We may be approaching, however tentatively and indirectly, the point where some AI reality takes hold, both in general and in terms of AI’s impact on the network. It’s not that actual enterprise AI types have changed, but that what most of them have known from the first is starting to influence broader media and even vendor planning channels.
Technology isn’t what it’s about, it’s the benefits of technology that drive change. All of our hype waves, including AI and 6G, are created because a new technology is a kind of blank slate, a “UFO” that because it’s not actually landing in your yard, lets you assign any properties to it that you find interesting or helpful. Hype builds on supposition, which is most powerful when reality isn’t messing up our flights of fantasy. We’re seeing a bit of that now, in vendor announcements.
Nokia’s AI lab announcement and Street success in May demonstrate that a WAN vendor can’t really participate in AI traffic/network needs. Wall Street comments that the success it’s earned comes from the leveraging of optical networking in the data center, where AI inarguably can drive massive traffic changes. Let’s run with that point for a moment.
AI drives traffic within the model, for sure. It can also drive traffic to the model, if the model consumes a large amount of data in training or through something like RAG. It’s less likely, for now at least, to drive traffic growth between model and user. So far, less than two percent of enterprises have reported any noticeable growth in network traffic AI-to-user. The reason is that AI, in chat or agent form, tends to deliver answers, which don’t generally require a huge amount of data. Hey, people struggle to absorb 500-word tech articles, so why would they want massive AI reports?
The corollary to this is that AI is going to impact networking of the models, not connection of users to results. That means that companies like Nokia need to participate in AI hosting infrastructure, not wait for AI traffic to emerge in the WAN, if they want to gain from AI. It also means that telcos are unlikely to play in the AI world we see today because they are WAN providers facing a LAN-connected AI revolution…sort of.
What about the RAG stuff? The problem with that, says enterprises, is that the majority of data they’d want AI to analyze to create business cases is the same data that they don’t want to host in the cloud for security and governance reasons. The majority of AI agents hosted by cloud giants, for example, operate on less sensitive data, supporting missions that improve productivity but don’t create revolutions in traffic. Of the three agent models enterprises have always recognized (interactive, embedded, and workflow), the first two can easily be applied to non-governed data but the last poses the same governance concerns that “moving everything to the cloud” poses.
If AI networking is to develop, everything we can see today says it must develop from a decision to host AI within the enterprise. This decision would create an explosion in AI data centers, which would mean that AI data center networking would become an enterprise challenge and not just one for the cloud providers. It would also release the governance concerns on core enterprise data, which would then permit AI to be integrated into more of today’s business-critical workflows.
We are seeing, enterprises tell me, less growth in self-hosting of AI than expected. The reasons are first that all the public focus on AI is on cloud-hosted models, which can’t fully address governance concerns, and second that the technology itself seems to enterprises to be in a state of flux. What’s really needed to self-host AI?
Broadcom seems to want to answer that, within VMware. Its VCF 9.1 release is aimed at production AI hosting, meaning to enterprises the workflow model of AI agents that would let them open business cases involving the application of AI to core, governed, data. In their press release, you’ll find this quote: “As more enterprises turn to AI for driving competitive advantage, they face three critical challenges: data and IP privacy concerns, surging infrastructure costs, and their readiness for the world of agentic AI,” said Krish Prasad, senior vice president and general manager, VMware Cloud Foundation Division, Broadcom. “VCF 9.1 is a single unified platform that addresses all three and delivers one of the most advanced infrastructure for Private AI. We enable zero-trust security for AI, reduce costs through intelligent infrastructure optimization and hardware choice, and provide the flexibility to run both agentic workflows and accelerated inferencing on the same platform.”
This may be the most important announcement for AI of recent times. Enterprises don’t want to blaze AI trails, they want to ride an AI superhighway, which somebody has to build for them. VMware is perhaps the most widely deployed virtualization and compute infrastructure framework, so making it AI-ready, and coming out with the specific solution to the three challenges that Prasad cites, gives enterprises at least a view of the on-ramp.
How does this impact 6G, though, and what about the Broadcom announcement of the first 50G PON edge AI portfolio? Let’s see.
There are two ways that AI could credibly generate more edge traffic. One way is that AI supports new and broadly used real-time applications, which would require sensor data to bring reliable real-world state into machine-accessed form. Another is to find non-governed AI missions that do generate more traffic, which would mean finding a credible collection of prospective consumers of those missions.
As all of my readers know, I’m a fan of real-time AI missions. My problem is that you can’t sell connectivity for them until they’re ready to be connected, and nobody in the telco world seems to recognize that. 6G without real-time applications is hype, just like 5G was. That should mean that telcos and their vendors should be looking to promote those real-time applications, not just connecting hypothetical ones. How do you connect a hypothetical application? With a hypothetical service. How do you pay for it? With hypothetical money. Do better, people, and it may be that Nokia’s dabbling into data center networking for AI will lead its lab initiative to look at all the self-hosting issues. VMware is there to help.
For 50G PON, Broadcom doesn’t need to sell every consumer or small business site on massive AI commitments, only sell some of them, perhaps ten to twenty percent. If that number in a PON span need capacity, then the span needs to provide it as an option. Right now, ten percent of consumers (roughly) and fifteen percent of workers, can justify cloud-productivity AI tool purchases. Drop the price in half and you triple those numbers.
These two trends, the expansion of the cloud-agent value proposition and the creation of a reliable and cost-effective self-hosting model for AI, are independent. Get either going and you have some form of AI success. Get them all going, and do you have that elusive AI revolution. Maybe, and we may find the answer to that in the next year.
