Trying to win an AI war is really, really, really expensive.

• AI wars are going to be really, really, really expensive.
• Demis Hassabis, head of Google DeepMind, has suggested that Google will spend more than $100 billion on AI development.
• His predictions show how much money tech giants will need to spend to make AI smarter.

Winning AI wars isn't cheap. Just ask Demis Hassabis.

To kick off this year's TED conference, Google's head of AI gave a reality check to his audience in Vancouver on Monday, offering his best guess at how much the search giant will spend developing AI: 1,000. That's over a billion dollars.

Hassabis, who heads Google's famous research institute DeepMind and is perhaps the most important person at the center of Alphabet's AI plans, revealed the astronomical numbers in response to a question about the nature of the competition.

Microsoft and OpenAI will develop a $100 billion supercomputer called “Stargate” with “millions of specialized server chips” to power the ChatGPT maker's AI, according to a report in The Information last month. He says he is making plans.

Unsurprisingly, when asked about the rival's rumored supercomputer and its cost, Hassabis was quick to point out that Google's spending could exceed that. That's over time. ”

While the generative AI boom has already sparked a huge investment surge (AI startups alone raised about $50 billion last year, according to Crunchbase data), Hassabis' comments are a sign of the race to lead the AI ​​sector. This suggests that it will be more expensive.

This is especially true for companies like Google, Microsoft, and OpenAI. All of these companies are in a fierce battle to emerge as the first to claim the development of artificial general intelligence (AI with capabilities that rival human reasoning and ingenuity).

thick chips

Still, the idea that a single company could spend more than $100 billion on a single technology that some believe may be overhyped is eye-opening. .

It's worth considering where that spending will go. First, a large portion of the development cost will be on the chip.

For companies investing in the race to develop smarter AI, these will be among the most expensive purchases. Simply put, the more chips you have, the more computing power you have and the more data you can train your AI models with.

Companies working on large-scale language models like Google's Gemini and OpenAI's GPT-4 Turbo rely heavily on third-party chips like Nvidia. But they are increasingly trying their own designs.

The general business of training models is also becoming expensive.

Stanford University's annual AI Index report released this week notes that “the cost of training cutting-edge AI models has reached unprecedented levels.”

The magazine noted that $4.3 million was used to train GPT-3 in 2020, while OpenAI's GPT-4 used “an estimated $78 million worth of computing for training.” Meanwhile, Google's Gemini Ultra training costs were $191 million. The cost to train the proprietary technology behind an AI model was approximately $900 in 2017.

The report also notes that costs are likely to rise if AGI is the end goal, as “there is a direct correlation between the cost of training an AI model and its computational requirements.” .