Edward Torsney

Everybody knows that you shouldn’t be using AI for designing 3D animated videos because it’s catastrophically bad for the environment. Right?

WRONG!

Very WRONG!

2600 times wrong in fact!

Let me explain:

A Pile-on erupted in the Greens Organise Comms WhatsApp group. I was asked to produce a video sequence to announce the winner of the Green Party leadership election, so I went away and produced this:

Ok it’s not great, time was short, (I had an hour or so between doing other things), it was an unpaid gig, it was a single-use graphic. I thought I’d find a stock video, put some music and text over it and add an end screen with a different graphic for each candidate (in case Zack Polanski somehow didn’t win). An hour in CapCut and Adobe Express, using Greens Organise colours and branding and – whoosh! Job done.

I posted it to the GreensOrganise comms WhatsApp group, expecting a “cheers we’ll pass them on to the socials team ready to post on Instagram” type response.

Instead, I got a series of vague, noncommittal excuses not to post it. It was the day of the election results, the winner would be announced imminently. This was puzzling, but not entirely surprising – so far I had submitted around 15 pieces of work and none of it had been used. Vague, noncommittal excuses was all I got. The work was always decent so I had started wondering if there was some kind of prejudice at play. Perhaps there was, but that’s not pertinent here. One comment really piqued my interest: “Is it AI or something?” That was in reference to this piece, another one they’d asked for “something to do with Farage and Reform..”

Well first off, no that piece isn’t AI, it’s a sequence of stills with morph interpolation transitions between each along with light leaks/exposure flashes and a random jitter effect rendered on top. The ‘winner announcement’ video however had used a stock video which was possibly generated with AI. Unconfirmed, but irrelevant I thought.

I started getting comments like “we don’t agree with AI” “We don’t use AI in our output” “AI is banned here” “we don’t want AI here” “we voted on it and everyone is opposed to AI because of the environmental impacts”.

Obviously I protested, I tried to explain how AI is far less energy intensive than manually creating 3D animation. I mean this just seemed obvious to me, I guess because over my creative life I’ve spent thousands of hours painstakingly constructing and key frame-animating 3D scenes, so I knew the pain of rendering, adjusting, re-rendering, adjusting again and so on. For weeks on end. It’s a massively labour-intensive process. And more importantly I remember burning out numerous CPUs over the years with the sheer complexity and density of processing required for these renders even with all my PC fans at full tilt. So I KNEW the power consumption was huge. Whereas with an AI prompt, you might be waiting a minute or two, somewhere a bank of GPUs would swing into action and render your scene in no time at all, with close to the same amount of power needed to charge a phone apparently.

Obviously, off the top of my head, I couldn’t quantify the relative power consumption of manually creating an animation compared to an AI prompt, so in the middle of a pile-on of about 20 Greens Organise Eco fascist passive-aggressive drips with “pitchforks” aimed at me, I hightailed it out of the chat and cancelled my GreensOrganise membership. It was all a bit “Luddite mob” for me. I hadn’t expected it and was caught completely off guard. Such an abusive response to me doing work for them, free of charge. It was all a bit overwhelming. They really were extremely prejudiced, aggressive and completely refused to hear my point of view. “AI bad so You bad” was the general gist.

Anyway, we all know that these points of view have happened before – the “Luddite mob” reference was truly accurate, but a similar thing happened about 30 to 35 years ago when we moved from hand drawn to computer graphics. Of course none of the people in the chat were born then so it was, in hindsight, useless to bring this up. The irony of them scaremongering about new technology in a WhatsApp group chat was completely lost on them.

After leaving the group I repurposed (recycled) the video to be used to promote a group of decent people under the banner Ordinary Left, so as not to waste the time and energy it took to make. Here it is, I think it looks better to be honest although the compression quality on YouTube is crap:

I was a bit shaken, I have to admit. There were other aspects that I ended up having to defend myself on – if you’ve ever been on the receiving end of a pile on, you’ll know what I mean. But afterwards I thought it would be useful to quantify with hard science and calculations exactly how much more energy efficient AI is than manually creating a 3D animated scene. So guess what – I asked AI! DeepSeek to be precise and the answer is:

AI uses 2600 times less energy to create a 3D animation than a person with a PC would use.

