Google reportedly allocated $191 million to Gemini Ultra training last year. That’s more than double the estimated cost of GPT-4 from OpenAI.

Google allocated $191 million to training its Gemini Ultra model last year. The analysts at Stocklytics know this. The investment is huge, especially considering OpenAI developed GPT-4 for less than half that budget. Note that training GPT-4 also comes at a huge cost, with an estimated investment of $78 million in 2023.

The numbers show how expensive it is to build and train basic models. They also make it clear that Google is not content with second place and is not afraid to invest heavily in AI. To keep pace with GPT-4, the company committed significant financial resources.

Not for small shrimp

Building large basic generative AI models is a very inaccessible practice given the price, reserved for the largest tech companies with access to GPU supercomputers and a strong war chest. Smaller players can start working with such models, innovating on them or, at best, developing smaller alternatives. Given the costs, a real competitor to the generative AI models of Google, OpenAI and Facebook will not suddenly come from a small start-up.

Hugging Face recently announced its intention to strengthen the role of smaller companies. To this end, the company is making $10 million of GPU computing power available to be used by small developers, startups and academics to develop their own solutions.