Google Could Spend $100 Billion Going All-In on AI

BY KIT EATON @KITEATON

The ongoing explosion of increasingly sophisticated AI systems is driven by vast amounts of hardware sitting in giant, faraway data centers, crammed with endless rows of servers running AI chips.

Google DeepMind CEO Demis Hassabis, mindful that current AI tech was costly to create and maintain, told a conference audience that developing the next generation of AI tech could cost Google more than $100 billion to achieve it.

Hassabis unveiled that number at the start of this year's TED Conference, Business Insider reports. He mentioned the size of Google's AI investment while respondIing to a question about what Google's rivals were spending--with mention of the rumored giant "Stargate" computer that OpenAI and Microsoft are said to be building. "We don't talk about our specific numbers," Hassabis said, but added that he thinks Google will be investing more than the estimated tenth of a $1 trillion cost of Stargate. 

DeepMind, started off as an AI research laboratory cofounded in 2010 by Hassabis. After growing thanks to investments, from entrepreneurs like Peter Thiel and Elon Musk, Google saw the potential of the company and acquired it in 2014 for an undisclosed sum somewhere above $400 million. Hassabis, explaining his notion about the acquisition, underlined that developing better AI really is about more computer power.

"We knew that in order to get to AGI we would need a lot of compute," Hassabis said, referencing the concept of Artificial General Intelligence--what the average person may think of as "full" AI, the kind of system seen in many a sci-fi movie, capable going beyond the limits of current AI tech. Google, Hassabis pointed out simply, "had and still has the most computers." 

The sheer cost of the equipment needed to build AIs is perhaps the main part of Google's investment. Estimates suggest that for the development efforts behind the current cutting edge GPT4 AI algorithms for OpenAI's ChatGPT chatbot, the company needed over 10,000 H100 AI processing units from market-leading AI chipmaker Nvidia. The massive demand for these chips has created a shortage, driving costs per unit per unit up to $40,000 apiece. Elon Musk, who is also developing AI systems as part of both his X and Tesla companies, estimated OpenAI may need as many as 50,000 H100 units to build Chat GPT 5. To capitalize on the soaring AI market, Nvidia has just revealed its next-gen "Blackwell" chips, which are much more powerful than the H100s.

Compared to the hardware costs, the price for training AI on real data may turn out to be much lower. AIs need vast amounts of information as inputs, to train algorithms to offer intelligent-seeming answers to user questions. There's ongoing controversy--and litigation--over how entities like Google and OpenAI may have infringed on some content owner's intellectual property rights to feed the data-hungry AIs they were developing.

Nevertheless, a few high-profile deals by AI companies to buy training data have already been executed. Apple is thought to have struck a deal for between $25 million and $50 million for access to photo site Shutterstock's image archive, and Google reached a deal with social media site Reddit to train its AIs on Reddit's huge archive of uploaded user posts--for a meager $60 million, according to reports. Though many such deals would be needed to train an AI on paid-for data, the costs don't reach anywhere near the scale of investment Hassbis says Google will need to make.

As all of this news is playing out, another cash figure underlines the massive costs of next-gen AIs, this time it concerns venture capital and smaller companies, rather than AI giants like Google. Andreessen Horowitz, one of the world's best known VC firms, just said it has raised $7.2 billion in a new venture fund. Cofounder Ben Horowitz explained that some $1.25 billion would be dedicated in investments in "infrastructure" which includes AI investments. The investment firm previously stated its intent to invest in future AI projects back in March, when cofounder Marc Andreessen criticized OpenAI for its weak security. 

Photo Credit: Getty Images.