Introduction
Artificial intelligence is hungry for data. High‑volume model training and inference send floods of traffic across enterprise networks, and interactive agents demand low latency and real‑time responses. Yet most networks were not built with these demands in mind. Research by Broadcom (reported in NetworkWorld) found that 99 % of organisations have a cloud strategy, but only 49 % believe their networks are ready to support AI, and 95 % lack adequate visibility into network traffic (networkworld.com). KPMG’s infrastructure survey echoes this, noting that just 17 % of companies have networks capable of handlingAI workloads, while more than halfreport only moderate scalability (kpmg.com). Without addressing these blind spots, even well‑designed AI initiatives will falter.
Why networks struggle with AI workloads
Explosive data flows
Training and inference on large models produce orders of magnitude more network traffic than traditional applications. Edge devices, sensors and agentic workflows create streams of data that must be aggregated, processed and stored in real time.Legacy networks with rigid capacity planning and hierarchical architectures cannot adapt quickly to such surges.
Lack of observability
Riverbed’s study shows that organisations deploy an average of 13 observability tools and that 95% see standardising data across applications and infrastructure as critical (riverbed.com). Yet documentation of how data is collected, transformed and used is often scattered or absent. Without clear provenance, AI models risk propagating errors and bias. Lack of observability
Siloed network and data teams
AI solutions span compute, storage, networking and security, yet these domains are often managed separately. Misalignment leads to bottlenecks: data scientists may design models that require latency guarantees the network cannot meet, while network teams are unaware of upcoming AI workloads.
Strategies for an AI‑ready network
Upgrade bandwidth and reduce latency
Invest inhigh‑throughput, low‑latency connections across data centres, clouds and edgelocations. Modern optical transport and software‑defined wide area networks(SD‑WAN) provide flexible bandwidth allocation and dynamic path selection.Evaluate emerging technologies such as 400G Ethernet and edge computing tobring data processing closer to where it is generated.
Adopt unified observability
IConsolidate disparate monitoring tools into a single observability platform that spans applications,infrastructure and user experience. Riverbed’s research suggests organisations are moving towards open standards like OpenTelemetry; 88 % have already started adoption and 94 % see it as a stepping stone to AI automation (riverbed.com). Unified observability allows teams to detect bottlenecks, correlate events and maintain servicelevels.
Embrace microservices and event‑driven architectures
Legacy monolithic networks create single points of failure and limit scalability. Microservices and event‑driven designs decouple components, enabling them to scale independently and handle asynchronous data flows. HPE’s networking division notes that combining microservices with agentic AI enables self‑driving operations and cross‑platform innovation (hpe.com). Event‑driven systems allow AI agents to subscribe to network events and act only when needed, reducing unnecessary traffic.
Strengthen security and governance
AI workloads introduce new attack surfaces. Cisco’s AI Readiness Index reports that only 31 % of organisations feel prepared to secure autonomous agents, signalling a need forzero‑trust architectures (kramerand.co). Implement network segmentation, authentication and encryption to protect sensitive data. Use policy‑based controls to ensure thatAI agents cannot overwhelm network resources or access unauthorised systems.
Coordinate cross‑functional planning
Bring network architects, data engineers and AI teams together when designing and deployingAI projects. Joint capacity planning ensures networks are scaled ahead of demand. Shared visibility through observability tools facilitates rapid troubleshooting and performance tuning.
Balanced perspective
Upgrading network infrastructure can be costly and complex. However, the alternative – running AIworkloads on networks ill‑equipped for them – risks failed deployments and wasted investment. By gradually integrating observability, automation and flexible architectures, organisations can enhance performance without massive overhauls. Cloud‑native networking services and consumption‑based models allow incremental upgrades aligned with business value.
Conclusion
AI initiatives will only succeed if the networks they rely upon can deliver high throughput, low latency and reliable visibility. Current surveys reveal significant gaps in readiness (networkworld.com), (kpmg.com). By investing in modern connectivity, unified observability, microservices and event‑driven patterns, and cross‑functional governance, organisations can ensure their networks are ready to support AI at scale.