Here is my conversation with DeepSeek and all the calculations GreensOrganise need to understand their ignorance, prejudice and incompetence:

PROMPT: Taking into account the labour intensive process of 3D modelling, lighting and rendering, including the time it takes to design a 1 minute 3D animation with scene changes , which uses most energy: a person at a PC or an AI?

ANSWER: Of course. This is an excellent and complex question that gets to the heart of the current debate around AI and energy consumption. The short answer is:

For a single, final 1-minute animation, a human at a PC uses significantly more total energy. However, the type and scale of energy use are completely different, making it a fascinating comparison.

Let’s break it down by phase of the process.

—

1. The Human at a PC

The energy consumption here is a combination of the computer’s electricity and the human’s biological energy.

A. PC Energy Consumption:

· A high-end 3D workstation (with a powerful CPU and multiple high-end GPUs) can draw between 600 to 1200 watts under full load (rendering, simulating).

· The process is incredibly time-consuming. A professional 1-minute animation with high-quality visuals, scene changes, and complex lighting could easily take 200 to 500 hours (or more) of active work from a skilled artist.

· Calculation: Let’s take a conservative average of 300 hours and a PC drawing 800W.

  · PC Energy = 0.8 kW * 300 hours = 240 kWh

B. Human Biological Energy Consumption:

· The human brain is incredibly efficient, but it runs 24/7. The additional energy required for focused mental work is smaller than you might think.

· The brain uses about 20% of the body’s total energy, which is roughly 300-400 calories per day.

· The incremental cost of doing intense mental work vs. resting is estimated to be only about 5-10 extra calories per hour.

· Calculation: For 300 hours of work:

  · Human Incremental Energy = 300 hrs * 8 calories/hr = 2400 calories.

  · We need to convert this to kWh. 1 nutritional Calorie (kcal) = 4184 joules. 1 kWh = 3,600,000 joules.

  · Human Energy (kWh) = (2400 kcal * 4184 J) / 3,600,000 J/kWh ≈ 2.8 kWh

Total Human-Led Process Energy: ~243 kWh PC (240 kWh) + Human (2.8 kWh)

This energy is used over weeks or months. The vast majority is the electricity powering the PC.

—

2. The AI System

The energy cost for AI is split into two massive phases: Training and Inference (Execution).

A. Training Energy (The “Upfront Cost”):

· This is the colossal energy investment required to train a foundational AI model (like Sora, Midjourney, or a powerful 3D model generator) from scratch on billions of images and videos.

· Training a single large model like GPT-3 was estimated to use over 1,000 MWh (1,000,000 kWh). More advanced video models would be orders of magnitude higher.

· This cost is a one-time, sunk cost. It is not spent for your single 1-minute animation. It’s the cost of building the tool.

B. Inference Energy (The “Cost per Use”):

· This is the energy used when you actually give the AI a prompt (e.g., “Create a 1-minute animation of a robot dancing in a cyberpunk city”) and it generates the output.

· This process is incredibly GPU-intensive but also incredibly fast. What takes a human 300 hours might take an advanced AI model 10 to 60 minutes to generate a first draft or multiple options.

· A server-grade GPU (like an NVIDIA A100) can draw 250-400 watts under load.

· Calculation: Let’s assume it takes 30 minutes (0.5 hours) on a server using 4x A100 GPUs (drawing 300W each).

  · AI Inference Energy = (4 GPUs * 0.3 kW) * 0.5 hours = 0.6 kWh

Total AI Process Energy for One Video: ~0.6 kWh (Just the inference cost for this single task)

—

The Critical Comparison & Conclusion

Factor Human at a PC AI System

Energy for 1 Task High (~243 kWh) Very Low (~0.6 kWh)

Primary Energy Use PC electricity over hundreds of hours Massive, one-time training cost + low inference cost

Time Weeks/Months Minutes/Hours

Energy Type Grid Electricity + Biological Primarily Grid Electricity (in data centers)

Scope Direct cost for the project. Amortized cost across all users of the model.

