The heavy buzz around all things AI got louder in the financial reports of networking vendors this quarter, even though AI hasn’t made a significant impact on most vendors’ financial performance and supply chain challenges remain a more immediate concern.
Vendors such as Cisco, Arista, Juniper, Extreme and HPE’s Aruba report that they are shipping more products, thanks to multi-month efforts that include significant product redesigns and relentless efforts by their supply-chain teams to address component shortages. But the situation is still challenging, and some enterprise customers still face order delays.
“Clearly backlog is coming down as we expected it to this year, but it still remains about 3X what we would normally expect,” Rami Rahim, CEO of Juniper, told Wall Street analysts this quarter. Juniper reported about a $2 billion backlog at the start of the year and expects that to be cut in half by the end of its fiscal year in December.
Cisco, too, is still reporting siazable backlog levels but says the situation has improved dramatically.
“The aging of our backlog has continued to improve as the supply situation normalizes, and as expected, increased customer deliveries reduced our year-end backlog to roughly double historical levels as we enter fiscal '24,” Cisco CFO Scott Herren told analysts at Cisco’s most recent earnings presentation. “That excess backlog will work down in the first half of fiscal '24 with the majority of that being worked off in Q1, by the way,” he said.
While backlog and supply chain issues are still a topic of concern, the subject of AI development opportunities was predominant for all vendors.
Cisco CEO Chuck Robbins said the company has taken some $500 million in orders for AI Ethernet fabrics, for example.
“The acceleration of AI will fundamentally change our world and create new growth drivers for us,” Robbins said. “Cisco's ASIC design and scalable fabric for AI position us very well to build out the infrastructure that hyperscalers and others need to build AI ML clusters. We expect Ethernet will lead in connecting AI workloads over the next five years.”
Cisco recently unwrapped new high-end programmable Silicon One processors aimed at underpinning large-scale AI/ML infrastructure for enterprises and hyperscalers. AI/ML models have grown from needing a few GPUs to needing tens of thousands linked in parallel and in series. The number of GPUs and the scale of the network are unheard of, Cisco said.
“The AI opportunity is exciting, and as our largest cloud customers review their classic cloud and AI networking plans, Arista is adapting to these changes and doubling down on our investments in AI,” Jayshree Ulall, CEO of Arista, told analysts at Arista’s recent financial call. “We expect larger clusters and production deployments in 2025 and beyond. In the decade ahead, AI networking will become an extension of cloud networking to form a cohesive and seamless front-end and back-end network.”
“We are in the middle of trials for back-end AI networks, leading to pilots in 2024,” Ulall added.
Arista and Cisco are betting big that Ethernet will be the tool of the AI networking trade in the future. They are both part of a recently announced group – the Ultra Ethernet Consortium (UEC), hosted by the Linux Foundation – that’s working to develop physical, link, transport and software layer Ethernet advances.
The group, which includes AMD, Broadcom, Eviden, HPE, Intel, Meta and Microsoft, aims to enhance today’s Ethernet technology in order to handle the scale and speed required by AI.
“AI traffic and performance demands are different as it comprises of a small number of synchronized high bandwidth flows, making them prone to collisions that slow down the job completion time of AI clusters as they connect 1000s of GPUs, generating billions of parameters,” Ulall said.
Arista has been developing features for its EOS networking software such as intelligent load balancing, and advanced analyzers to report and [other] tools that can achieve predictable performance and established Ethernet and IP technology will ultimately be the underpinning architecture to handle that, Ulall said.
Cisco’s goal is to combine these enhanced Ethernet technologies and take them a step further to let customers set up what it calls a Scheduled Fabric. In a Scheduled Fabric, the physical components – chips, optics, switches – are tied together like one big modular chassis and communicate with each other to provide optimal scheduling behavior.
“As we get scheduled fabric out and these customers get more comfortable moving from InfiniBand to Ethernet, I think that's when we'll start to see the real impact of AI. And maybe it's late '24, but I would suspect into '25 for sure,” Robbins said.
In the meantime, one networking competitor that says AI is already impacting its bottom line is Juniper.
“Customers are recognizing Juniper’s leadership when it comes to AI-driven operations delivered via a modern microservices cloud,” Juniper’s Rahim said. “Revenue from the Mist segment of our business, which are products driven by Mist AI [Juniper’s core cloud-based management system], had a record quarter growing by nearly 100% year over year in the Q2 timeframe, with orders growing by nearly 40% year over year.”
Juniper recently integrated the ChatGPT AI-based large language model (LLM) with Mist’s virtual network assistant, Marvis. Marvis can detect and describe myriad network problems, including persistently failing wired or wireless clients, bad cables, access-point coverage holes, problematic WAN links, and insufficient radio-frequency capacity.
By adding ChatGPT capabilities, Juniper is expanding the role of Marvis and augmenting its documentation and support options to help IT administrators quickly get the necessary assistance with problems or challenges, Juniper stated.