Course:ELEC300E/CaseStudies/AIDatacenterImpacts
Context: The Carbon Footprint of AI
| Vancouver's AI Data Centers |
|---|
 |
| Content / Topics |
| AI Data Centers |
| Environmental Impacts |
| Policy |
| Local Issues |
| Relevant Analysis Strategies or Tools |
| Systems Mapping |
| Recognizing Biases & Positionality |
| Recognizing logical fallacies |
| Decision Impact Analysis |
What have you heard about the environmental impacts of AI and AI datacenters throughout the past couple years?
To get us into the spirit of discussion, and to begin thinking about rhetoric, context on the carbon footprint of AI is formulated as an argument.
Should we be concerned about the carbon footprint of AI?
Person A: Not really, the energy used by AI is trivial. A typical LLM response requires only around 0.42Wh - far less than the amount of energy required to boil water for a cup of coffee, which requires around 50Wh.
Person B: But you likely use AI (including AI based search) much more often than you boil water. And what about training? Training a new AI model requires hundreds of megawatt hours of electricity, which corresponds to around 300 tons of C02 emissions. That is the total lifetime emissions of 5-6 cars!
Person A: Training isn't done for every AI query though. That cost is amortized over millions of queries, meaning that the marginal cost of training is very low.
Person B: Maybe the cost of a single query is small, but ChatGPT processes around 2.5 billion prompts each day! That adds up to 383 GWh each year for ChatGPT alone. That amount of electricity could supply 3 million people in Sub Saharan Africa.
Person A: If we are looking at the total global energy requirements of AI, we should compare with total global energy consumption: Datacenters accounted for only 1.5% of total global energy consumption in 2024, and only a fraction of those datacenters are specifically for AI.
Person B: That might be a small percentage, but the energy demands of datacenters are expected to double by 2030, in large part because of AI. If you look at the US specifically, data centres are predicted to account for 12 per cent of energy use by 2028.
Person A: You are only looking at the environmetnal downsides of AI, without considering the potential positive impacts. AI has the potential to significantly improve the efficiency of various industries, and could be leveraged as a tool to decrease emissions.
Person B: Certainly there are possible positive uses for specific types of AI - that doesn't mean that typical uses of GenAI are sustainable, efficient or justifiable. Recent independent researchers "did not find a single example where popular tools such as Google’s Gemini or Microsoft’s Copilot were leading to a “material, verifiable, and substantial” reduction in planet-heating emissions".
Person A: In theory though, all uses of GenAI could be efficient, if the data centers run on renewable energy! Investment in AI companies could fund research to develop more efficient data centers that have minimal carbon emissions and environmental impacts.
Person B: Unfortunately that is not what is currently happening. In the first half of 2026 alone, there was a 76% increase in the construction of new gas-fired power generation projects in the United States, and around half of these projects are directly linked to AI data center requirements. We need to be decreasing our reliance on fossil fuels, not building new gas-fired power generation plants!
Person A: This is a good argument for developing efficient AI data centers in places like Vancouver, where much of our energy comes from renewable sources!
Person B: But is it fair for specific communities to bear the brunt of the energy requirements of AI? 71% of people in British Columbia would oppose a datacenter being constructed in their neighborhood. These increased demands could put strain on the power grid and risk brownouts during heatwaves or snowstorms, and might increase consumer power bills. Would you really want a data center in Vancouver?
That brings us to our case study!
Case: Vancouver's AI Data Centers
The idea of building AI data centers in Vancouver isn't just theoretical: two Vancouver AI data centers are currently planned for construction by Telus. The plans are part of Canada's “Enabling Large-Scale Sovereign AI Data Centres” program, through which the government is investing in Canadian AI infrastructure projects. A 100,000 square foot facility is planned in Mount Pleasant and a 400,000 square foot location near BC Place, two densely populated neighborhoods.
These plans proved controversial, sparking large protests in Vancouver. City councillors Lucy Maloney (OneCity) and Sean Orr (COPE) put forward a motion to pause the plans, assess the impacts of the planned data centers and create regulations on AI data centers. This motion was defeated 7-3 in a city council vote, with Mayor Ken Sim (ABC) and the other ABC councillors voting against the motion.
