Here’s how a global cloud provider improves data delivery with F5

Hunter SmitSenior Manager, Product Marketing | F5

Object storage was designed to be an affordable, durable place to store data, not to be the busiest service in the data center. Until the AI era, that distinction didn't matter much because the traffic hitting it, things like backups, archives, media, and application assets, was forgiving. Then model training, fine-tuning, and retrieval-augmented generation (RAG) workloads arrived, and the read/write patterns changed from occasionally retrieve to continuously saturate.

That shift creates an economics problem. Your GPU cluster is likely the most expensive asset on your infrastructure balance sheet, and it earns nothing while it waits for data. Every stall in the storage path is amortized across hardware that costs more per hour than almost anything else you run, which means the layers have inverted. The delivery tier in front of your storage now determines the return on your most expensive compute, and that path is no longer plumbing so much as a product in its own right.

What one hyperscale storage operator did about it

One of F5’s customers, a global cloud provider, runs an S3-compatible object storage service across dozens of data centers. The service is the primary repository for unstructured customer data and, increasingly, as the feeding ground for those customers' model-training pipelines. Every read and write flows through a load-balancing tier before it ever touches an indexing or storage node, and that tier, built on Citrix NetScaler appliances, had become the constraint.

The requirements weren't exotic. They were ordinary requirements made brutal by the economics of meeting them at that scale. The provider needed to terminate and re-encrypt TLS at line rate, route at Layer 7 on S3 bucket name, and enforce rate limits consistently across an active-active cluster, applying both global limits at the service level and targeted limits per bucket so that one tenant's ingest job can't starve everyone else's. It also needed global server load balancing across sites and FIPS support for regulated tenants. Multi-tenant fairness, it turns out, is a delivery-tier policy problem, because by the time traffic reaches the storage nodes the damage is already done.

The provider replaced its legacy application delivery controller (ADC) estate with F5 BIG-IP deployed on F5 rSeries. TLS terminates and re-encrypts at line rate. Requests route at Layer 7 on the S3 bucket name, and bucket-level policy is now defined in one place rather than reproduced across the fleet. Rate limits are defined once and applied uniformly across the cluster rather than hand-tuned per box. Global server load balancing steers traffic across sites, and regulated tenants keep FIPS coverage.

The hardware changed as much as the policy implementation did. The new estate occupies an 80% smaller physical footprint, carries 2.4x the rated throughput, draws roughly a fifth of the power per rated gigabit, and leaves 222 fewer devices to patch, certify, and troubleshoot. Headroom is normally bought with the two things a data center has least of, watts and rack units. Here it arrived alongside a reduction in both, which makes the recovered capacity an option rather than a purchase. In a facility where AI capacity is bought in power and space, the provider now holds more of both without having committed in advance to what it will spend them on.

This is what F5 means by AI data delivery. The path between storage and compute is a boundary, and a control point inserted there is what makes reachability, policy, and delivery resilience enforceable. It means treating that path as engineered control and data planes with their own performance, security, and governance requirements rather than an afterthought bolted to the storage array.

Why this matters even if you never train a model

Almost nothing in this customer’s situation is specific to machine learning or training frontier models, and that's the point. Because S3 rides on HTTP, every discipline F5 spent three decades building for web traffic, from intelligent health-aware load distribution and protocol inspection to encryption offload and per-tenant throttling, applies directly to object storage the moment you put a capable proxy in front of it. The same front door that keeps a training pipeline stable and saturated also keeps backup windows predictable, analytics queries steady, and a noisy application from degrading storage for everyone else. AI data delivery is the workload that finally made the front door impossible to ignore.

Three questions for whoever owns your storage

As you evaluate your own organization’s storage path, here are three questions to consider:

1. When a storage node degrades, what reroutes traffic, and how fast? If the answer is that the client retries, you don't have a delivery tier, you have a delivery architecture based on hope.
2. Can you rate-limit a single tenant or bucket without touching the storage cluster itself? If you can't, your fairness policy is a suggestion.
3. How much of your storage refresh budget assumes the front door stays fixed? This provider found that the delivery tier, not the storage behind it, was where the leverage was hiding.

The takeaway is that capacity and delivery are different decisions. Capacity is something you buy. Delivery is something you architect. As data pipelines become the workloads your business depends on most, the teams pulling ahead are the ones treating the path to storage with the same rigor they've always applied to the storage itself.

F5’s focus on AI doesn’t stop here—explore how F5 secures and delivers AI apps everywhere.

About the Author

Hunter SmitSenior Manager, Product Marketing | F5

Hunter Smit is a senior manager of product marketing for solutions and strategy at F5. He leads solutions marketing across the F5 portfolio, including AI infrastructure and application security, and is driving F5’s entry into the emerging AI data delivery market. Hunter is a member of the Whitworth University School of Business Advisory Board and has served as an adjunct instructor for product marketing. He earned an MBA and a Bachelor of Business Administration degree in management from Whitworth University.

More blogs by Hunter Smit