Delivering the bandwidth density and efficiency needed to scale AI compute clusters to 1,000 accelerators.
By Nandita Aggarwal and Nicholas Chang
As AI models drive compute demand, servers keep getting bigger. Rack‑scale AI systems (such as the 72-GPU systems from NVIDIA or AMD) enable many GPUs to work together through system-level optimization. They push beyond the limits of single-chip performance and meet the soaring compute needs of the AI era.
But this is just the beginning. The next step is to link racks together and unlock shared efficiencies to achieve the best total cost of ownership for data centers. These AI compute clusters will soon be scaling beyond 1,000 accelerators.
To fully utilize these accelerator clusters, high bandwidth and low latency are needed. That is why next‑generation AI systems are increasingly shifting to co-packaged optics (CPO). Traditional copper interconnects increasingly constrain performance, system growth, and power efficiency. At the speed these interconnects operate, copper can only reach a few meters without signal degradation. Conventional pluggable optics, found in data centers today, are too inefficient to be viable as scale-up fabrics push toward 1,000+ accelerators and ever-higher bandwidth density.
CPO is emerging as the only solution capable of delivering the bandwidth density and efficiency that multi-rack scale-up demands.
According to Kinny Chen, senior manager at Wiwynn, “To fully utilize these accelerator clusters, you need both extremely high bandwidth and very low latency. Co‑packaged optics significantly reduces latency and enables much higher bandwidth. CPO is no longer just a ‘nice‑to‑have’ technology – it’s becoming essential for next‑generation AI and HPC systems.”
Building such a rack-scale system adds many new design challenges.
As Vivek Gupta, Ayar Labs’ chief strategy officer, explained, “Scale-up networks are going from 72 GPUs to 576, which is where the copper topology breaks. Retimers and copper cannot physically handle the speed needed to connect 576 GPUs. This is the perfect moment for CPO technology to improve connection bandwidth and reduce connection power, leading to a paradigm shift in the way GPUs are connected.”
Cooling becomes a significant concern, and fiber management can be incredibly complex. Potential customers need proof that these new, unfamiliar optical components function as intended in real-world conditions. And they expect these new systems to be as serviceable and manageable as their current ones.
This is why Ayar Labs and Wiwynn recently announced a partnership to bring CPO to rack-scale AI systems that support next-generation hyperscale workloads. By combining Ayar Labs’ CPO solution with Wiwynn’s rack-level system design and manufacturing capabilities, we can build systems that are not constrained by the bandwidth and reach limitations of copper. Just as importantly, our collaboration focuses on the practical requirements, like cooling, serviceability, and fiber management, needed to deploy CPO in production.
Fig. 1: Ayar Labs and Wiwynn’s collaboration features an optically connected rack with integrated CPO technology for rack-scale AI infrastructure.
Wiwynn brings over a decade of experience delivering rack-level IT solutions to top cloud service providers, with strong expertise in board design, system integration, and high-volume L10/L11 rack delivery. Wiwynn has shipped general and AI servers to more than 750 data centers worldwide, backed by manufacturing operations in Taiwan, the U.S., Mexico, Malaysia, and the Czech Republic.
Ayar Labs and Wiwynn announced their partnership at the Optical Fiber Communication (OFC) Conference in March 2026. We demoed a complete rack-scale AI infrastructure reference design, from the L10 chassis to the L11 rack. The solution incorporates a 100% liquid-cooled architecture optimized for high-power operation, including support for external laser small form factor pluggable (ELSFP) light sources, advanced fiber management, and serviceable system designs required for hyperscale environments. It also incorporates a high-voltage DC (HVDC) power architecture to support the power and scalability demands of next-generation accelerators.
We’re sharing additional details here to demonstrate that any design challenges are addressable in real-world operating conditions for hyperscalers. To see this innovation in action, Ayar Labs and Wiwynn will be showcasing the full rack-scale infrastructure at the Computex event in Taipei next month.
At the L10 level, we bring together high-power AI ASICs, co-packaged optics, advanced fiber management, ELSFP light sources, and cold-plate liquid cooling into one integrated design.
Each compute tray combines two AI ASICs with a CPU, and we use a high-voltage DC power architecture to support the power and scalability needs of next-generation accelerators. Each accelerator can deliver more than 100 Tbps of optical connectivity. They consist of TeraPHY optical engines and SuperNova light sources.
The rack can support up to 32 AI compute trays, linking 64 ASICs together through optical interconnects. 16 racks are connected into an AI scale-up cluster supporting 1,024+ accelerators. We use high‑bandwidth, ultra‑low‑latency optical links between AI accelerators across systems and racks. This lets accelerators in different racks communicate as if they are in the same system, enabling efficient scale‑up to thousands of GPUs.
Chen noted, “Overall, this platform is purpose-built for co-packaged optics and high-density accelerator scale-up, with all the key supporting technologies designed from day one.”
