In my blog earlier this week, I noted that the self-hosted AI space was currently dominated by computer vendors, but that another driving player might emerge. Yesterday, Juniper made an announcement that clearly shows it wants to be one of those AI-dominating players. They may achieve that goal. Coverage of the announcement was decent; HERE is a nice story.
Juniper’s briefing deck (“Seize the AI Moment: Networking for AI”) notes the driver for Juniper’s interest, which is that enterprises want to host their own AI in most of the missions that are likely to generate real projects, and that networking for AI is unique. I agree with both points, as the blog I reference illustrates, based on what enterprises have been telling me. While AI self-hosting clearly has to be created around servers, the connectivity is what really molds the servers into a model host, and connects it with both training/source data and users. Given that, a network player has a realistic chance to be a driving AI player.
Juniper had three elements in its announcement; the Ops4AI Lab, multi-vendor Juniper-validated designs (JVDs) for AI, and the automatic tuning of the AI cluster fabric for performance and efficiency. Each of these seems to address some of the issues enterprises have reported with AI self-hosting.
The AI lab is something Juniper has been working on internally, and with this announcement is also available (through the Juniper sales teams) to customers. The lab is supported by relationships with players like AMD, cerebras, Broadcom, Intel, Nvidia, UltraEthernet Consortium, Vast, and WEKA, and it’s built on Juniper’s data center virtualization software, Apstra. Open source models (like Llama2) are available, or customers can bring their own models. All the required software is incorporated, and customers can run their own applications (if they are compatible with the configuration) or run standard AI/LLM benchmarks.
The lab plays into the validated designs, which can be configured onto the lab setup. The JVDs are IMHO the key to this all, in that they offer enterprises a proven AI hosting configuration that can actually run stuff and be deployed by them at almost any desired scale. These designs lay out an entire AI cluster with all the switching and routing needed for intro-cluster connection and connection to databases and users. That all makes JVDs a blueprint for an AI cluster, which is just what enterprises are looking for, and why computer vendors have tended to be the dominant player in self-hosted AI so far.
The lab and validated designs include the announced fabric optimization and tuning capability Juniper also announced, to deal with the challenges of handling AI data flows. Juniper combines priority-based flow control (RoCEv2) and explicit congestion notification to create its Data Center Quantized Congestion Notification (DCQCN) to maximize QoE for AI while maintaining reasonable network investment levels. This is an area that most enterprises haven’t even considered, but that can make or break an AI deployment. One of the comments current self-hosted AI user made to me this year was that they tended to overbuild on the network side of their AI deployments for fear of having the network constrain their applications.
This could be very important to Juniper. While I can’t find any justification for the IDC prediction that the top two thousand companies will dedicate 40%of their core IT spending to AI (the most any company suggested to me for next year didn’t hit half that level), it’s very possible that the largest factor in data center spending growth over the next five years will be AI self-hosting. Driving decisions here could be worth a lot.
One obvious question here is the impact of the proposed acquisition of Juniper by HPE. HPE is an AI server player, but not a giant in the space according to enterprises who have offered me their views. If the HPE deal goes through, the Juniper announcements could combine with HPE to give the new HPE a major advantage in the AI self-hosting space, providing as usual that they handle the marketing/positioning effectively. Juniper was insightful in the briefing they gave me (and presumably others in the press/analyst community) and I’d hope that HPE could take advantage of this.
The acquisition still isn’t a sure thing, and may not come about until early 2025, and so Juniper has to be prepared for either of these eventualities. I think the announcement does that, and even though it wasn’t public until yesterday, 13 enterprises said they’d gotten advance notice and 38 offered comments on it overall. The consensus was positive, particularly among the 13 who had advanced knowledge, so if Juniper can keep up the momentum on this, with or without the HPE deal going through, they should have a good shot at exploiting the self-hosting AI space.
In fact something like this could be necessary for the space to reach full potential. Right now AI has a lot of senior management support (many CIOs say it has too much, and that senior management has too many AI dreams and not enough plans) and so getting something realistic approved would be easier. It’s possible that barriers to self-hosting, including a simple lack of CIO confidence that a plan can be created that would work out, will develop this year if there are failures in projects created by AI ignorance that a model like Juniper’s might address, even dispel. That could stall AI investment short of even my more modest suggested levels.
AI will never transform businesses in its public-model generative AI form. Self-hosting, particularly self-hosting of something like Llama2, is essential to control AI cost and data security/privacy risks. Juniper’s announcement gives them the technical foundation for a credible shot at the real AI opportunity, even if the HPE deal isn’t approved. The key will be in the marketing/positioning. The sales initiatives so far seem to be generating appropriate interest, but a broader base is essential. HPE could help to provide that if the deal closes, but if it does not then Juniper will need to prepare marketing/positioning to take up the slack.
