On April 18th, it became apparent that leading the race in artificial intelligence doesn't come cheap. Demis Hassabis, head of Google's AI division, revealed at this week's TED conference that Google's investment in AI research may have surpassed $100 billion.
Hassabis, currently in charge of Google's research outfit DeepMind, stands as a key figure in Alphabet's AI endeavors. When questioned about AI competition during the conference, he dropped this staggering figure.
Last month, reports surfaced that Microsoft and OpenAI were planning to invest $100 billion to construct a supercomputer dubbed "Stargate," equipped with "millions of dedicated server chips" to bolster OpenAI's AI models.
During the TED conference, when asked about rival supercomputers and their costs, Hassabis hinted that Google's expenditure might exceed that sum: "We don't disclose specific figures, but I think over time, our total investment will surpass that amount."
The rapid advancement of artificial intelligence has triggered a massive wave of investment. According to Crunchbase data, AI startups raised nearly $50 billion last year. Hassabis's remarks suggest that the cost of vying for dominance in the AI field will only increase.
In this competition, companies like Google, Microsoft, and OpenAI aspire to be the first to announce the development of general AI.
However, the idea of a company pouring over $100 billion into what is considered an overhyped technology still comes as a shock.
The primary uses of these investments merit contemplation. Firstly, a significant portion of R&D costs will go toward chips.
For companies dedicated to developing more advanced AI systems, chips represent a major expense. Simply put, more chips equate to greater computational power, enabling the training of AI models on larger datasets.
Companies like Google and OpenAI, developing large language models, rely heavily on third-party chip suppliers like NVIDIA, but they're also striving to develop their own chips.
The cost of model training is also on the rise.
Stanford University's annual AI Index report released this week indicates that the cost of training state-of-the-art AI models has reached unprecedented levels.
The report states that the training cost for OpenAI's GPT-4 model is estimated at $78 million, compared to $43 million for training GPT-3 in 2020. Meanwhile, Google's Gemini Ultra project incurred training costs as high as $191 million. In 2017, the original technical training cost of an AI model was around $900.
The report also highlights that the training cost of AI models is directly related to computational demands, hence, if achieving general AI is the ultimate goal, costs will continue to spiral upwards.
