The debate about AI replacing humans in architectural design just got more interesting. I've spent weeks analyzing the energy costs of AI-generated 3D models in architecture, engineering, and construction. The results are mind-blowing - and not in the way you might expect.
If you need more than 2.3 full iterations, human modelers will be less expensive than AI, which means that AI not only is completely wasteful in energy, it will most likely also be completely wasteful in cost compared to the labor side.
🎧 Listen To This AEC_VC Episode
The Raw Numbers: AI Energy Consumption in 3D Modeling
Let's start with some everyday benchmarks to put things in perspective. One hour of Netflix streaming consumes about 0.8 kilowatt hours (kWh). A single load of laundry uses 4.6 kWh. Now, here's where it gets interesting - AI image generation alone consumes nearly 3 kWh for just one image. That's more energy than baking in your oven for an hour. And we're not even talking about 3D models yet.
When we look at 3D modeling for architecture, the numbers become staggering. I analyzed four different reference projects: a 150 square meter single-family home, a six-apartment building, a 10,000 square meter office building, and an international airport. The energy consumption for AI to generate these models scales dramatically with complexity. An international airport model would consume between 700,000 to 1.2 million kilowatt hours - equivalent to running 21.5 million hours of a 40-watt halogen light bulb. Even a modest single-family home would require energy equivalent to 6,500 hours of air conditioning.
To be fair, I had great discussions with several of you (eg. Francesco at Augmenta, Raphael Riad) who also pointed out the potential better energy efficiency of eg. transformer models in AI for 3D. Fair enough. Raphael was kind enough to run some quick tests on his H100 for models with 80-200 polygons inspired by my analysis, with the following results (credits to Raphael Riad):
- H100 can go up to 19,000 tokens per second
- Models are about 1000-3000 tokens
- H100 is rated up to 700 watt usage is around 500
- Models I’ve generated are 80-200 polygons, we can assume 100
- 2000T/model/19000T/s =0.1s/model
- 0.1s/3600s/h=2.77e-5 h/model
- 2.77e-5 h/model*500w/1000w/kw=1.39e-5kwh/model
- 1.39e-5kwh/model/100polys/model=
- 1.38889e-7kwh/polygons
At first sight, this could suggest orders of magnitude lower energy use for extremely simple and light models on a good rig with almost no detailing. However, transformer models’ energy uses scale superlinearly, and 100 polygons is NOTHING in 3D for BIM and AEC. To give you an idea, a super simple single family home will easily have 75’000 to 200’000 polygons, and with superlinear energy use scale on transformer models, it will likely not be the exact number kWh founder above. It is enocuraging though. So we will still need to see similar studies for real-world AEC models.
The Human Brain's Remarkable Efficiency
Here's where the comparison gets fascinating. A trained human modeler using Revit can create about 1,000 polygons per hour, consuming only about 83 kilocalories of energy. When we do the math, humans need just 13 kWh to model a single-family home and around 1,000 kWh for an international airport. Compare that to AI's consumption of hundreds of thousands of kilowatt hours.
This translates to humans being 875 times more energy efficient than AI in creating 3D models. That's not a typo - it's almost three orders of magnitude more efficient. The human brain, evolved over millions of years, remains an incredibly optimized machine for spatial reasoning and design tasks.
The Cost Equation Changes Everything
At first glance, AI appears to have a cost advantage. Using New York City architect salary rates, AI-generated models are about 2.3 times less expensive than human-generated ones when considering only energy costs versus salary. However, this calculation assumes perfect execution - getting the model right on the first try.
The reality is more complex. If you need more than 2.3 iterations to perfect a model (which is common in architectural design), humans become more cost-effective than AI. This doesn't even account for the nuanced understanding and creative problem-solving that human architects bring to the table.
Implications for the Future of Architectural Design
The data points to a clear conclusion: AI-generated 3D architectural models will be most cost-efficient (but far less energy efficient) for simple, repetitive projects like standard residential or small commercial buildings. This makes sense because these projects often have standardized elements and fewer unique details.
For complex, highly detailed projects, AI generation shows no substantial advantage in either cost or energy efficiency beyond very early feasibility studies and low Level of Detail (LOD) designing. The future likely lies in collaboration - using AI to generate initial versions or specific parts of models, with humans handling the fine-tuning and complex design decisions.
You Can Find More Analysis On AEC_VC
Spotify: https://open.spotify.com/show/3fvTj23GIMlAT6CCm4t1lT
Apple: https://podcasts.apple.com/de/podcast/aec-vc/id1740915855
Foundamental: https://www.foundamental.com/
Subscribe to the Newsletter: https://www.linkedin.com/newsletters/aec-vc-7186650851766083584/
Follow The AEC_VC
Patric Hellermann: https://www.linkedin.com/in/aecvc/
Web: https://aecvc.com/
#AIinConstruction #EnergyEfficiency #ConstructionTech #3DModeling #ArchitecturalDesign #Sustainability #FutureOfConstruction #GreenTechnology