April 15, 2024

sullivanprogressplaza

It's the Technology

Meta axes two Danish datacenters amid AI push • The Register

Meta axes two Danish datacenters amid AI push • The Register

Meta has canned two datacenters under development in Denmark as part of a broader plan to deepen investments in artificial intelligence.

In a statement provided to The Register, Peter Münster, Meta’s head of communications for the Nordics, confirmed the termination of the project.

“We are currently expanding our datacenter in Odense with a third building that we expect to complete over the course of the summer of 2023,” he said. “We have decided not to move forward with the planned expansion beyond three buildings and we will be ending construction work on the two additional buildings.”

The decision comes just weeks after Meta announced it would lay off approximately 11,000 employees – about 13 percent of the company’s global workforce. The layoffs followed a string of difficult quarters for the Social Network. During the most recent quarter, revenues slid four percent year over year while profits slid 52 percent to 4.4 billion.

While Meta shed human capital, the org indicated it was doubling down on datacenter infrastructure. During its Q3 earnings call this fall, CFO Dave Wehner said Meta was in an “investment cycle to build more headroom in our datacenters.”

Those investments included a larger emphasis on AI/ML workloads, which Wehner said Meta would be “carefully evaluating” to determine the scale of future deployments. In the weeks since making these statements, Zuck’s mothership appears to have decided to focus future deployments around AI infrastructure – to improve engagement both with ads and with user-generated content.

“Over the last several weeks, we’ve announced a number of steps that will make us a more streamlined organization. A big part of that is a shift of more of our resources to higher priority growth areas, including a strategic investment in artificial intelligence,” Münster said.

Thus the switch in Danish strategy. “AI workloads at scale require a different type of datacenter than those built to support our regular online services.”

GPU-accelerated compute nodes commonly deployed for AI/ML workloads can consume an order of magnitude more power than a general compute server. It’s not uncommon for a single 5U GPU server to consume north of 6.5kW, and next-gen systems are expected to consume closer to 11kW. As a result, datacenters must be equipped to handle these systems’ power and thermal requirements.

Despite terminating the datacenter projects, Meta claims it remains committed to Denmark’s Odense region. ®