Enterprises are rapidly moving beyond AI experimentation to production-scale deployments, but the journey is fraught with challenges. Building an AI infrastructure that is secure, scalable, and efficient requires careful orchestration of compute, networking, storage, security, and observability. Cisco and NVIDIA have jointly developed a solution that addresses these requirements head-on: the Cisco Secure AI Factory with NVIDIA. This modular reference design is built to help organizations turn AI investments into tangible business value while mitigating risks and operational complexities.
Three core challenges in enterprise AI
According to Abhinav Joshi, leader of AI solutions and product marketing at Cisco, organizations face three primary obstacles when building enterprise AI infrastructure: deployment complexity, security vulnerabilities, and performance bottlenecks. These challenges are amplified by the rise of agentic AI, which relies heavily on inferencing and places greater demands on every layer of the infrastructure stack.
Deployment complexity arises from the need to rapidly operationalize an AI infrastructure that seamlessly integrates compute, networking, storage, security, and observability. Additionally, a Kubernetes-based container management platform and a robust AI software toolchain are essential for consistent development, testing, and deployment of containerized AI applications. Many organizations lack the in-house expertise to design and validate such a stack from scratch, leading to delays, misconfigurations, and suboptimal performance.
Security vulnerabilities represent the second major hurdle. AI models, frameworks, applications, and the supporting infrastructure are attractive targets for attackers. Without integrated security measures, organizations expose themselves to threats such as prompt injection, model poisoning, data leaks, and manipulation of large language models (LLMs) through malicious inputs. As AI agents become more autonomous and ingest diverse data sets, the attack surface expands considerably. Attackers can disrupt operations, extract sensitive information, or compromise the integrity of AI-driven decisions.
Performance bottlenecks, particularly related to networking, constitute the third challenge. AI workloads produce enormous amounts of network traffic during pre-training, post-training, fine-tuning, retrieval-augmented generation (RAG) pipelines, and inferencing. High-speed interconnects between GPU servers, data throughput to storage layers, and real-time response delivery to end users all require high-performance networking. Without it, GPUs may remain underutilized, jobs take longer to complete, and token economics suffer—meaning organizations pay more for every useful token generated. High-performance networking keeps workloads moving efficiently as agents retrieve context, coordinate tools, and execute multi-step workflows.
A unified approach: Cisco Secure AI Factory with NVIDIA
Cisco and NVIDIA jointly address these three challenges with a modular reference design that integrates high-performance compute, networking, and storage infrastructure with Kubernetes and AI software. The solution is pre-validated and compliant with NVIDIA Enterprise Reference Architectures, reducing deployment risk and accelerating time to value. Enterprises can choose the components that best meet their immediate needs, with the confidence that they can add capacity later as demands grow.
Security is embedded at every layer of the full stack, from models and applications to agents, providing protection from the supply chain to runtime. Key Cisco products delivering this protection include Cisco AI Defense, Cisco Hybrid Mesh Firewall, Cisco Isovalent Runtime Security, and Splunk Enterprise Security. Tight integration among these solutions enables quicker response to critical exposures. Cisco’s Live Protect capability, for example, puts guardrails around AI jobs, allowing them to continue running even when vulnerabilities are discovered—a critical feature given that model training can take days to complete.
Another challenge many organizations face is a lack of in-house IT talent with AI expertise. Cisco addresses this through professional services from its own teams and channel partners. Additionally, at a recent Cisco Live event, the company announced new deployment automation software called Stack Automation by Quali. This tool further reduces deployment time from days to hours, enabling both professional services teams and customers to stand up secure AI infrastructure quickly and reliably.
The modular nature of the reference design provides flexibility for a wide range of AI use cases, from training and fine-tuning to inferencing and agentic workflows. By leveraging NVIDIA’s accelerated computing platforms, including GPUs and networking such as NVIDIA Quantum InfiniBand, the solution delivers the performance required to keep AI workloads moving efficiently. Combined with Cisco’s robust networking portfolio—including Cisco Nexus switches, Cisco UCS servers, and Cisco Silicon One–based routing—the infrastructure is designed to eliminate bottlenecks and maximize GPU utilization.
