Cisco’s announcement of its G300 chip was positioned to address AI workloads, but these days about the only thing we’re not claiming is driven by AI is politics, and I’m not sure about that either. In any event, anything that’s positioned relative to AI has to be examined in two ways. First, does it actually provide value in AI evolution? Second, how is that value derived, from AI or from something broader that happens to at least possibly consume it?
The first question above may be the hardest to answer with any authority, given the range of things that are claimed to be AI missions. There seem to be four AI models out there, best understood by arranging a row/column structure with a 2×2 dimension. You can draw this out or visualize it, as you wish.
For the rows, label the first “cooperative” and the second “autonomous”. For the columns, label the first “cloud” and the second “self-hosted”. OK?
The cooperative cloud model, our first square, represents the AI that most people who use AI are thinking about. There’s a huge data center complex hosting LLMs, and users connect to it almost always via the Internet. This connection lets the users ask questions, typically ones whose answer is derived from general knowledge that is culled from the Internet itself.
The autonomous cloud model, down one row in the “cloud” column, is where many of today’s cloud giants are trying to get to. “Agentic” AI is positioned by them as a form of AI that does something rather than tells people something. The challenge with that, in a value sense, is that it’s not clear how much of it can be done without gathering in what the user would consider personal, or what a worker’s company would see as proprietary.
Now for the self-hosted column, starting with the “cooperative” row. Here we have missions that use language models hosted by a company or, in theory, a person. In roughly 80% of applications reported so far, this is done to overcome data sovereignty issues or personal privacy concerns, and in the other 20% because either cost or QoE demands dedicated handling rather than a resource pool. However, this group of missions still works interactively with a user.
In the self-hosted autonomous box, we have applications where the AI does something similar to what a traditional application or component would do, which is to process input and produce output without continual human interaction. This is the AI model that enterprises have generally found capable of making a business case, but it’s also variable in terms of how it uses data, just as applications/components are.
Generally, Cisco’s chip opportunities, the justification for and value of the G300 line, would arise out of either hosting AI or delivering large quantities of data to/from the models, wherever they were hosted. Obviously, it’s massive data movement that justifies the G300. Massive data movement would tend to rule out our 1:1 box, the conversational cloud, but would likely play in at least some of the AI applications of the other boxes.
Generally, I think, Cisco’s chip opportunities are also greatest where the network impact is across many buyers, particularly because Cisco has a broad network incumbency. That means it likely derives more opportunity from the self-hosted column than the cloud column. The cloud-autonomous combination might involve data movement, but the capacity of the user-to-cloud connection limits the value of being able to push data faster from its enterprise source.
To me, this says that Cisco gains from the G300 to the extent that enterprises host their own AI and apply it to locally sourced data. If the cloud model prevails, then the best Cisco could hope for is that they’d sell G300 gear to hyperscalers who demanded a massive discount. The worst is that those buyers would use generic chips from people like Broadcom or Nvidia, and Cisco would see nothing from it. So, the G300 is not a bet on AI in general, but on self-hosted AI. Cisco’s press release, linked above, shows that with the heading “Silicon One G300: The Networking Foundation for the Agentic Era.” Yes, for sure that’s kissing the most credible current AI baby in a PR sense, but it’s also a positioning statement.
Does Cisco see what enterprises have said all along? I think that’s likely, and for sure Cisco’s broad AI positioning would seem to apply mostly to enterprises who plan to host their own, in their own data centers. However, Cisco would surely not send hyperscaler hosts of AI away, and might well also hope to grab up any smaller ones that evolve, such as within telcos. So, there is AI value to the G300.
On to our second point, which is whether Cisco can add value to AI, or even outside AI. The answer to that, I think, is a clear “Yes!”
The biggest risk to any network-mission value proposition is congestion that impacts QoE. Traffic management is complicated and costly, but essential if capacity is limited. Something like 80% of router code is dedicated to it, and enterprises estimate that it generates three-quarters of user complaints and two-thirds of their network operations costs. If you simply made networks very fast, raised the capacity, you’d have a profound impact on QoE, business cases, and costs.
What would the optimum data center network for the 2020s era look like? Enterprises would say “infinite capacity, 100% availability, and free”, but they all know that’s not realistic. What they actually would like is one with a capacity so high that congestion becomes unlikely, where alternate paths can be created to respond to failures but don’t congest in the process, and that offers autonomous recovery but high netops visibility. In short, they’d like what Cisco seems to be promising the G300 and the rest of its portfolio can approach. And since Cisco says (in the same release) “To enable AI network builders of all sizes – hyperscale to enterprise – Cisco is introducing the next generation of Cisco N9000 and Cisco 8000 fixed and modular Ethernet systems, powered by Silicon One, and designed for the extreme power and thermal demands of AI workloads,” they’re promising a level of capacity higher than enterprises would likely need themselves.
The most important point, though, is that the value of capacity doesn’t stem from AI, it stems from the increasing need to organize general business value from specific applications and data. AI is itself a tool in doing that, but enterprise needs for QoE, resilience, and cost-efficiency existed long before AI came along.
Traffic management is needed when traffic conditions need to be managed, and those conditions drive not only most of the opex but most of networking’s complexity. Where traffic conditions are most critical is in the data center, where traditional transactional data, increasing real-time data, and business intelligence-gathering are creating a web of information flows that are complex, QoS-dependent, and business-critical. Those who build the data center of the future would love to trade a modest capex increase for a reduction in opex and complexity. If a major data center networking vendor doesn’t offer that, then there’s always white boxes, so Cisco is on the right track as long as they don’t get so caught up with the AI hype wave that they’re caught in a collapse.
