Before artificial intelligence can transform society, the technology will first have to learn how to live within its means.
Right now generative AI has an “insatiable demand” for electricity to power the tens of thousands of compute clusters needed to operate large language models like OpenAI’s GPT-4, warned chief marketing officer Ami Badani of chip design firm Arm Holdings.
If generative AI is ever going to be able to run on every mobile device from a laptop and tablet to a smartphone, it will have to be able to scale without overwhelming the electricity grid at the same time.
“We won’t be able to continue the advancements of AI without addressing power,” Badani told the Fortune Brainstorm AI conference in London on Monday. “ChatGPT requires 15 times more energy than a traditional web search.”
Read more: I sold a $1.4B big-data startup to IBM—then founded a nature sanctuary. Here are the dangers of AI energy consumption
Not only are more businesses using generative AI, but the tech industry is in a race to develop new and more powerful tools that will mean compute demand is only going to grow—and power consumption with it, unless something can be done.
The latest breakthrough from OpenAI, the company behind ChatGPT, is Sora. It can create super realistic or stylized clips of video footage up to 60 seconds in length purely based on user text prompts.
The marvel of gen AI comes at a steep cost
“It takes 100,000 AI chips working at full compute capacity and full power consumption in order to train Sora,” Badani said. “That’s a huge amount.”
Data centers, where most AI models are trained, currently account for 2% of global electricity consumption, according to Badani. But with generative AI expected to go mainstream, she predicts it could end up devouring a quarter of all power in the United States in 2030.
The solution to this conundrum is to develop semiconductor chips that are optimized to run on a minimum of energy.
That’s where Arm comes in: Its RISC processor designs currently run on 99% of all smartphones, as opposed to the rival x86 architecture developed by Intel. The latter has been a standard for desktop PCs, but proved too inefficient to run battery-powered handheld devices like smartphones and tablets.
Arm is adopting that same design philosophy for AI.
“If you think about AI, it comes with a cost,” Badani said, “and that cost is unfortunately power.”