The training of AI models requires millions of dollars. It is not simply a matter of training AI on a large number of datasets alone. Better chips, expert staff and research are an important part of the whole process. These factors take the total cost of training AI models up to millions of dollars.
The recent 2025 AI Index Report by HAI has revealed the cost of training major AI models owned by major tech companies. Some of these models cost more than 100 million dollars. So what appears to be a simple AI model is actually a million-dollar project.
As per the report, DeepSeek from China and Llama 2-70B from Meta cost $6 million and $3 million for training them, respectively. Though still expensive, they are relatively cheaper when compared to other AI models from major companies.
PaLM 2 model from Google cost $29 million, and Mistral Large from Mistral cost around $41 million. Their cost is high, but they are still nowhere near the most expensive AI models being used today.
More complex models like OpenAI's ChatGPT-4 that uses artificial neural networks to guess sequences of words cost $79 million. Its new models like o1 and o3 give even better answers to users due to their test-time compute strategy. But the better version is even costlier. Its Pro o1 monthly subscription is available for $200.
If I talk about the costliest versions available, Grok-2 from xAI cost around $107 million. Similarly, Llama 3.1-405B from Meta and Gemini 1.0 ultra from Google cost $170 million and $192 million respectively.
These are pretty huge numbers, and this is just the initial stage of AI models. It is not difficult to predict that future AI versions will cost companies more due to costlier chips and research. For users, it simply means higher subscription rates.
Read next: Which Countries Are Allocating the Highest Percentage of GDP to Research and Development Investment?
The recent 2025 AI Index Report by HAI has revealed the cost of training major AI models owned by major tech companies. Some of these models cost more than 100 million dollars. So what appears to be a simple AI model is actually a million-dollar project.
As per the report, DeepSeek from China and Llama 2-70B from Meta cost $6 million and $3 million for training them, respectively. Though still expensive, they are relatively cheaper when compared to other AI models from major companies.
PaLM 2 model from Google cost $29 million, and Mistral Large from Mistral cost around $41 million. Their cost is high, but they are still nowhere near the most expensive AI models being used today.
More complex models like OpenAI's ChatGPT-4 that uses artificial neural networks to guess sequences of words cost $79 million. Its new models like o1 and o3 give even better answers to users due to their test-time compute strategy. But the better version is even costlier. Its Pro o1 monthly subscription is available for $200.
If I talk about the costliest versions available, Grok-2 from xAI cost around $107 million. Similarly, Llama 3.1-405B from Meta and Gemini 1.0 ultra from Google cost $170 million and $192 million respectively.
These are pretty huge numbers, and this is just the initial stage of AI models. It is not difficult to predict that future AI versions will cost companies more due to costlier chips and research. For users, it simply means higher subscription rates.
Read next: Which Countries Are Allocating the Highest Percentage of GDP to Research and Development Investment?
