The real bottleneck of AI infrastructure: data delivery, not GPU computing power.

Industry Background

Enterprise AI is moving from the experimental phase to the operationalization phase. Over the past 18 months, companies have rushed to procure GPUs, deploy large language models (LLMs), and build AI toolchains. However, according to IDC's 2025 Spotlight report, organizations are shifting from one-time AI deployments to repeatable, scalable production architectures. As AI becomes embedded in core business operations, performance, security, reliability, and operational consistency are as important as model innovation.

Yet, many AI projects fall short of expectations. Technology decision-makers often blame insufficient computing power, so they add more GPUs, expand clusters, or switch to more powerful models. But infrastructure teams find that the problem often lies at the data level: expensive GPUs are not lacking in compute power, but in data.

Market Impact

GPU idle costs are extremely high. According to Uptime Institute's 2025 annual outage analysis, more than half of enterprises have experienced a major outage costing over $100,000, and one in five have faced costs exceeding $1 million. When data pipelines are blocked, GPU utilization plummets, directly eroding AI investment returns.

Nirav Shah, Senior Vice President of Product Marketing at F5, compares modern AI infrastructure to an iceberg: above the surface are the visible LLMs, AI applications, orchestration frameworks, and GPU clusters (about 10%); below the surface lie the storage, networking, traffic management, security controls, and data movement systems that determine whether investments yield real value (90%). Enterprises focus too much on the tip of the iceberg while ignoring the true bottleneck.

Competitive Landscape

This shift in perception is reshaping the AI infrastructure market. Traditional GPU vendors (e.g., NVIDIA) remain core beneficiaries, but data delivery solution providers—such as F5, NetApp, Pure Storage, etc.—are gaining new growth opportunities. Application delivery controllers (ADCs) are evolving into application delivery and security platforms (ADSPs), integrating traffic engineering, security controls, and visualization into one.

In contrast, architectures that rely solely on direct storage access face challenges. Tightly coupled designs expose performance vulnerabilities at AI scale: latency spikes, throughput congestion, and traffic surges have little impact on traditional applications but can exponentially amplify AI performance issues.

Enterprise Insights

• Enterprises should reassess the "below the waterline" part of their infrastructure. First, shift from tightly coupled to loosely coupled architectures: insert an intelligent control layer (e.g., ADC) between storage and compute to handle TLS termination, certificate management, traffic optimization, and policy enforcement, freeing up the storage system's CPU resources to focus on data services.Second, build resilience from three dimensions:
• Availability: Ensure AI workloads can always access healthy storage resources, with automatic failover in case of faults.
• Policy: Prevent traffic anomalies such as the thundering herd effect and retry storms, shaping traffic and maintaining a security posture through the intelligent control layer.
• Delivery: Isolate maintenance changes of clients and storage nodes, keeping data flow uninterrupted.

Case: A global large financial services institution deployed dedicated physical ADCs, achieving a performance improvement of over 5x for object creation, read, and delete operations, reducing deletion latency by an order of magnitude with no performance regression.

Future Outlook

In the next 12-24 months, AI infrastructure investment will shift from GPU cluster expansion to data pipeline optimization. Enterprises will focus more on the return on investment of the "data delivery layer", driving ADC to evolve into ADSP. Within 24-36 months, most large enterprises will adopt loosely coupled architectures, and data delivery will become a key competitive advantage for AI operations. Investors should pay attention to infrastructure vendors that provide data delivery, security protection, and intelligent traffic management, rather than focusing solely on compute power providers.