For most of Cisco’s history, the company occupied a well-defined place in the technology stack. Cisco built the networks that connected everything else. Servers evolved. Storage evolved. Applications evolved. Cisco moved the packets between them.
That position created one of the most successful companies in technology history.
It may not be enough for the AI era.
Reading through Cisco’s recent Silicon One announcements and related discussions at Cisco Live, I found myself less interested in the specifics of agentic AI or scale-across networking than in a broader pattern emerging across the industry. Cisco is increasingly describing itself not as a networking company, but as an AI infrastructure company. Silicon One sits at the center of that effort.
The distinction matters.
For the better part of two decades, much of the networking industry operated on a model that separated hardware systems from the silicon that powered them. Cisco, Juniper, Arista and others built products around merchant silicon supplied by companies such as Broadcom. That model worked well because networking was largely a game of performance, reliability and software differentiation.
AI is changing the equation.
Today’s infrastructure conversations are increasingly dominated by companies that control critical layers of the stack. NVIDIA is the most obvious example. NVIDIA’s position in AI is not simply the result of having powerful GPUs. The company controls the silicon, the networking, the software frameworks and an ecosystem optimized around those assets.
That level of vertical integration has become a strategic advantage.
Cisco appears to have reached a similar conclusion from the networking side of the industry.
Viewed through that lens, Silicon One looks less like a chip initiative and more like a long-term strategic necessity.
The recent Cisco blogs discussing agentic AI and scale-across networking provide useful examples. On the surface, these pieces focus on emerging infrastructure requirements. Autonomous agents create new traffic patterns. AI workloads require more visibility. Security policies must operate at machine speed. Massive training clusters are forcing operators to connect GPUs across multiple data centers.
Those are all legitimate challenges.
What stands out, however, is how often Cisco returns to the same theme. Silicon One is presented as the foundation that enables everything else. Security is tied to the silicon. Telemetry is tied to the silicon. Adaptability is tied to the silicon. Scale-across networking is tied to the silicon. Even optical networking is increasingly discussed alongside Silicon One as part of a converged architecture.
The message is difficult to miss.
Cisco is making the case that future AI infrastructure cannot be assembled from loosely connected components. It must be designed as an integrated system.
That argument becomes easier to understand when viewed against the backdrop of how AI infrastructure is evolving.
For the last several years, the AI conversation has been dominated by GPUs. Investors, media outlets and technology buyers have focused on compute capacity because that is where the bottleneck initially appeared. The assumption was that bigger AI models required more GPUs.
That remains true.
The difference is that organizations are discovering GPUs are only one part of the challenge.
Power availability has become a constraint. Cooling has become a constraint. Data center capacity has become a constraint. Increasingly, networking is becoming a constraint as well.
The largest AI deployments now resemble infrastructure projects more than traditional IT projects.
A modern frontier-model training cluster can require tens of thousands of GPUs and consume well over 100 megawatts of power. In some cases, the power requirements exceed what a single facility can reasonably support. As a result, operators are exploring architectures that distribute compute resources across multiple sites while attempting to preserve the behavior of a single cluster.
This is the problem Cisco describes as scale-across networking.
Whether scale-across becomes the industry’s preferred terminology is almost beside the point. The underlying challenge is real. AI infrastructure is growing beyond the physical limits of individual data centers.
Once that happens, networking moves from supporting actor to critical infrastructure.
The network is no longer simply carrying traffic between applications. It is participating directly in the operation of the AI system itself.
That shift creates an opportunity for Cisco.
It also creates a challenge.
Historically, networking vendors benefited from a relatively modular ecosystem. Customers could choose switching platforms, optics, software and silicon from different suppliers and integrate them into a functional architecture. AI infrastructure is pushing the market in a different direction. As performance requirements increase and operational complexity grows, vendors are increasingly promoting tightly integrated stacks.
NVIDIA has done this successfully with GPUs, networking and software. Hyperscalers often pursue similar strategies internally. Even cloud providers are investing heavily in custom silicon to reduce dependence on outside roadmaps.
Cisco’s response appears increasingly clear.
If networking is becoming strategic infrastructure, then the underlying silicon becomes strategic as well.
Owning the silicon roadmap allows Cisco to optimize hardware and software together. It allows the company to build telemetry capabilities directly into packet processing. It allows security features to operate closer to the hardware. It creates opportunities to coordinate networking systems with optical technologies in ways that are difficult to achieve when multiple vendors control critical components.
Most importantly, it gives Cisco greater control over its own future.
That may ultimately be the most important aspect of Silicon One.
Technology companies rarely invest billions of dollars developing custom silicon because they enjoy designing chips. They do it because they believe controlling that layer will become strategically important.
The AI era appears to be validating that belief.
There is another aspect of Cisco’s strategy worth watching.
The company is increasingly discussing networking, optics and silicon together rather than as separate businesses. The scale-across discussion is particularly revealing in this regard. Connecting geographically distributed AI clusters is not purely a networking problem. It is simultaneously a routing problem, an optical transport problem, a power-efficiency problem and a systems-design problem.
Cisco’s argument is that these elements should be optimized together.
Whether customers agree remains to be seen.
Many enterprises and service providers continue to prefer open architectures and best-of-breed approaches. Others may find value in reducing complexity through greater integration. The balance between those two philosophies has shifted back and forth throughout technology history.
AI may push the pendulum toward integration once again.
That is why I think the most interesting story in Cisco’s recent announcements has little to do with agentic AI. It is not even scale-across networking.
The more significant story is Cisco’s effort to redefine its role in the market.
For decades, Cisco connected the infrastructure that powered business and the internet. In the AI era, the company appears to be pursuing a larger ambition. It wants to become part of the infrastructure itself.
Silicon One is not simply a networking chip. It is Cisco’s attempt to secure a seat at the table where the future of AI infrastructure is being built.
The industry has already learned that controlling the compute layer creates enormous strategic leverage. Cisco appears to be betting that the same principle applies to networking. If AI infrastructure becomes a full-stack competition, companies that own more of the stack will have more influence over where the market goes next.
That may be the real reason Silicon One matters. Not because it helps Cisco build better switches or routers, but because Cisco understands that becoming an AI infrastructure company requires controlling more than the network. It requires controlling the foundation on which the network runs.
