Key Takeaways

• Ayar Labs and Wiwynn are moving co-packaged optics from component-level innovation toward rack-scale AI system deployment.
• The joint architecture integrates TeraPHY optical I/O chiplets and SuperNova remote light source into Wiwynn’s rack platforms, targeting >100 Tbps per accelerator and clusters exceeding 1,000 accelerators.
• The partnership directly addresses copper interconnect limits in bandwidth, power, and reach—now a primary bottleneck in hyperscale AI clusters.
• The announcement signals a broader shift: optical I/O is transitioning from R&D to system-level commercialization readiness.

Ayar Labs and Wiwynn have announced a strategic partnership to deliver optically connected, rack-scale AI systemsfor next-generation hyperscale workloads. Instead of focusing only on optical engines or chip-to-chip links, the partnership is aimed at how those technologies can be integrated into manufacturable, liquid-cooled, rack-scale platforms.

As AI clusters grow larger, the limiting factor is increasingly not just compute performance but how efficiently thousands of accelerators can communicate across boards, racks, and entire systems. Ayar Labs and Wiwynn are explicitly positioning their collaboration as a response to the bandwidth, reach, and power constraints of copper interconnects, which have become harder to manage as hyperscale AI deployments expand.

From Optical Components to Rack-Level Integration

According to the companies, the joint solution integrates Ayar Labs’ TeraPHY optical engines, powered by the SuperNova remote light source, into Wiwynn’s rack-level architecture for next-generation AI data centers. The stated focus is not just faster connectivity, but solving the deployment issues hyperscalers face in practice, including optical fiber management, integration of CPO-enabled AI ASICs, thermal management, power efficiency, and manufacturability.

Ayar Labs says the new optically connected infrastructure is designed to scale to 1,024 AI accelerators and beyond, with more than 100 Tbps of optical connectivity per accelerator, enabling thousands of accelerators to operate as a unified system across multiple racks. The announcement also says the design incorporates a liquid-cooled architectureoptimized for high-power operation, along with support for ELSFP remote light sources and serviceable system designs suited to hyperscale environments.

Traditional copper interconnects become increasingly difficult to scale as AI workloads demand more bandwidth, longer reach, and better power efficiency. Ayar Labs is pitching optical I/O as a way around that bottleneck. In its own product materials, the company says the combination of TeraPHY and SuperNova can deliver 5x to 10x higher bandwidth, 10x lower latency, and 4x to 8x better power efficiency compared with conventional interconnect approaches based on pluggable optics and electrical SerDes. Ayar also says SuperNova can provide up to 16 wavelengths and support up to 16 ports.

The question is becoming less about whether one chip is faster than another, and more about whether the overall system can move enough data efficiently across increasingly dense, power-hungry clusters. This partnership is part of that shift. The inference here is based on the companies’ emphasis on rack-level integration, liquid cooling, and manufacturability, not just optical performance.

Why Wiwynn’s Role Gives the News More Weight

Wiwynn’s presence makes the announcement more credible from a deployment standpoint. The company says it has shipped general and AI servers to more than 750 data centers worldwide, supported by manufacturing operations in Taiwan, the United States, Mexico, Malaysia, and the Czech Republic. The Ayar Labs announcement also highlights Wiwynn’s capabilities in board design, system integration, and high-volume L10 and L11 rack delivery.

Its financial scale also reinforces that point. Wiwynn reported NT$950.663 billion in fiscal 2025 revenue, up 163.7% year over year, and said AI-related products contributed more than half of revenue. The company also said it is continuing to invest in computing, thermal, and power-efficiency technologies while expanding global production capacity, including a U.S. factory ramp.

Just days earlier, on March 3, 2026, Ayar Labs announced a $500 million Series E round led by Neuberger Berman. The company said the funding will be used to scale high-volume production and test capacity, expand global operations including its Hsinchu, Taiwan office, strengthen ecosystem partnerships, and accelerate deployment of its co-packaged optics solution. Ayar said the new round brought its total funding to $870 million and its valuation to $3.75 billion.

Aligned With Broader Industry Direction

At GTC 2026, Wiwynn showcased NVIDIA Vera Rubin NVL72 infrastructure and emphasized its role in rack-scale integration, accelerated computing, storage, and liquid cooling. Wiwynn described customers as needing integrated rack-scale solutions that harmonize compute, storage, networking, and liquid cooling.

NVIDIA has also been pushing photonics more aggressively. In March 2025, NVIDIA unveiled Spectrum-X Photonics and Quantum-X Photonics networking switches, saying they are intended to connect millions of GPUs across sites while delivering 3.5x energy savings and 10x resilience in AI factories. NVIDIA’s Vera Rubin NVL72 platform itself is described as a rack-scale AI supercomputer built around tightly integrated compute and networking.

Taken together, these announcements suggest that optical interconnects are becoming part of the mainstream AI infrastructure roadmap. That is an inference, but it is a well-supported one based on the direction multiple vendors are now signaling.

What the Industry Should Anticipate Next

The next phase to watch is whether partnerships like this progress from architecture demonstrations and ecosystem positioning into named platform wins, deployment timelines, and production ramps. So far, the public announcement does not identify a hyperscale customer, commercial shipment volumes, or a specific accelerator customer program. What it does show is that the industry is now treating co-packaged optics as a rack-scale system design issue rather than only a chip-to-chip interconnect experiment.

Optical I/O adoption at rack scale, liquid-cooled system architectures, higher-density power delivery, serviceability, and the manufacturing readiness of optical-enabled AI systems. If these areas continue advancing together, co-packaged optics could become a more practical option for next-generation AI infrastructure rather than a niche photonics story.

The next phase of AI infrastructure may depend as much on interconnect architecture, cooling, and manufacturability as on silicon performance alone. Slowly the ecosystem is preparing for a future in which rack-scale AI performance is shaped by how well vendors integrate optics, thermals, networking, and power into a deployable platform.

Need help source components for high-performance computing, networking, thermal, or power infrastructure programs? 

Connect with IBS Electronics to support your next design cycle with supply chain insight, component sourcing support, and global procurement solutions.