The next AI wave: what’s standing in the way?

Data centers are becoming more and more power-hungry. What do we do?

During the Capital Markets day in November, Nokia shared that consumer AI traffic is expected to grow by 20%, and enterprise and industrial AI traffic by 50% over the next decade.

How will we generate enough power to meet the demand? Power itself might not be the problem, John Harrington, Nokia’s executive vice president (EVP) of Network Infrastructure, whom I met in Lisbon at the annual Atlantic Convergence conference, thinks.

“It's the grid. It's how you adapt the grid if you're putting a data center in a location that's now consuming a lot of power; the grid has to be reconfigured. And this takes a long, long time,” he told Cybernews.

Nokia is working to reduce power per bit – the amount of data required to transmit or process a single bit of data.

“In every generation of technology that we produce, every chipset, every router and every product, we're always looking at how we can improve the performance and reduce the power by doing more integrated technology, more integrated componentry,” he explained.

According to Nokia’s December blog post, over the last 20 years, “optical transceiver technology has cut power-per-bit from 5 W/Gb to 0.04 W/Gb.” That is more than a 100-fold reduction.

“Data center facilities are often separated by long-haul distances of more than 1,000 kilometers and require DWDM optical transport networks that combine ultra-low power-per-bit efficiency with operational simplicity,” the company explained.

This is what Harrington called scaling up.

“Scaling up basically means I want to improve the performance of the box. I want to make it faster. I want to make it better. And then scaling out means just adding more boxes. So this is consuming more power,” he said.

According to a recent report by an investment bank, Goldman Sachs, global power demand from data centers is expected to grow 175% in 2030 vs. 2023 levels, or, as they put it, “equivalent of adding another Top 10 power consuming country.”

Experts believe the electric grid will require significant investment, and expect “roughly $790 bn of grid spend through 2030, up from $780 bn in our last update.”

And while reports like this from MIT, claiming 95% of AI pilots fail, ruffle the market feathers for a bit, Harrington seems to think bigger and names big issues such as data security, sovereignty, and the supply chain as potential obstacles that could slow down the technological boom.

What data can travel long distances, and what data should be kept domestically? How to ensure data can be transmitted in a secure way? Harrington also wonders about the supply chain.

“We have a big backlog of orders. There's no lack of demand. I'm not saying that we're in a COVID-like supply crunch. I'm just saying we cannot produce the things quickly enough,” he said.

In November, Nokia announced plans to invest $4 billion in research and development and manufacturing for “AI-ready network connectivity.”

Companies like Nokia find themselves at the eye of an AI supercycle.

“Hyperscalers, high-performance compute companies, infrastructure connectivity providers like us are finding that the demand is so fast and that we all have to operate at a much faster clock speed,” Harrington said.

Unlock exclusive Cybernews content on YouTube