Fig. 2: A L10 chassis designed and manufactured by Wiwynn that features Ayar Labs co-packaged optics and ELSFP light sources, as well as AI ASICs, advanced fiber management, and cold-plate liquid cooling.
Liquid cooling is a vital part of this design. High-powered AI racks use above 100 kW per rack, reaching power densities that challenge the limits of air cooling. All those high-wattage GPUs, CPUs, and memory components crammed into a small space create concentrated hot spots. Fans are not enough to cool them sufficiently.
According to Nandita Aggarwal, senior staff engineer for laser product development at Ayar Labs, “As bandwidth requirements increase, we’re hitting the practical limits of air cooling. Liquid cooling brings coolant closer to heat-generating components, which removes heat much more efficiently.”
Liquid cooling offers 3,000x the volumetric heat capacity of air cooling in a smaller, more compact form factor. This reduces the overall energy consumption for data center cooling infrastructure and enables consistent, high-powered operations. This type of compact, efficient cooling system will be essential for future AI systems.
Liquid cooling introduces new design challenges, such as deciding what to cool first and how to route the water to the front of the chassis. Wiwynn’s engineering talent was invaluable in solving these problems.
Validating the mechanical and thermal performance of this design is also important. Photonics components are sensitive to temperature, and it has been unclear how well they would function in real-world environments. Photonics components and ELSFP are highly sensitive to vibration and alignment, so the system must balance thermal performance with mechanical design, field serviceability, and a path to manufacturing automation. This demonstration provides a design reference for future high-density CPO systems.
Nicholas Chang, technical supervisor at Wiwynn, said, “Photonics components and ELSFP are very sensitive to temperature, vibration, and alignment. When paired with liquid cooling architecture, balancing thermal performance with robust mechanical design, field serviceability, and automation brings challenges and is key to design decisions.”
He added, “The Wiwynn team is experienced in system integration and liquid cooling architecture, while the Ayar Labs team has expertise in photonic components and ELSFP. Together, we can quickly create a solid design and validate the performance smoothly and with confidence.”
Fig. 3: A liquid-cooled high-power ELSFP demonstration.
This demonstration is one of the first times liquid-cooled light sources are being used in a rack architecture. Specifically, the demo showcases a liquid-cooled, high-density ELSFP board design in a “belly-to-belly” placement. The cold plate is designed to support OIF-spec high-power ELSFP with full serviceability. With this demonstration, we aim to validate mechanical and thermal performance and provide a deployment-ready design for high-density CPO systems soon.
“By combining our optical technology with Wiwynn’s system expertise, we are delivering a pre-validated, liquid-cooled AI architecture that simplifies next-gen connectivity and speeds time-to-market,” Aggarwal explained.
Because data centers already have deep operational familiarity with pluggable module handling, ELSFP supports a service model that matches how operators manage optics today while enabling CPO performance at the accelerator.
As one might imagine, 1,024 accelerators, with 8 optical engines each and 32 fiber cables per engine, add up fast. At this scale, fiber routing and service operations must be designed in from the start and not layered on after the fact.
“Hyperscalers like the promise of CPO, but they want proof that reliability and serviceability considerations to run this technology in data centers are met in the form factors they understand,” Gupta added.
Our design prioritizes clean routing, predictable bend-radius management, and service workflows that make it easier to unplug and replace cables if they fail. The ELSFP light sources can be unplugged in the same way without system downtime. This kind of serviceability is non-negotiable for large-scale production deployment of CPO.
As the ecosystem matures, detachable connector options can further support serviceability and fiber management at rack scale. Ayar Labs is vendor-agnostic and works across the ecosystem to support multiple connector options based on customer requirements and deployment needs.
“AI infrastructure is outgrowing the limits of copper, and hyperscalers need a fundamentally new approach to scale,” said Mark Wade, CEO and co-founder of Ayar Labs. “Optically connected racks eliminate the interconnect bottleneck and unlock the next order of magnitude in performance and efficiency.”
This partnership shows that CPO can be engineered into rack-scale systems with the thermal, mechanical, and operational characteristics hyperscalers require to bridge the gap between silicon-ready innovation and system-ready deployment. Since hyperscalers tend to procure and deploy infrastructure at rack- and cluster-scale, we must validate cooling, fiber routing, serviceability, and a clear path to manufacturable integration.
“CPO brings a whole set of new mechanical and thermal challenges,” said Chen. “You’re dealing with sensitive optical components that need precise alignment, proper heat management, and clean fiber routing—all inside a dense, high‑power system. Our systems engineering team connects the optical technology to the realities of the rack. That’s where our expertise really stands out.”
—Nicholas Chang is a technical supervisor at Wiwynn.
 |
 |
 |
 |
 |  |
 |  |
 |
Leave a Reply