Agentic AI: a new frontier requiring robust infrastructure
Agentic AI represents the next evolution of enterprise AI, where models not only generate responses but also take actions, interact with tools, and pursue goals autonomously. This paradigm places even greater demands on infrastructure. Agents must retrieve context from multiple data sources, coordinate with other agents, and execute multi-step workflows in real time. Without a solid foundation, latency and throughput issues can cripple agent performance and undermine user trust.
The Cisco Secure AI Factory with NVIDIA is purpose-built to support agentic AI workloads. Its integrated security ensures that agents operate within defined guardrails, preventing unauthorized actions and data leaks. Observability features, including Splunk, provide visibility into agent behavior, performance, and security events, allowing IT teams to monitor and optimize operations continuously.
Moreover, the solution’s support for Kubernetes-based container orchestration simplifies the management of agentic AI services, enabling seamless scaling, updates, and failover. The pre-validated stack reduces the risk of integration errors, allowing organizations to focus on building and deploying agents that deliver business value rather than troubleshooting infrastructure issues.
Overcoming deployment complexity with automation
One of the most significant barriers to enterprise AI adoption is the time and skill required to deploy infrastructure. Cisco’s partnership with Quali brings Stack Automation to the table, a tool that automates the deployment of secure AI infrastructure. This automation reduces the deployment window from several days to just a few hours, dramatically shortening the path from planning to production. It also minimizes human errors that can lead to security gaps or performance issues.
Enterprises can take advantage of this automation whether they are building new data centers or extending existing ones to support AI workloads. The modular reference design allows them to start small and grow incrementally, adding compute nodes, storage, or networking capacity as demand increases. This flexibility is particularly valuable for organizations that are still exploring AI use cases and want to minimize upfront investment.
Performance at scale: the role of networking
Performance bottlenecks in AI factories often stem from networking, not just compute. Cisco and NVIDIA address this by integrating high-speed, low-latency networking throughout the infrastructure. Cisco’s Nexus switches and Silicon One–based routers provide the backbone for GPU-to-GPU communication, storage access, and user-facing traffic. NVIDIA’s Quantum InfiniBand or Spectrum Ethernet switches further accelerate inter-node communication for distributed training and inferencing.
By eliminating network bottlenecks, the solution ensures that GPUs are fully utilized, jobs complete faster, and token economics improve. For enterprises deploying agentic AI, where low-latency responses are critical, high-performance networking is a must. The combined Cisco-NVIDIA stack delivers the throughput and low latency required to support real-time, interactive AI applications.
Security as a foundation, not an afterthought
Security is often treated as an add-on in AI infrastructure, but the Cisco Secure AI Factory integrates it at every layer. Cisco AI Defense protects models and data during training and inference. Cisco Hybrid Mesh Firewall enforces segmentation and traffic inspection across the data center and edge. Isovalent Runtime Security provides kernel-level security for containerized workloads, while Splunk Enterprise Security offers advanced threat detection and response.
Live Protect is a standout capability that allows AI jobs to continue running even when vulnerabilities are discovered, reducing downtime and productivity loss. This is especially important for long-running training jobs that cannot easily be interrupted. The integrated security approach also helps organizations comply with regulatory requirements and protect sensitive data.
In the context of agentic AI, security becomes even more critical. Agents that act autonomously must be prevented from being manipulated or from accessing unauthorized data. The Cisco Secure AI Factory enforces policies that constrain agent behavior, monitor for anomalies, and respond to incidents in real time. This provides the confidence needed to deploy AI agents in production environments without exposing the organization to undue risk.
As enterprises move from experimentation to production-scale agentic AI, success will depend on more than raw compute. Organizations will need AI factories that securely deliver valuable outcomes while operating efficiently at scale. The Cisco Secure AI Factory with NVIDIA provides a sound foundation for these projects, integrating security, performance, and automation into a modular, pre-validated stack. By addressing deployment complexity, security vulnerabilities, and performance bottlenecks simultaneously, it enables enterprises to realize the full potential of AI—turning technology investments into measurable business outcomes.
Source: Network World News