At Mobile World Congress 2024 (MWC24) last week in Barcelona, artificial intelligence (AI) was, as expected, a big topic. In fact, of the nearly 40 meetings I had across three days, AI was the only topic that came up in each one. Regarding networking, there are two types of AI – AI for networking and networking for AI. The former uses AI to run the network, and the latter deploys a network to support AI. Organizations building a network to run AI should first implement AI to operate the network.

Current networks were never designed for the demands of AI. Years ago, I recall a conversation with Charlie Giancarlo just after he became CEO of Pure Storage several years ago. Having known Giancarlo as an executive who has always thrived in capitalizing on market transitions, I was surprised he went to a storage company. He predicted that as AI matured, processors, networks, and storage would need to evolve, and that’s now true. The processor evolution is evident as the GPU is now at the center of AI strategies, and consequently, Nvidia has left its once formidable rival, Intel, in the dust.

Regarding storage, Pure Storage differentiates itself on greater performance, agility, and energy management than legacy vendors, such as Dell. As a result, Pure Storage's stock has nearly doubled in the past year.

And then there’s the network, which hasn’t quite hit the inflection point yet, but it’s coming. On its last earnings call, Cisco talked about the upcoming networking for AI trends. CEO Chuck Robbins stated, “We are clear beneficiaries of AI adoption. The majority of that $1 billion in orders (mentioned earlier in the call) will turn into revenue in our FY25.” Juniper and Arista have both echoed similar statements regarding the AI opportunity.

AI evolution of networking vendors

All the major network vendors are now working on evolving their products to meet the unique demands of AI. One of the positive signs for the industry is the formation of the Ultra Ethernet Consortium, where rival companies such as Arista, Cisco, Juniper, and HPE have come together to create an Ethernet standard that can challenge InfiniBand as the technology that can support AI. I don’t believe Ultra Ethernet will universally displace InfiniBand, but many customers do want an option, and having the industry come together on a standards-based offering will create a rising tide where everyone wins.

This means for networking for AI, the technology is here, or at least right around the corner. What’s missing is an operational model that can support AI. Without change, an organization’s AI plans could be derailed for the following reasons:

• Slow configuration and change management. The current operational model is slow and error-prone. Network engineers often must make changes to the network on a box-by-box basis, leading to long lead times when implementing a network-wide change. AI requires speed, and people-centric operations can no longer keep up with the accelerated pace that AI brings. AI automates tasks and can make changes in a fraction of the time that people can.
• Excessive downtime. My research shows that human errors are the largest cause of unplanned network downtime. Network professionals often work in “firefighting” mode, particularly when troubleshooting problems. This leads to mistakes, which can cause unplanned downtime, which in a network supporting AI could lead to massive amounts of money being wasted, given the expense of data scientists and the cost of processing data. With AI-based operations, there is no chance of typos, syntactical errors, or other actions that lead to people-related downtime.
• Impact on sustainability goals. There’s no question that AI flies in the face of sustainability as GPU-enabled servers use considerably more power than ones without. Also, AI systems require massive scale, causing companies to expand their data center footprint for the first time in a decade. With AI comes more power consumption and perhaps a step backward with sustainability efforts, but AI can minimize the impact by optimizing the network to be as energy efficient as possible.
• Security risks. You can’t talk about AI without talking about security and privacy. The concept of “guardrails” has become commonplace as there is a greater awareness of biased AI and hallucinations. One aspect of AI that does not get enough attention is how the infrastructure gets secured to protect the massive amount of data companies have in their AI systems. Finding threats today has been compared to finding a needle in a stack of needles. People can’t sift through security information and correlate it quickly, but machines can.

Networking for AI is coming, and the best way companies can prepare for this is by first implementing AI for networking. Even if there are no immediate plans to build a network for AI, AI can help streamline operations and troubleshoot problems faster, which will positively impact employee and customer experience. Network engineers should not fear AI but consider it a tool that can make a good engineer great.