See this visualization first on the Voronoi app.
Use This Visualization
Charted: The Surging Cost of Training AI Models
This was originally posted on our Voronoi app. Download the app for free on iOS or Android and discover incredible data-driven charts from a variety of trusted sources.
Key Takeaways
- While detailed information on AI model training costs is limited, it is estimated that Google spent $192 million on Gemini 1.0 Ultra—the highest across leading models.
- These costs were calculated based on cloud compute rental prices, which can rake up millions of dollars quickly since companies rent out thousands of supercomputers that run nonstop for weeks.
- By contrast, the DeepSeek-V3 reportedly cost $6 million to train, however this cost remains disputed.
Today’s state-of-the-art AI models can cost $100 million or more to train.
Yet as companies pour millions into improving model performance, the escalating costs are raising serious questions in the industry. DeepSeek, a new competitor, cited training costs of just $6 million. Meanwhile, an s1 model from Stanford and the University of Washington cost just $6 to train.
This visualization is part of Visual Capitalist’s AI Week, sponsored by Terzo, shows the dollar cost of training AI models, based on analysis from the 2025 AI Index Report.
Training AI Models is Not Cheap
Below, we show the estimated cost for leading models—from OpenAI’s GPT-4 to xAI’s Grok-2:
| Year of Release | Model | Maker | Training Cost(Inflation-Adjusted) |
|---|---|---|---|
| 2023 | GPT-4 | OpenAI | $79M |
| 2023 | PaLM 2 | $29M | |
| 2023 | Llama 2-70B | Meta | $3M |
| 2023 | Gemini 1.0 Ultra | $192M | |
| 2024 | Mistral Large | Mistral | $41M |
| 2024 | Llama 3.1-405B | Meta | $170M |
| 2024 | Grok-2 | xAI | $107M |
As we can see, OpenAI’s GPT-4 cost $79 million, using models containing artificial neural networks that guess the sequence of words in a string of text.
OpenAI has since released new models o1, o3, and o4-mini, which use a “test-time compute” strategy. This means that the longer the model thinks about an answer, the better the answer it spews out. Today, OpenAI charges $200 per month for a pro o1 subscription, which is reportedly running at a net loss given the scale of queries exceeding the compute costs budgeted to run them.
With a $192 million price tag, Google’s Gemini 1.0 Ultra cost more than several major models. For Gemini Ultra, a large chunk of the costs were for research and development staff salaries (including equity)—making up to 49% of the final cost. Meanwhile, AI accelerator chips accounted for 23% of the total cost, followed by 15% for other server components.
Meanwhile, the Grok-2 model from xAI can answer queries on current events in real-time using data from X. Overall, the model cost $107 million to build, which is now integrated into the Grok AI chatbot on X.
To dive into all the AI Week content, visit our AI content hub, brought to you by Terzo.
Learn More on the Voronoi App
To learn more about this topic from a user perspective, check out this graphic on the most popular AI tools in 2025.