Ten years ago, data center needed 10 megawatts (MW) of electricity to operate, or enough juice to power 2,000 households for a day. Today, that figure is closer to 300 MW. What’s to blame? Artificial intelligence is a reasonable culprit.

A majority of data center power used to support AI workloads comes from servers, which account for 61% of the modern data center’s energy consumption, VP of Cisco’s Engineering Sustainability Office Denise Lee told SDxCentral.

But too many data center are unprepared for the jump in power, compute and networking demands associated with AI and machine learning workloads. Rack densities, for example, are creeping into the triple digits and rendering traditional air cooling ineffective.

“The projected requirements have growth expectations far beyond the planned capacity,” Lee said.

How to manage AI energy demands

To that point, the industry needs new platforms and systems, like direct-to-chip or liquid cooling, that are designed to manage the surge in AI workloads and high performance computing (HPC) (HPC) needed to run those applications.

Another way to manage the growing impact of AI is to incorporate circular design principles into data center technologies and ecosystems. Circular design is a model wherein waste is designed out of the process and materials are reused when possible.

“This way, companies can support emissions reduction and reduced energy usage while achieving the performance needed to support AI demands,” Lee said.

In addition, advanced energy management and control tools are fundamental priorities for IT teams that need visibility, control and performance monitoring. “Organizations can see what their output is, control it through software and policy and understand its impact on performance,” she said.

Reducing data center power use with Cisco

Lee highlighted Cisco’s Silicon One chip as another avenue to lower the environmental impacts of AI. Silicon One customers can build a data center network that uses 50% less optics, 40% less switches and one less layer of networking, which translates to 1 MW of power savings per AI cluster.

“Cisco’s Silicon One chip enables networks that are fast, flexible, energy efficient and ready for the increasing demands of AI,” Lee said. “At a time when data centers are power hungry, Silicon One helps set the new standard for energy use and sustainability to contend with the sharp rise of AI.”

In addition, Cisco’s partnership with Nvidia to join the latter’s Blackwell GPUs with the former’s Ethernet networking addresses the unique requirements of AI on data center infrastructure.

Nvidia’s Blackwell GPU architecture touts faster performance and lower energy consumption for AI models than the company’s previous GPUs, which makes it “a sustainable choice for our work and the work of our customers,” Lee said.

“This partnership is the start of an engineering relationship where we will work together to deliver solutions that help customers automate and optimize fabrics and compute for sustainable AI workloads,” she said.

The bigger picture

Besides helping its customers increase the sustainability of their data center, Cisco acknowledges its responsibility to the broader community. “It’s important for tech companies to develop a holistic approach to environmental sustainability, which includes how companies operate their business, engage with suppliers and help customers and communities reduce their overall environmental impact to adapt to the growing demands of AI,” Lee said.

The first step in that approach is accelerating and supporting the transition to clean energy. “We must move away from fossil fuels toward wind, solar and renewable energy,” she said. “This allows businesses to evolve from a linear to a circular economy based on recycling, reusing and reducing the resources used.”

Investments in local communities and ecosystems are also crucial to this aim. “Develop[ing] technologies, capabilities and talent to be climate-resilient will be key in addressing these challenges as we move forward as an industry,” Lee said.