6SZpMbuhlvo0daHlE_M/
13.
Source: youtube.com
Link:https://www.youtube.com/watch?v=8M3MciEmC5c
Source snippet
[https://www.youtube.com/watch?v=8M3MciEmC5c](https://www.youtube.com/watch?v=8M3MciEmC5c)...
14.
Source: youtube.com
Link:https://www.youtube.com/watch?v=WNt_1bSODIo
15.
Source: youtube.com
Link:https://www.youtube.com/watch?v=XtT5i0ZeHHE
16.
Source: youtube.com
Link:https://www.youtube.com/watch?v=ckoi0RTEgcY
17.
Source: youtube.com
Link:https://www.youtube.com/watch?v=cz3AYYZBiGs