Redesigning the data center for the AI era

Sometimes change happens so rapidly that it’s difficult to keep up, we must make concessions or face tough obstacles. GPUs set the pace, as AI began to take over the world, but data centers struggled – some couldn’t meet time-to-market demands, while others paid a high price.

A stark disparity in the construction of AI infrastructure builds can be seen if we look back a few years, or even compare it to today’s standard HPC loads. Despite their power, a typical 20 to 30 kW setup is no longer enough to support AI workloads.

Why today’s data centers are reaching their limits

With the latest popular B200s, the GB series, using 100–130 kW per rack, the world is racing to meet the new benchmark: the 600 kW Rubin Ultra. We are now looking at a load range of 20 kW to 600 kW per square meter – essentially compressing a 15 to 20 MW datacenter into just 1,000 sqm of white space.

From a real estate or environmental standpoint, this may seem like a positive shift. But it brings serious engineering and operational challenges – ones that are keeping operations and project teams up at night. Here are some of the challenges.

  • Load tension or leakage:

    The recommended load tension for 100 kW racks is two to three hundred kilo per square meter. This simply means it’s difficult to design risers that can support such weight. As a result, cooling pipes are often installed above the floor or above the racks as a workaround. But this isn’t a long-term, adaptable solution. Solving one problem tends to create another, and maintaining data center uptime becomes a real challenge.

  • Data center expansion:

    As infrastructure becomes denser, large data centers can no longer utilize land to its full potential. One might assume this means higher deployment density within a smaller footprint – but that’s not the case. Government regulations and the limits of electrical engineering constrain how much power can be delivered to a single site. In my view, the “Land Use Effectiveness” of data centers is declining by the day.

  • Project planning:

    Data centers now need strong mechanical and civil infrastructure capable of supporting any solution. However, designing for a wide range of capacity loads is not economically ideal. Budgets rise significantly, and ROI becomes questionable, especially when low-capacity deployments occupy the same high-spec space.

  • Hardware obsolescence:

    Outdated hardware makes project planning even harder. With systems lasting just two to three years, data centers are already struggling with the earlier issues. What’s worse, upcoming systems aren’t compatible with the old ones, leaving no room for reuse or adaptation.

Rethinking the data center stack

Given all these challenges, it’s time to rethink and re-architect the four key pillars of a typical data centre:

  • Telecommunication:

    Future data center networks will be hyper-dense. Each NVIDIA SuperPod might require 22 km of optics. Expect sidecars the size of a single rack and up to 5,000 cables on a Rubin board with a 72-layer PCB. Are your Meet Me Rooms ready to support this?

  • Electrical:

    Power is the lifeblood of any system. Racks are arriving with built-in power, and data centers are moving beyond the traditional N or 2N setups. Systems now require odd numbers of connections, which makes redundancy unpredictable, whether for PDUs, busbars, ATS, or N(x) transformers.

  • Architecture:

    Each new iteration brings changes: cable exhausts, cooling liquid networks, load-bearing considerations. Every component must be accounted for, and no single failure should affect the rest. Data centers must scale from HPC to next-gen AI with minimal redesign.

  • Mechanical:

    Risers capable of supporting two to three hundred kilos or more in the future, plenum space for massive power and network cables, and allowances for extensive cooling infrastructure are daily challenges for project teams. Every aspect of physical engineering must be considered, and all mechanical requirements met.

 

*The author handles end to end operations and strategy for the Compute Nordic. He has around 13 years of experience with various major DC players in India and Norway.

**This piece first appeared in Issue 9 of W.Media’s Cloud & Datacenters magazine. Click the image below and head to page 22-23 to read the story.

Publish on W.Media
Author Info:
Gaurav Dixit
Site Manager, Compute Nordic
Share This Article
Related Posts
Other Popular Posts
Northeast Asia News
Northeast Asia News