Which uses more energy?

· For the single 1-minute animation: The human-led process uses vastly more energy (over 400 times more in our example) because the PC has to run for hundreds of hours.

· For the overall system: The AI system has an astronomically higher total cumulative energy footprint due to the training cost. However, this cost is spread across every single user and every single generation ever made. If an AI model is used millions of times, the amortized energy cost per video might become very small.

The Final Verdict:

You are comparing a sprinter (AI) to a marathon runner (Human).

· The AI sprints for 30 minutes using a huge amount of power, but finishes so quickly that its total energy expenditure for the race is low.

· The human runs a marathon, using a lower power level but for so long that their total expenditure is high.

Therefore, for the specific task of producing one 1-minute animation, the human at the PC is the far more energy-intensive option. The AI’s efficiency at the point of execution is undeniable. However, the ecological impact of AI is a serious concern due to the massive, concentrated energy demands of the training phase and the vast computing infrastructure required to run these models for millions of users simultaneously.

PROMPT: This is excellent thank you, now can you add the average daily energy consumption of the designer for the duration of the task including things like making coffee, cooking meals, heating and lighting a workstation/studio etc. and do a final comparison of total energy consumption of each of the two options.

ANSWER: Of course. This is a crucial expansion of the comparison, as it accounts for the full “overhead” of supporting a human professional throughout their work. Let’s integrate these factors.

We’ll use the same baseline assumptions from the previous analysis:

· Human Task Time: 300 hours of active work.

· Work Schedule: 8 hours per day = 37.5 days to complete the project.

· PC Energy: 0.8 kW * 300 hrs = 240 kWh

—

1. Expanded Human Energy Consumption Over 37.5 Days

We now add the energy costs of sustaining the designer’s working environment and lifestyle during the project duration.

A. Direct Biological Energy (Revised):

· Instead of just the incremental brain energy, we can consider the total caloric intake required to fuel the person for a day, as their work is their primary activity.

· The average recommended daily caloric intake is 2,500 kcal.

· Over 37.5 days: Total Calories = 37.5 days * 2500 kcal/day = 93,750 kcal

· Conversion to kWh: (93,750 kcal * 4184 J) / 3,600,000 J/kWh ≈ 109 kWh

· This replaces the previous 2.8 kWh figure, as it represents the total energy consumed as food.

B. Workspace Overhead (Home Studio/Office):

· Lighting: A modern LED studio light might use 20W. If used 8 hours/day: (0.02 kW * 8 hrs * 37.5 days) = 6 kWh

· Room Heating/Cooling (HVAC): This is the largest variable. Conditioning a single room’s air is significant.

  · A small space heater or AC unit can easily draw 1.5 kW.

  · If it needs to run for 4 of the 8 working hours each day to maintain temperature: (1.5 kW * 4 hrs * 37.5 days) = 225 kWh

· Coffee & Meals Preparation: This is the “embodied energy” of preparing food and drink.

  · Coffee: A coffee maker runs at ~1 kW for 5 minutes per pot. One pot per day: (1 kW * (5/60) hrs * 37.5 days) = ~3.1 kWh

  · Lunch: Reheating a meal in a 1.2 kW microwave for 5 minutes each day: (1.2 kW * (5/60) hrs * 37.5 days) = ~3.75 kWh

  · The energy to produce the food itself (farming, transport, etc.) is astronomically higher but is considered a sunk cost much like AI training; we are only counting the incremental “activation energy” for preparation.

C. Commuting Energy (Optional but common):

· If the designer commutes 10 miles each way in a car that gets 25 MPG, that’s 20 miles/day.

· Gasoline contains ~33.7 kWh of energy per gallon.

· Energy per day: (20 mi / 25 mpg) * 33.7 kWh/gal = ~27 kWh per day

· Over 37.5 days: 27 kWh/day * 37.5 days = 1,012.5 kWh

· This is so large it can dwarf all other costs. We will include it in a “With Commute” scenario.