Recommended Readings and Resources
- 2 page document: Vancouver city council motion to pause data center construction: https://council.vancouver.ca/20260715/documents/cfscmotion5.pdf
- ~3 minute news article reporting on several arguments made primarily in favor of the development: https://betakit.com/steel-concrete-and-code-feds-and-telus-announce-three-ai-data-centres-in-bc/
- ~3 minute opinion piece (Optional) presenting several arguments against the developments: https://betakit.com/steel-concrete-and-code-feds-and-telus-announce-three-ai-data-centres-in-bc/
Discussion Questions
Here are a few questions to help you begin analyzing the issues discussed in this case study. The purpose of these questions is to prompt reflection and further consideration: there are no right or wrong answers!
- Personal response: Take stock of your immediate response to this case study: outrage, discomfort, boredom, confusion, frustration, hope, interest? How might your unique personal experiences and perspectives inform this response? How can you leverage this response in a positive way?
- Contradictory viewpoints and outcomes: As you were reading the discussion between Person A and Person B, and the different sides of the AI data centers debate, which arguments did you find convincing? Did you notice any cognitive biases, assumptions, weak arguments or logical fallacies?
- Positive action: What do you think of Vancouver's AI data centers, after reading through the case study? What concrete actions would you like to see from Vancouver city council?
- Interruptions and invitations: Whether you support the construction of these data centers or are more critical of the plans, how can you advocate for your position and effect change in your community?
- Physical connections: This case study is local! Does your relationship to the city, and your proximity to the proposed projects impact your understanding or perception of the problem?
- Positive practices: What is one concrete action that you could take, or a practice that you could adopt in your personal, academic or professional life to address some of the issues discussed in this case study?
Bonus Personal Reflection Question:
- Social and Historical Context: Engineering is often inherently political, but the political aspects of this case study are very explicit. How does this case study relate to your previous understandings or assumptions about Vancouver municipal politics? Do you have any broader takeaways about the political dimensions of conversations around AI data centers? How do you see your work relating to your personal political views?
Appendix
Just the facts, without the Person A / Person B script.
Individual Scale:
The carbon footprint of a single AI query depends on the type of query (classification, image generation text generation etc), the hardware used to run the model, the model used, and what kind of energy it runs on (renewable, gas powered etc). In any case, the energy required by a single query, and the associated carbon footprint is small:
You might compare this with boiling water for a cup of coffee, which requires ~50Wh.
Note: training an AI model requires hundreds of megawatt hours, but when divided by the total number of times the model is used, the marginal cost of each query is very low.
Global Scale:
The footprint of a single AI query is small, but aggregated over billions of prompts it becomes much more significant. ChatGPT processes around 2.5 billion prompts each day.
ChatGPT energy use per year ~= 2.5 billion x 0.42Wh x 365 = 383 GWh
For comparison, according to a UN report, "This amount of electricity would be enough to meet the annual domestic electricity demand of nearly 3 million people in Sub-Saharan Africa".
When you look at the total energy consumption of AI datacenters compared with total global energy consumption however, the impact of AI again appears small. Datacenters accounted for 1.5% of total global energy consumption in 2024, and while that percentage is increasing significantly (set to double by 2030), it accounts for a small percentage of the global total (47666 TWh).
Note that all electricity consumption is not equal when it comes to environmental impact. The amount of electricity used In the first half of 2026 alone, there was a 76% increase in the construction of new gas-fired power generation projects in the United States. These changes are outlined by the Global Energy Monitor in an independent 2026 analysis of global power generation projects.
Half of this change can be linked to the growing energy requirements of AI datacenters. While the energy requirements required for a single AI inference are small, t. As described in the Guardian:
"The stampede by tech companies, such as Google, Open AI and Amazon, to build datacenters has contributed to a backlog of the most efficient gas turbines used for accompanying power plants. This has led several companies, including Elon Musk’s xAI, to quickly switch to smaller, less efficient turbines which are more heavily polluting."
The energy demands of AI are also highly localized, meaning that even if the total global energy consumption of AI datacenters is a small percentage of a country's energy use, they could cause severe strain on power grids at a local level.