The chips needed to run AI models have a typical lifespan of one to three years. When these NPUs are under heavy load, they can burn out in just 54 days. Consequently, the entire AI industry must constantly produce new chips simply to maintain its current capacity, a requirement that hinges on access to rare earths.
How is this sustainable in the long term? Even for a country that dominates the rare earth supply chain, eventually the rare earth deposits will be exhausted, no? Same problem with the rapid expansion of electric vehicles and renewables: batteries and solar panels have significantly shorter life spans than fossil fuel engines and turbines in power plants, and rely on rarer materials than just steel and copper. Therefore one assumes that a priority should be research and investment into recycling technology, as China is doing for batteries and solar panels.
Also, what do the logistics look like of data centers with millions of chips if chips have to constantly be replaced? At some point even this replacement process will necessarily have to be automated, else human technicians will not be able to keep up with the 24/7 replacement. Or is replacement done in bulk rather than on a per chip basis as they burn out?
12
☆ Yσɠƚԋσʂ ☆ - 1mon
You can recycle them, but yeah this is definitely a problem. Although, we're already in that scenario with other kinds of data centres too. The servers are designed so you can replace things on the fly, but you still have to constantly make chips. With AI, it's more of a problem cause chips tend to run on higher loads more of the time. So, making more efficient models would be one way to address that.
yogthos in technology
How China's State-Led Strategy Outmaneuvered Market-Driven Tech
https://dialecticaldispatches.substack.com/p/the-planning-advantageHow is this sustainable in the long term? Even for a country that dominates the rare earth supply chain, eventually the rare earth deposits will be exhausted, no? Same problem with the rapid expansion of electric vehicles and renewables: batteries and solar panels have significantly shorter life spans than fossil fuel engines and turbines in power plants, and rely on rarer materials than just steel and copper. Therefore one assumes that a priority should be research and investment into recycling technology, as China is doing for batteries and solar panels.
Also, what do the logistics look like of data centers with millions of chips if chips have to constantly be replaced? At some point even this replacement process will necessarily have to be automated, else human technicians will not be able to keep up with the 24/7 replacement. Or is replacement done in bulk rather than on a per chip basis as they burn out?
You can recycle them, but yeah this is definitely a problem. Although, we're already in that scenario with other kinds of data centres too. The servers are designed so you can replace things on the fly, but you still have to constantly make chips. With AI, it's more of a problem cause chips tend to run on higher loads more of the time. So, making more efficient models would be one way to address that.