Building A Production-Ready Optically Connected Rack For AI Scale-Up

Why It Matters

The move to CPO eliminates bandwidth bottlenecks and latency penalties of copper, enabling hyperscalers to expand AI clusters cost‑effectively. This partnership proves that high‑density, liquid‑cooled optical racks can be manufactured at scale, accelerating AI infrastructure deployment.

Key Takeaways

• Ayar Labs and Wiwynn partner to deliver CPO rack‑scale AI systems.
• Each rack supports 64 ASICs, 1,024+ accelerators across 16 racks.
• Liquid‑cooled architecture reduces cooling energy and handles >100 kW per rack.
• Co‑packaged optics provide >100 Tbps per accelerator with ultra‑low latency.
• Design includes serviceable fiber routing and ELSFP modules for hyperscale ops.

Pulse Analysis

The rapid growth of generative AI models has driven data centers to assemble ever‑larger GPU clusters, exposing the physical limits of traditional copper interconnects. Copper’s signal degradation beyond a few meters and its rising power draw make it unsuitable for the multi‑rack fabrics that will soon host thousands of accelerators. Co‑packaged optics (CPO) integrates photonic transceivers directly onto silicon, delivering terabit‑scale bandwidth with microsecond‑level latency, and positioning itself as the only viable path for next‑generation AI and high‑performance computing (HPC) infrastructures.

Ayar Labs and Wiwynn’s collaboration tackles the engineering challenges of bringing CPO to production. Their reference rack combines two AI ASICs per tray, a CPU, and a high‑voltage DC power architecture, all cooled by a 100% liquid‑cooled cold‑plate system capable of dissipating over 100 kW per rack. The design incorporates external laser small form‑factor pluggable (ELSFP) light sources, advanced fiber‑management modules, and serviceable connectors that mimic existing pluggable optics workflows, ensuring hyperscalers can maintain familiar operational procedures while gaining optical performance.

For the market, this partnership signals a shift from incremental copper upgrades to a wholesale optical redesign of AI data centers. By validating thermal, mechanical and serviceability aspects at OFC 2026 and the upcoming Computex showcase, Ayar Labs and Wiwynn provide hyperscalers with a proven, manufacturable blueprint that can be deployed at scale. The ability to interconnect over 1,000 accelerators with sub‑nanosecond latency and terabit bandwidth will lower total cost of ownership, accelerate AI model training cycles, and set a new standard for future AI infrastructure investments.