iPronics and the Barcelona Supercomputing Center have started a two-year collaboration to test programmable optical networking inside a research GPU and high-performance computing environment. In its September 21 announcement, iPronics said its ONE platform will be integrated with BSC infrastructure so the partners can study how network behavior can respond to AI workloads.
The project targets a problem that becomes harder as accelerator clusters grow: the processors may be fast, but the links between them can leave capacity idle. iPronics describes ONE as a rack-ready optical circuit switching platform, while BSC will develop software above the switch-management layer. Together, the teams plan to connect application behavior with reconfigurable optical links instead of treating the network as a fixed layer beneath the compute.
Why the interconnect is becoming part of the AI system
Large language models and mixture-of-experts systems do not keep all of their work inside one accelerator. They repeatedly exchange model state, activations, routing decisions, and other data across a cluster. That makes communication patterns part of the workload, particularly when different jobs create different bursts of traffic. A static network can be provisioned for a peak that is rarely sustained, or it can force competing jobs to share paths that do not match what they are doing at that moment.
Optical circuit switching approaches the problem by changing which endpoints are connected as demand changes. The goal is not simply to replace every electrical link with an optical one. It is to give the software managing the cluster another way to shape traffic, while keeping the compute environment and its existing software stack in place. iPronics says the BSC work will examine dynamic, workload-aware control across the platform’s hardware and software, with the intended benefits of higher GPU utilization, lower latency, and lower energy consumption.
That places the announcement in the same wider hardware story as Huawei’s optical approach to its Atlas 960E SuperPoD and the optical connectivity work in Qualcomm and Amazon’s custom AI infrastructure partnership. Those efforts are not interchangeable products, but they show why networking, memory, and packaging are now central to the performance of large AI systems. Our earlier report on Avicena’s 1 Tbps optical links covers the same pressure from a different point in the rack.
A research collaboration, not a finished product launch
The iPronics and BSC announcement sets out a testbed and an engineering program rather than a customer deployment timetable. iPronics will provide the programmable optical switching platform, low-level software, and application programming interfaces. BSC will build software above the management layer and use its research-class GPU and HPC infrastructure to explore how the pieces behave together.
The work will cover AI training and inference, including large language model and mixture-of-experts workloads. That scope matters because a useful result has to account for more than the raw switching hardware. The system must observe workload behavior, decide when a topology change is worthwhile, and make that change without turning the network into a new operational bottleneck. The partners say their experiments will produce system-level insight for future AI infrastructure design, but the announcement does not publish benchmark results from the joint environment.
iPronics said the collaboration follows its recent expansion in AI infrastructure, including a Series B financing round and a new United States office. Those details frame the BSC project as part of the company’s commercial development, but they do not establish that the platform is already deployed across hyperscale production fleets. The immediate news is narrower and more useful: a supercomputing center is giving programmable optical networking a real GPU and HPC environment in which its control model can be tested.
If the approach can improve utilization without forcing operators to replace their software stack, optical switching could become a practical part of the next generation of AI clusters. The BSC program should provide evidence about that question, but the September announcement is a starting point, not a performance verdict.
