itstargetnews
Presented by F5
As enterprises pour billions into GPU infrastructure for AI workloads, many are discovering that their precious compute resources are sitting idle longer than expected. The reason is not the hardware. It’s this often invisible data delivery layer between storage and compute that starves GPUs of the information they need.
"While people focus their attention, rightly so, on GPUs, because they are very important investments, this is rarely the limiting factor," says Mark Menger, solution architect at F5. "They can do more work. They are waiting for the data."
AI performance is increasingly dependent on an independent, programmable control point between AI frameworks and object storage – one that most enterprises do not intend to architect. As AI workloads scale, bottlenecks and instability occur when AI frameworks are tightly coupled to specific storage endpoints during scaling events, failures, and cloud transitions.
"Traditional storage access standards are not designed for highly parallel, heavy, multi-consumer AI workloads," said Maggie Stringfellow, VP, product management – BIG-IP. "Efficient data movement in AI requires a separate data delivery layer designed to abstract, optimize, and secure data flows independent of storage systems, as the GPU economy makes it difficult to see quickly and expensively."
Why AI workloads prefer object storage
These bidirectional patterns include multiple ingestions from continuous data acquisition, simulation output, and model checkpoints. Combined with reading-intensive training and inference workloadsthey emphasize the tightly coupled infrastructure upon which storage systems depend.
While storage vendors have done significant work in scaling data throughput into and out of their systems, that focus on throughput alone creates knock-on effects across the entire transfer, traffic management, and security layers associated with storage.
The stress of S3-compatible systems from AI workloads are multidimensional and very different from traditional application patterns. It’s less about raw throughput and more about concurrency, metadata pressure, and fan-out considerations. Training and fine-tuning create more challenging patterns, such as multiple parallel readings of small to medium-sized objects. These workloads also include repeated transmission of training data throughout the periods and periodic checkpoint write bursts.
RAG workloads introduce their own complexity by increasing the request. A single request can flip dozens or hundreds of additional pieces of data, progressing to more detailed, related pieces, and more complex documents. Stress concentration is less about capacity, storage system speed, and more about request handling and traffic shaping.
The risks of tightly coupling AI frameworks to storage
When AI frameworks directly connect storage endpoints without an intermediate delivery layer, it can easily increase operational vulnerability during scaling events, failures, and cloud migration, which can have significant consequences.
"Any instability in the storage service now has an empty blast radius," Menger said. "Anything here would be a system failure, not a storage failure. Or frankly, bad behavior in one application can have ramifications for all consumers of that storage service."
Menger described a pattern he saw with three different customers, where tight coupling resulted in complete system failure.
"We see a lot of training or maintenance workloads that fill up the storage infrastructure, and the storage infrastructure goes down," he explained. "At that scale, recovery cannot be measured in seconds. Minutes if you’re lucky. Usually hours. GPUs are currently not powered. They are hungry for data. These high cost resources, all the time the system is down, negative ROI."
How an independent data delivery layer improves GPU utilization and stability
The financial impact of introducing an independent data delivery layer goes beyond preventing catastrophic failures.
Decoupling allows data access to be optimized independently of storage hardware, improving GPU utilization by reducing idle time and contention while improving cost predictability and system performance as scale increases, Stringfellow said.
"It enables intelligent caching, traffic shaping, and protocol optimization closer to computing, reducing cloud exit costs and maximizing storage," he explained. "Operationally, this isolation protects storage systems from unlimited AI access patterns, resulting in more predictable cost behavior and robust performance under growth and variability."
Use a programmable control point between compute and storage
The answer to F5 is to set it Application Delivery and Security Platform, powered by BIG-IPas a "storage front door" which provides health-aware routing, hotspot avoidance, policy enforcement, and security control without requiring application rewrites.
"The introduction of a level of transmission between calculation and storage helps to define the boundaries of liability," Menger said. "Computing is about execution. Saving is about durability. Delivery is about reliability."
Programmable control points, which use event-based, conditional logic rather than generative AI, enable intelligent traffic management to go beyond simple load balancing. Routing decisions are based on real backend health, using intelligent health knowledge to detect early signs of trouble. This includes monitoring for early signs of trouble. And when problems arise, the system can isolate bad components without removing the entire service.
"An independent, programmable data delivery layer becomes necessary because it allows policy, optimization, security, and traffic control to be applied equally to ingestion and consumption paths without changing storage systems or AI frameworks," Stringfellow said. "By decoupling data access from storage implementation, organizations can safely absorb bursty writes, optimize reads, and protect backend systems from unlimited AI access patterns."
Managing security issues in AI data delivery
AI isn’t just pushing teams to save throughput, it’s forcing them to treat data movement as a performance and security problem, Stringfellow said. Security is no longer an afterthought simply because the data is located deep in the data center. AI introduces automated, high-volume access patterns that must be authenticated, encrypted, and managed with ease. That’s where the F5 BIG-IP comes into play.
"F5 BIG-IP sits directly on the AI data path to provide high-throughput access to object storage while enforcing policy, inspecting traffic, and making payload-informed traffic management decisions," Stringfellow said. "Faster feeding of GPUs is necessary, but not sufficient; Today’s storage teams need confidence that AI data flows are optimized, controlled, and secure."
Why data delivery will define AI scalability
Looking ahead, the requirements for data delivery will only intensify, says Stringfellow.
"AI data delivery will move from bulk optimization towards real-time, policy-driven data orchestration in distributed systems," he said. "Agent-based and RAG architectures require fine-grained runtime control over latency, access scope, and delegated trust boundaries. Businesses must start treating data delivery as programmable infrastructure, not a storage or networking product. Organizations that do this early are faster and have less risk."
Sponsored articles are content produced by a company that pays to post or has a business relationship with VentureBeat, and is always clearly labeled. For more information, contact [email protected].