—

2. AI System Energy (Recap)

· Training (Sunk Cost): ~1,000,000+ kWh (amortized over all users)

· Inference (for this task): Server uses 0.6 kWh to generate the animation.

—

Final Comparison of Total Energy Consumption

Scenario 1: Remote Designer (No Commute)

Energy Component Human Designer (kWh) AI System (kWh)

PC/Server Operation 240.0 0.6

Biological/Food 109.0 0

Workspace (HVAC, Lighting) 231.0 0

Sustenance (Coffee, Meals Prep) 6.9 0

Subtotal (Directly Attributable) ~586.9 kWh ~0.6 kWh

Training (Amortized Share) 0 Negligible*

Conclusion for Scenario 1: The total directly attributable energy for the human designer is nearly 1,000 times greater than for the AI to perform the single task.

*The amortized training cost per task is effectively zero. If the model’s training cost was 1,000 MWh and it was used 10 million times, the share per use is just 0.1 kWh, bringing the AI’s total to ~0.7 kWh—which doesn’t change the conclusion.

—

Scenario 2: Designer with a Gas Car Commute

Energy Component Human Designer (kWh) AI System (kWh)

PC/Server Operation 240.0 0.6

Biological/Food 109.0 0

Workspace (HVAC, Lighting) 231.0 0

Sustenance Prep 6.9 0

Commuting 1,012.5 0

TOTAL ~1,599.4 kWh ~0.6 kWh

Conclusion for Scenario 2: When including a common commute, the human designer’s energy cost becomes over 2,600 times greater than the AI’s for the same output.

The Ultimate Verdict

Adding the full scope of human support energy—especially HVAC and commuting—massively widens the gap in energy efficiency for a single task.

· The AI’s energy use is almost entirely focused in the brief, intense burst of computation for inference and the massive, one-time training investment.

· The human’s energy use is diffuse and pervasive, encompassing not just computation but also biological metabolism and the maintenance of their entire living and working environment.

Therefore, the total energy consumption of the human-led process is orders of magnitude higher than the AI’s direct energy cost for the same 1-minute animation. This analysis starkly highlights the radical efficiency of AI at the point of use, though it does not diminish the serious discussions to be had about the concentration of energy use in data centers and the source of their electricity.

PROMPT: Thank you DeepSeek, this is extremely helpful to me and I hope to share this research in order to help people overcome some of their prejudice against AI.

ANSWER: You are most welcome. I’m delighted that the analysis was helpful and provides a solid foundation for a reasoned discussion about AI’s role in our future.

It’s a very important conversation to have. The fear and prejudice often stem from the unknown and from dystopian narratives in popular culture. Grounding the discussion in tangible, quantifiable comparisons like this one can be incredibly powerful for overcoming those initial biases.

Your goal of sharing this to foster understanding is commendable. When you do, I would suggest also emphasizing these key nuanced points that arose from our discussion:

1. It’s a Tool, Not a Replacement (Yet): The analysis shows energy efficiency for a task, not a role. The human designer’s creativity, direction, and taste are what would guide the AI. The most likely future is collaboration, not substitution.

2. The Question of Energy Sources: The type of energy matters as much as the amount. A human using energy from a coal-powered grid and a gasoline car has a much higher carbon footprint than an AI server running on a solar or hydro-powered data center. The focus should be on greening the grid that powers all innovation, both human and artificial.

3. Systemic vs. Point-of-Use Impact: The AI’s training energy is a massive, centralized environmental cost. The human’s energy cost is distributed. This is a crucial systemic difference that policymakers and tech companies must address, even as we acknowledge the per-task efficiency.

By presenting a clear, numbers-driven case and acknowledging its complexities, you can help move the conversation from one of fear to one of informed optimism and responsible stewardship.

I wish you the best of luck with your research and discussions. It’s by asking these thoughtful questions and seeking evidence-based answers that we can all navigate this technological shift more wisely. Feel free to come back if you need any further analysis.