Fortinet Brings AI Data Centre Security to NVIDIA DPUs

As enterprises race to operationalise generative AI, the modern data centre is being reshaped around accelerated computing—especially GPU-rich infrastructure that can train and run large models at scale. That shift is also changing the security equation. Traditional, host-based security and perimeter-only approaches can struggle to keep pace with the east-west traffic patterns, multi-tenant workloads, and high-throughput demands that come with AI clusters.

Against this backdrop, Fortinet has announced an integration that brings its data centre security capabilities to NVIDIA BlueField DPUs, aiming to improve how security and networking are applied in AI-focused environments without sacrificing performance. The move reflects a broader industry trend: security controls are increasingly being offloaded and accelerated so they can operate closer to the data path and keep up with line-rate traffic.

Why AI data centres are forcing a rethink of security

AI infrastructure often concentrates enormous compute resources—and valuable data—in a small number of clusters. That creates a compelling target for attackers and raises the cost of downtime. At the same time, AI workloads can generate massive, continuous flows of internal traffic as datasets are staged, models are trained, checkpoints are written, and inference services are scaled across nodes.

In practical terms, AI-driven data centres face a combination of challenges:

High east-west traffic that can bypass traditional north-south security chokepoints.
Performance sensitivity, where latency and throughput directly affect model training times and inference responsiveness.
Operational complexity, including mixed workloads, containers, virtual machines, and multi-tenant environments.
Rising compliance expectations around sensitive data handling, particularly as AI systems interact with proprietary or regulated datasets.

This is why many organisations are embracing architectural patterns such as zero trust segmentation, microsegmentation, and hardware-assisted isolation—approaches that can reduce attack surfaces while keeping performance predictable.

What DPUs change in the security and networking stack

A Data Processing Unit (DPU) is designed to offload infrastructure functions—networking, storage, and security—from the CPU, helping free compute resources for applications (in this case, AI workloads). NVIDIA’s BlueField DPUs are widely positioned for modern data centres that need accelerated networking and isolation between tenants or workloads.

By placing security services on DPUs, organisations can potentially:

Apply security controls closer to the traffic, improving visibility and enforcement for east-west flows.
Reduce CPU overhead by offloading inspection and policy enforcement to dedicated hardware.
Support stronger separation between workloads, which is particularly useful in shared AI environments.

In an era where AI compute is expensive and in high demand, freeing CPU cycles and avoiding bottlenecks is not just a technical optimisation—it’s an economic one. The faster the cluster can train or serve models, the better the utilisation of high-cost infrastructure.

What Fortinet is delivering on NVIDIA BlueField

Fortinet’s announcement centres on extending its data centre security capabilities to NVIDIA BlueField DPUs, aligning with the needs of AI-centric environments. The intent is to deliver high-performance security enforcement that can scale with accelerated computing deployments, while supporting modern segmentation requirements.

While organisations will evaluate the specifics based on their architecture and tooling, the value proposition is clear: bring security functions into the accelerated data path so that protection does not become the limiting factor for AI infrastructure.

This approach also fits into the broader shift toward distributed security, where enforcement is embedded across the fabric rather than concentrated at a single perimeter gateway. For AI clusters, that can mean better containment if a workload is compromised, and more consistent policy enforcement across nodes.

What this means for enterprise buyers and data centre operators

For CISOs, platform teams, and infrastructure leaders, security decisions around AI are increasingly tied to architecture choices. Integrations that combine networking acceleration with built-in security controls can reduce the need for complex traffic hairpinning or oversized CPU allocations for inspection.

In practical evaluation terms, organisations may want to consider:

Deployment model: how security policies are managed and updated across clusters.
Performance impact: measurable latency/throughput changes under AI-scale east-west traffic.
Segmentation strategy: mapping policies to tenants, projects, and model pipelines.
Operational fit: integration with existing Fortinet tooling and data centre workflows.

As AI adoption continues, the market is likely to see more partnerships between security vendors and silicon/platform providers. The goal is to keep security aligned with how compute, networking, and storage are actually being built for AI—modular, accelerated, and highly distributed.

Conclusion

Fortinet’s move to bring AI data centre security capabilities to NVIDIA BlueField DPUs highlights an important direction for the industry: security must scale at the speed of accelerated infrastructure. As enterprises invest in GPU clusters and AI platforms, solutions that offload and distribute security controls can help protect high-value workloads without undermining performance. In a landscape where both compute time and risk exposure are costly, integrating security into the data centre’s accelerated fabric is becoming a strategic necessity, not an optional enhancement.