On Friday the 13th, Bloomberg published a piece by Lynn Doan entitled ‘AI Wants More Data, More Chips, More Real Estate, More Power, More Water, More Everything.’ It’s an eye-catching screed that brings you into the conversation about how we’re going to sustain large language models with data centers in the years to come.
Doan’s piece also asks a number of questions that I’ve heard people asking over the last weeks and months. Here are some of the highlights that I think are most relevant to what we’re dealing with right now in the artificial intelligence world.
New Energy Solutions: We’re Serious, All of a Sudden
Doan points out that a lot of the brand-new initiatives to invest in new kinds of power for data centers were languishing prior to AI creating this demand.
Maybe the best example is nuclear power. We knew that nuclear power is clean, and the technology has advanced since the grim days of a potential Three Mile Island meltdown in the 1970s. However, for whatever reasons, America had been dragging its heels to develop new nuclear facilities, until this dramatic new need emerged in the past few years. Now Companies like TerraPower are rushing to put new nuclear solutions in place – and most of the country seems to be on board.
In terms of overall power demand, Doan points out that data centers are projected to use up to 8% of all power by the end of the decade, and notes that this will require a significant shift in thinking about energy.
As an interesting little side note, she quotes the CEO of Constellation Energy talking about corporate plans that seem “functionally impossible” – and points out that Constellation is in charge of reviving Three Mile Island for Microsoft.
Data Styles of the Rich and Powerful
Another point that the author comes back to more than once is that those with more resources will have more clout over the results that generative AI provides.
There’s the reality of richer groups of people having better LLMs at their disposal, and then there’s the disparities around the data that systems need to function.
Late in the essay, Doan shows how most available high-quality data is applied to English speaking societies with certain European heritage patterns. Experts are suggesting that if we don’t have more data for diverse and non-white society groups, they will lose out in the AI revolution. That leads back to discussions about bias, issues that engineers have been wrestling witgh ever since these technologies emerged.
Thirsty Chatbots and the Water Problem
Doan also estimates data centers will need 500,000,000 gallons of water per day by 2030, and shows how this conflict over water might hurt communities that need their own drinking water as well. In an interesting metric, she suggests that asking ChatGPT a series of 10 to 50 questions will require the equivalent of a 16 ounce bottle of water.
You can imagine large language models quaffing their own Dasani or Fiji bottles the same way humans do in meetings. But the result is that we’re going to have some pretty strenuous demand for drinking-quality water.
Networking Demands: Sending Data Where it Needs to Go
Then there’s all of the networking capability that we need for these systems as well. Doan talks about companies like Verizon ramping up, and how the new data center demand is going to affect the market. She also mentions the corporate move to the cloud, and how that got companies used to anticipating greater networking needs even before AI came in to play.
“Dramatic increases in computing performance over the years continue to be reflected in the network traffic they generate,” James Donovan wrote in April of 2020 at The Cabling Science Institute. “The potential of devices to utilize network bandwidth is still far from fully realized by current software, but the upward trend is clearly evident. (A device’s) growing ability to handle voice, video and multimedia, as well as data, is causing all these elements to converge on the same networks. This convergence, combined with rapid growth in the individual areas, is increasing demand for network capacity.”
Quarreling Over Precious Metals
Here’s another point that brings up something that I was covering last week…
Amid all of the chip wars that we’ve seen regionally and globally, there’s also the hunt for raw materials and the geopolitical tensions that have an impact on demand and supply.
Specifically, the U.S. and China are starting a kind of protectionism around technology, and that has led China to ban the export of rare earth metals gallium and germanium to the U.S. As Doan points out, China also has the lion’s share of silicon resources, and so American companies and administrations have to keep that in mind as they’re crafting policy.
The Trouble with Brain Drain
Finally, Doan also suggests companies are going to need a great deal of human talent.
It’s not just a question of taking masses of unemployed people and making them data engineers. It’s getting access to the skilled professionals who will be needed to usher in the next era of AI.
Think about all of these issues as we move forward. AI does indeed want a lot, and how it gets it will matter.
