Matilda Bailey stands at the intersection of traditional infrastructure and the explosive growth of next-generation networking solutions. As a specialist who has watched enterprise networking evolve from simple switching to complex software-defined layers, she offers a unique perspective on how artificial intelligence is currently dismantling and rebuilding the industry’s core foundations. The conversation explores a significant shift in the market: the return of physical infrastructure as a central priority. After years of the industry trying to abstract away the network, the demands of distributed AI have made the underlying hardware more critical than ever before.
The discussion covers a wide range of transformative themes, beginning with how AI functions as a high-speed backplane for modern data centers. We delve into the internal shifts at major technology providers, where tens of thousands of developers are using AI to manage codebases of staggering complexity. The dialogue also touches on the revolutionary use of the Linux kernel for real-time security, the strategic bridge between legacy virtual machines and cloud-native Kubernetes environments, and the emerging challenge of controlling autonomous AI agents that operate within enterprise systems.
Distributed AI systems now require memory, compute, and storage to work together across physical infrastructure at scale. How is the network evolving to act as the essential backplane for this new era?
In the current landscape, the network has reclaimed its role as the central nervous system of the data center, functioning much like the PCI bus does inside a single server. When you are orchestrating distributed AI systems, you aren’t just sending data from point A to point B; you are facilitating a constant, high-speed dialogue between GPUs, massive memory banks, and storage arrays. This physical infrastructure must work in perfect synchronization at a massive scale to prevent bottlenecks that would otherwise render expensive compute resources idle. The result is a new operating model where the network is no longer a commoditized layer hidden by software, but the foundational backplane that makes high-performance AI possible. It is the one component that enterprise customers are realizing they must count on if they want their next-generation applications to function with any degree of reliability.
With a development team of roughly 12,000 people, the internal software process has clearly changed. What specific impacts have AI coding tools had on managing the millions of lines of code found in modern networking hardware?
The shift inside large development organizations has been nothing short of transformative, especially when you consider the sheer volume of code we are dealing with today. A standard Catalyst switch or a sophisticated firewall can contain anywhere from 50 to 100 million lines of code, a scale that previously hit a “context ceiling” for older AI models. In the past, we saw small teams of five or ten people accomplish tasks that used to require a hundred developers working for a full year, but that was mostly on greenfield projects. Now, with newer frontier AI tools, we can finally process the deep context of legacy products, allowing for a massive acceleration in software development across the entire product portfolio. It feels like we’ve finally broken through a wall, moving from simple automation to a sophisticated understanding of complex, decades-old systems.
Frontier AI models are now being used to find vulnerabilities at a scale that was previously impossible for human teams. How does this change the traditional approach to securing data center infrastructure?
The traditional “locked-down” model of data center security, where you test a configuration and then leave it alone for months, is effectively dead. We are seeing frontier models, such as Anthropic’s Claude Mythos, comprehending entire, multi-million-line codebases and finding vulnerabilities that have remained hidden from human eyes for years. This isn’t just a one-time improvement; it’s an ongoing, relentless cycle of discovery that will continue to uncover new risks at a pace we’ve never experienced. Because switches and routers are high-performance, inline systems, you can’t simply take them offline every time a new vulnerability is found without causing massive disruptions. We’ve had to rethink the entire security stack from the kernel up to ensure that protection is as dynamic and continuous as the threats themselves.
The use of eBPF technology and the Isovalent platform suggests a much deeper integration with the Linux kernel. What does this look like in practice for a network administrator managing a live system?
By leveraging eBPF technology, we can actually step inside the Linux kernel to inspect memory and intercept every system call or function call as it happens. For a network administrator, this translates into a tangible sense of control through tools like Live Protect, which are integrated directly into operating systems like NXOS or IOS. When a vulnerability is flagged on a dashboard, the admin can simply click a button to apply a compensating control to a specific process ID and file. This shield blocks the malicious action without affecting any other part of the system or requiring a reboot. It’s an incredible shift to be able to modify the behavior of a running system in real-time without ever touching the underlying binaries or disrupting the flow of traffic.
Most enterprises are still heavily reliant on virtual machines despite the push toward AI and containers. How can organizations bridge the gap between twenty-year-old VM technology and a Kubernetes-driven future?
It is a reality of the modern enterprise that while AI is exciting, the vast majority of current workloads are still running on virtual machines that have been the standard for over twenty years. The friction usually occurs because these VMs operate at Layer 2, while Kubernetes was born in the cloud at Layer 3, making migration a networking nightmare that often requires reengineering IP addresses. We are solving this by using a software bridge that allows VMs to migrate into a Kubernetes environment one at a time without changing their identity. This allows an organization to see their legacy VM-based workloads and their modern container-based apps as peers on the same dashboard. It removes the pressure of a forced, “all-at-once” migration and lets different generations of technology live together on a single, unified infrastructure.
As AI agents start performing tasks on behalf of users, the traditional model of password-based access seems insufficient. What kind of controls are necessary to prevent these agents from overstepping their bounds?
When you have an AI agent acting for a human, the standard six-month password rotation and broad application access are simply too permissive and dangerous. If an agent is authorized to file an expense report, there is no reason it should have the credentials to access procurement systems or make unauthorized purchases—we certainly don’t want an agent buying a Porsche on the company dime. We are implementing task-scoped and session-specific controls through SSE solutions and hybrid mesh firewalls to address this exact problem. This creates an ephemeral layer of security that follows the agent only for the duration of a specific task and within a very narrow set of permissions. It’s about moving toward a model where access is granted based on the “intent” of the action rather than just the identity of the user.
What is your forecast for the networking industry over the next twelve months?
My forecast is that we are heading toward a total convergence where the distinction between “legacy” and “next-gen” infrastructure finally disappears. Within the next year, I expect customers to realize they can power their AI applications of tomorrow, their Kubernetes apps of today, and their VM workloads of yesterday all with one single architecture and one design. We will see the network move from being a complex series of silos into a unified, intelligent fabric that automatically adjusts to the specific needs of the workload, whether it’s a massive AI training job or a simple database query. This consolidation will make the “startling” advancements we are seeing now feel like the standard operating procedure for every modern enterprise.
