iPronics and BSC collaborate on programmable AI networking

Key Takeaways

• iPronics and the Barcelona Supercomputing Center (BSC‑CNS) have signed a 2‑year strategic partnership to fuse programmable optical switching with GPU‑centric AI/HPC workloads.
• The iPronics ONE rack‑ready optical circuit switch (OCS) eliminates the need for custom software stacks, delivering plug‑and‑play integration for hyperscalers.
• Joint development will produce workload‑aware networking that dynamically matches AI traffic patterns (LLM, MoE) to reconfigurable optical paths, cutting latency by up to 30 % and power draw by 15 % in testbeds.
• The collaboration follows iPronics’ $125 M Series B raise (total funding $177 M) led by Maverick Silicon, Light Street Capital, and NVIDIA.
• A new U.S. office in Santa Clara, CA will support deployments and accelerate the rollout of the ONE platform to cloud providers.

iPronics ONE: Plug‑and‑Play Optical Switching for GPU Clouds

What the ONE Platform Delivers

• Form factor: 2U rack‑mount, 10‑port 400 Gb/s optical circuit switch.
• Switching speed: Sub‑microsecond (≈ 800 ns) reconfiguration, enabling real‑time traffic steering.
• Power envelope: 45 W per unit, roughly 15 % lower than legacy electronic packet switches of comparable capacity.
• Software stack: Pre‑installed low‑level drivers, RESTful APIs, and a Python SDK that expose switch state without altering existing orchestration tools (Kubernetes, Slurm, etc.).

Why It Matters for Hyperscalers

Traditional programmable optics require custom firmware and deep integration with the data‑center control plane, a barrier that adds months of engineering effort. The ONE platform’s rack‑ready, zero‑touch approach lets operators retrofit existing GPU clusters—such as NVIDIA H100 or AMD Instinct‑MI250X arrays—while preserving their current software ecosystem.

The BSC‑CNS Collaboration: Merging Hardware with AI‑Centric Software

Joint Development Roadmap

Technical Focus Areas

• Dynamic topology: The switch can rewire 10 × 10 ports on‑the‑fly, allowing the network to morph from a fat‑tree to a directed‑acyclic graph optimized for specific AI communication patterns.
• API hierarchy: iPronics supplies low‑level control (port‑level bandwidth throttling, wavelength assignment), while BSC builds a high‑level orchestration layer that ingests job‑level metadata from AI frameworks (PyTorch, TensorFlow).
• Energy accounting: Early measurements on BSC’s testbed show a 15 % drop in total energy per training epoch when the optical fabric replaces a 100 Gb/s Ethernet spine.

Market Context: How iPronics ONE Stands Against Conventional Solutions

Numbers reflect typical vendor specifications and iPronics public data (2024‑2025).

Strategic Implications for Cloud Providers

1. Higher GPU utilization – By aligning optical paths with AI communication bursts, the system can keep GPUs busy up to 95 % of the time, versus the 70‑80 % ceiling of static Ethernet fabrics.
2. Reduced CAPEX/OPEX – The ONE platform’s modest power draw and rack‑level footprint translate into ≈ $0.02/kWh savings per PB of AI traffic.
3. Future‑proofing – The programmable nature of the switch means it can adapt to emerging AI models (e.g., 1‑trillion‑parameter MoE) without hardware refresh.

Bottom Line

The two‑year alliance between iPronics and the Barcelona Supercomputing Center puts a rack‑ready, low‑latency optical circuit switch into the hands of researchers and cloud operators looking to squeeze every ounce of performance from GPU clusters. By delivering a plug‑and‑play solution that integrates with existing AI stacks, the partnership promises measurable gains—up to 30 % lower latency and 15 % energy savings—while keeping software continuity intact. Backed by a robust $177 M funding base and a new U.S. foothold, iPronics is poised to become a pivotal supplier for the next generation of AI‑focused data centers.