East-West Is the New North-South: Rethink Security for the AI-Driven Data Center

In modern data centers, traffic patterns have changed significantly. Distributed AI systems rely on real-time data exchange among training nodes, orchestration tools, and storage systems. These connections occur deep inside the network between workloads rather than across networks. 

For example, traffic may go from Server 1 to Server 2, which may be in the same server rack. Traffic never leaves the rack or traverses the network to cross the firewall (or legacy segmentation solution).

A single AI model training session can generate terabytes of east-west traffic as datasets move across GPU clusters, container environments, and scalable compute fabrics. The tolerance for latency is extremely low. Traditional perimeter tools weren’t designed for this level of throughput or complexity.

More significantly, most perimeter tools can’t see this traffic at all. As a result, it’s important to ask these questions:

• Do our current firewalls protect our AI pipelines, or just our front door?

• Is our cybersecurity strategy built for east-west traffic, or is it still anchored in north-south assumptions?

Imagine your data center as a large, high-security office building.

North-south traffic is like visitors entering and exiting the building. They pass through physical security, badge readers, or reception. Access is logged and monitored. This is where traditional firewalls operate, screening what comes in from the outside.

But once inside, the visitors move freely throughout the building. Some enter human resources offices, others go into the finance or engineering departments. But not everyone should have access to payroll systems or sensitive information.

That internal movement, which represents workload-to-workload communication, is east-west traffic. And in most data centers, it is barely secured, if at all.

Here is what that means for your AI infrastructure:

• Lateral movement becomes easy. Threats can quickly move from one function to another without detection.

• Insider threats go unnoticed. Internal access is often assumed to be trusted, increasing the chance of tampering or data leaks.

• AI pipelines and datasets remain exposed. Sensitive data and metadata used in training models aren’t monitored with the same rigor as external traffic.

• Compliance scope expands. Without proper segmentation, proving compliance with frameworks like National Institute of Standards and Technology (NIST) or Health Insurance Portability and Accountability Act (HIPAA) becomes difficult and expensive.

Even with endpoint agents, virtual firewalls, or traditional network security measures, internal traffic among services often bypasses deep inspection because of performance trade-offs.

If the majority of your data center traffic flows east to west, your security controls must be positioned accordingly.

That is exactly what the integration of Akamai Guardicore Segmentation with the Aruba CX 10000 smart switch, powered by AMD Pensando data processing unit, is designed to deliver.

This approach embeds microsegmentation enforcement directly into the data center switch at the top of every rack. As traffic enters the network fabric, it is inspected, validated, and matched to policy in real time. There’s no need to redirect traffic to centralized firewalls, hairpin traffic, or consume resources on software agents.

With this model, you can gain:

• High-speed, in-fabric enforcement built for both modernized applications and AI workloads

• Automation of policy creation and lifecycle management

• End-to-end visibility across the AI infrastructure ecosystem

• Reduced attack surface without sacrificing performance

It’s a smarter and more scalable way to defend against unauthorized access, cyberattacks, and supply chain vulnerabilities. This model is especially critical in hyperscale and AI-driven environments.

• How do we monitor and control east/west movement between workloads today?

• Are our existing security measures optimized for AI, or are they built for yesterday’s data center?

• Can we detect security incidents across dynamic containers, APIs, and model training environments?

• What happens to sensitive data once it moves past the perimeter?

Zero Trust doesn’t stop at access control. It requires continuous visibility and enforcement across every connection, including internal ones.

In the third and final part of this series (coming soon!), we’ll explore how Zero Trust switching enables just that. You’ll see how Akamai Guardicore Segmentation and AMD Pensando together deliver a high-speed, embedded security architecture for AI-ready environments. This approach avoids the latency, overhead, and visibility gaps of traditional firewalls.

Akamai is working closely with security teams, network architects, and service providers to rethink segmentation for real-world AI systems. If you’re building next-generation AI data centers, managing large-scale machine learning workloads, or addressing risk management and threat detection across critical infrastructure, we’d love to connect.

Let’s define what secure AI should really look like — starting from the inside out.