F5, a global leader in application delivery, API delivery, and security, has announced a significant strategic step to expand its existing collaboration with NVIDIA. Announced from Dubai, United Arab Emirates, on March 25, 2026, this collaboration aims to accelerate the performance of infrastructure powering AI models during live usage and elevate operational efficiency to unprecedented levels that keep pace with fast-moving digital demands.
Technical integration details and keeping pace with the agentic era
This expanded integration combines advanced networking solutions—specifically F5 BIG-IP Next for Kubernetes—with powerful NVIDIA BlueField-3 data processing units. This innovative fusion aims to form an intelligent infrastructure layer based on real-time operational data analysis. Undoubtedly, this development directly contributes to raising the output processing rate of artificial intelligence models, improving GPU utilization efficiency, and reducing latency. It also provides tremendous support for running multi-user AI platforms within secure, flexibly scalable environments, which is a fundamental pillar for companies seeking to enhance their competitive capabilities in this vital field.
Token economics and new efficiency standards
In generative AI systems, “tokens” represent the basic unit of measurement for model outputs, whether these tokens are words, signals, or pieces of data generated and processed during the inference phase. The size of these tokens and their production speed are critical factors in determining end-user experience quality, the efficiency of the adopted infrastructure, and the financial return achieved from each accelerator processor.
As enterprises and GPU-as-a-Service (GPUaaS) providers continue striving for rewarding returns from AI technologies—transitioning from initial trial phases to delivering actual revenue-generating services—infrastructure efficiency has become one of the most important strategic criteria. Success is no longer measured solely by deployed computing capacity, but by more precise indicators tied to token processing economics, sustainable generation rates, time to first token (TTFT), cost per token, and the return realized from each accelerator.
Smart infrastructure and workload routing
The transition from traditional application-based inference models to sophisticated workflows driven by agentic AI necessarily requires adopting new architectural approaches that enhance token generation rates and reduce operational costs. Herein lies the importance of the BIG-IP Next for Kubernetes solution, which leverages NVIDIA NIM data, runtime environment signals (Dynamo), and GPU performance metrics to make intelligent, inference-aware routing decisions prior to execution.
By aligning workloads with the most appropriate accelerators in real time, this innovative approach helps improve token processing economics, raise sustainable utilization levels, reduce latency, and curb the frustrating need for reprocessing.
«AI infrastructure is no longer limited to providing access to GPUs or expanding their deployment scale; rather, it is anchored on maximizing the economic return per processor. In collaboration with NVIDIA, we are now enabling what are known as AI factories to treat token generation as a measurable business metric. Our innovative solution provides the intelligence and governance frameworks required to increase GPU productivity, lower the cost per token, and scale shared AI platforms with absolute confidence».
— Kunal Anand, Chief Product Officer at F5
Documented results and a qualitative leap in performance
Performance indicators clearly reflect this technological progress. In rigorous tests verified and certified by the independent firm The Tolly Group, the BIG-IP Next for Kubernetes solution, enhanced with NVIDIA BlueField-3 data processing units, achieved a staggering increase in token generation rate of up to 40 percent. It also accelerated time to first token (TTFT) by 61 percent, alongside a 34 percent reduction in total request latency. These positive results are not limited to minor incremental improvements; rather, they represent a clear qualitative leap in overall operational efficiency.
Offloading network workloads to boost productivity
By offloading complex networking tasks, secure TLS encryption, AI-aware load balancing, intensive traffic management, and data handling to NVIDIA BlueField-3 data processing units, the system preserves valuable processing capabilities in the central processing units. This allows GPUs to focus exclusively on their core function—executing inference operations with high efficiency, sustainable generation rates, and at very large scale.
Consequently, this improves GPU utilization levels, reduces wait times, and increases token generation, lowering the cost per token within the existing infrastructure. Most importantly, these major gains were achieved without requiring any code modifications to the models, facilitating direct deployment within existing AI runtime environments. For major enterprises and NeoCloud providers fiercely competing over token economics, this difference marks the dividing line between an infrastructure that restricts AI output and one that accelerates its pace.
«The combination of NVIDIA’s accelerated computing infrastructure and F5’s application delivery and security platform—which is AI-aware—enables advanced levels of token economics within AI factories, while supporting large-scale, cost-effective inference execution without requiring any model modifications. Working closely with F5, we are empowering organizations to scale inference operations within AI factories with exceptional operational and economic efficiency».
— Kevin Dierling, Senior Vice President of Networking at NVIDIA
FAQs
What is the primary goal of the collaboration between F5 and NVIDIA?
The collaboration aims to accelerate infrastructure performance for running AI models during the inference phase and improve GPU utilization efficiency to increase token generation rates and reduce latency.
How does this integration help reduce AI costs?
By offloading networking tasks, encryption, and load balancing to NVIDIA BlueField-3 data processing units, GPUs are freed up exclusively for inference operations, increasing token productivity and lowering the cost per token without needing to modify models.
What documented benefits did tests show for this joint solution?
Tests by The Tolly Group showed the solution achieved a 40 percent increase in token generation, accelerated time to first token (TTFT) by 61 percent, and reduced total latency by 34 percent.
What is meant by the term “token economics” mentioned in the article?
It refers to measuring the efficiency of AI infrastructure based on the cost, speed, and volume of tokens (words or data) produced by the model, alongside the financial return realized from each data accelerator processor.
