Tech benchmarks are exciting — we love seeing LLMs get smarter, faster, and more versatile with every iteration. But while we celebrate the breakthroughs, it’s important to remember the staggering amount of resources required to train these models. Imagine a city’s worth of power, water, and more going into each cycle of training. That’s the scale we’re talking about here.
As AI becomes more integral to our lives, so does its environmental impact. Training a single leading-edge AI model isn’t just a computational feat; it’s a resource-intensive process that draws on:
- Power🔌
- Water💧
- Carbon Emissions💨
- Land🖼️
Unfortunately, official data on these metrics is often scarce. Companies provide limited transparency about the resource consumption of their AI models, leaving us with gaps in understanding their environmental costs.
This post aims to bridge that gap by estimating resource consumption using publicly available data, industry reports, and third-party analyses. Here’s what I’ll focus on:
- Power Consumption: How much electricity is consumed during the training of leading AI models? This includes GPU power consumption, the number of GPUs used, total GPU usage in hours, server power consumption, the number of servers involved, total training time, the approximate power cost ($/MWh), and the probable source of power (e.g., grid or off-grid).
- Water Usage: How much water is needed to cool the infrastructure? (If data is available.) This includes GPU water consumption, the number of GPUs used, total water usage in liters, and approximate water management cost ($/l).
- Carbon Footprint: What are the greenhouse gas emissions tied to training? (If data is available.)
- Land Consumption: How much physical space is required for data centers? What is the location selected for training? (If data is available.)
The next time you marvel at an AI model’s ability to write, code, or analyze, remember the resources behind it. This isn’t just about environmental consciousness; it’s about making informed decisions as consumers and developers of AI technology. By understanding the hidden costs, we can push for more meaningful and sustainable AI practices.
In the next sections, I’ll break down the numbers, discuss trends, and explore potential solutions for reducing resource consumption in AI. Stay tuned!
For this comparison, I wasn’t able to obtain data on water and land consumption, but I plan to include it in future versions. However, I’ve researched and estimated the power consumption and carbon footprint associated with model training.
.png)
.png)
.png)
The race to train the most powerful AI models has often come with a hefty price tag — both in terms of computational resources and environmental impact.
DeepSeek V3 redefines what’s possible in AI by combining exceptional performance with economical training. It achieved groundbreaking accuracy and outperformed competitors in key benchmarks using just one-tenth of the GPUs required by industry giants like Meta and OpenAI. While Meta and OpenAI spent $2–4 million on power costs alone, DeepSeek managed the same feat for just $450K — a staggering 80% savings.
Mistral NeMo is another trailblazer in efficiency. Trained with only 3,072 NVIDIA H100 GPUs — just 12% of what Meta and OpenAI required — its estimated power cost was similarly low at $470K.
NVIDIA Nemotron strikes a middle ground, employing 6,144 H100 GPUs with training power costs totaling $1.2 million. While more economical than LLaMA 3.1 or GPT-4, it remains less efficient than DeepSeek or Mistral in terms of resource and cost optimization.
The total training budget tells an even more transformative story: DeepSeek V3 came in at $5.576 million, while Mistral NeMo is estimated at $4.42 million — just a fraction of the billions spent by larger players. These models not only set new benchmarks for performance but also prove that the future of AI can be smarter, smaller, and more power-efficient.
As we look to the future of AI, the path forward seems clear: smaller models, smarter resource use, and less environmental impact. Just as past tech revolutions have driven progress through innovation and optimization, it’s time for AI to follow suit. This is a motivating and optimistic shift toward a more sustainable and cost-effective future.
The evolution of AI doesn’t have to be fueled by ever-increasing consumption. It can — and should — be about making AI more accessible, practical, and scalable for a wide range of applications, all while reducing our reliance on excessive resources. DeepSeek V3 is proof that we can have both cutting-edge performance and environmental mindfulness in the same package.
I believe that Smaller, smarter, and more power-efficient models are the key to shaping the next wave of AI advancements. The future looks brighter — and greener — than ever before.
Models compared v1: DeepSeek V3, Llama 3.1, and GPT 4.
Models compared v2: DeepSeek V3, Mistral NeMo, Llama 3.1, and GPT 4.
Models compared v3: NVIDIA Nemotron-4, DeepSeek V3, Mistral NeMo, Llama 3.1, and GPT 4.