Wiwynn took advantage of Computex 2026 to showcase where the physical infrastructure supporting the next generation of artificial intelligence is headed. The Taiwanese company, specialized in servers, storage, and rack-scale solutions for data centers, presented a proposal that goes far beyond simply adding more GPUs to a rack. Their clear message: AI is no longer just a computing problem, but a system engineering challenge as well.
Wiwynn’s exhibit combines rack-scale AI servers, advanced liquid cooling, optical interconnection, high-voltage direct current (HVDC) power distribution, and high-performance storage for vector databases and graph workloads. It offers a snapshot of a shift already underway in large data centers: power density, heat management, and data movement have become limits as critical as the chips themselves.
From Standalone Servers to AI Factories
Wiwynn places the concept of an “AI Factory” at the core of its proposal. While the idea is not new, it increasingly defines the market: an AI data center isn’t like a traditional facility with generic servers. It’s a tailored infrastructure designed to train, run, and support large-scale models, with complete racks optimized as unified units.
One of the most notable announcements is Wiwynn’s readiness, in collaboration with Wistron, for the NVIDIA Vera Rubin NVL72 platform — a fully liquid-cooled rack-scale system that integrates 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs. According to Wiwynn, this platform is intended for training, inference, and edge reasoning, offering up to a 10x improvement in performance per watt due to the system’s extreme co-design.
The company also showcased an AMD Helios solution based on the OCP ORv3 Open Rack Wide specification, featuring AMD Instinct MI400 GPUs, sixth-generation AMD EPYC CPUs, AMD Pensando networking technologies, and the ROCm software stack. While different from NVIDIA’s approach, it targets the same goal: building open, scalable AI infrastructure capable of large-scale deployment in hyperscale environments.
| Platform | Approach | Key Features |
|---|---|---|
| NVIDIA Vera Rubin NVL72 | Closed, fully liquid-cooled rack-scale system | 72 Rubin GPUs and 36 Vera CPUs |
| AMD Helios | Open architecture based on OCP ORv3 | Instinct MI400 GPUs, sixth-gen EPYC, Pensando, ROCm |
| Storage Prototype | Data-intensive workloads for AI | 96 liquid-cooled E3.S SSDs |
| Optical Scale-up | Optical interconnection for AI | CPO, TeraPHY optical engines, remote light sources |
| HVDC 800 V | High-density power distribution | Power rack, liquid busbar, 800V-to-50V PDB |
The message is clear: the rack has become the new fundamental unit of design. Previously, a company could consider servers, switches, and storage as separate modules. In advanced AI, those boundaries blur. CPUs, GPUs, memory, storage, networking, power, and cooling must be designed holistically.
Electricity and Heat as Part of the Product
A standout aspect of Wiwynn’s proposal is the 800 V HVDC power distribution at the rack level, developed with partners like Delta and TE Connectivity. As power density per rack increases, traditional distribution methods face practical limits. More power means more losses, heat, cabling, complexity, and operational risk.
The architecture includes a power rack, a liquid-cooled busbar, and a distribution board that converts 800 V to 50 V. This design aims to boost efficiency and prepare data centers for much denser AI racks. It’s not a minor detail: in AI infrastructure, stable and efficient power delivery can be as critical as selecting the right GPU.
Cooling remains a major frontier. Wiwynn introduced a dual-sided cold plate for ASICs up to 6 kW, using Liquid Metal TIM and a PCB architecture with a cavity designed for vertical power delivery. The company claims this design shortens power pathways and improves electrical distribution efficiency by over 80% in compact environments. Additionally, employing liquid metal as the thermal interface material could improve thermal efficiency by more than 30%.
It also demonstrated solutions based on diamond composite material—offering high thermal conductivity with lower weight compared to copper. Shown in a microchannel cold plate, this material could have future applications even at the chip encapsulation level. This underscores a clear trend: the future of AI hardware will depend not only on more powerful silicon but also on new materials, rack formats, and innovative heat evacuation methods.
Optical Interconnection: The Next Bottleneck
Wiwynn emphasized optical interconnection through CPO, or co-packaged optics. The reason is straightforward: AI models require moving vast data volumes between accelerators, memory, storage, and networks. While copper has sufficed for years, high-density configurations now face limits related to distance, power consumption, and bandwidth.
The company collaborates with partners like Ayar Labs and GUC to demonstrate an optical-scalable rack design integrating AI ASICs with advanced 2.5D packaging, TeraPHY optical engines, remote light sources like ELSFP SuperNova, fiber routing, and optical interconnects. They also developed a proprietary liquid cooling solution for ELSFP to maintain reliable optical connections within the chassis.
This focus points to a major upcoming shift: optical solutions won’t just connect racks or data halls—they will increasingly move closer to the chip and package. Data movement within an AI factory will become a significant energy and technological cost, and if bandwidth becomes a limiting factor, optical interconnection will transition from an enhancement to a necessity.
| Challenge of Large-Scale AI | Wiwynn’s Technological Response |
|---|---|
| Higher power consumption per rack | 800 V HVDC distribution |
| Increased heat per chip | Dual-sided cold plates and liquid metal |
| Greater thermal density | Diamond composite materials |
| More data movement | Optical interconnection and CPO |
| Higher storage demands | Servers with 96 liquid-cooled E3.S SSDs |
| Deployment complexity | Integrated rack-scale design |
The storage prototype also fits into this vision. Wiwynn presented a server with 96 liquid-cooled E3.S SSDs, designed to provide high IOPS performance and to support NVIDIA GPUs capable of saturating PCIe connections to the drives. It targets workloads like graph databases and vector-based databases—components increasingly relevant in AI applications, semantic search, and RAG (Retrieval-Augmented Generation).
AI Data Center Demands System Co-Design
Wiwynn’s presentation at Computex 2026 reflects a deeper industry shift. For years, infrastructure progress was explained mostly by incremental improvements: faster processors, more memory, better networking, or denser storage. AI demands a different approach. When a rack can consume tens or hundreds of kilowatts, everything becomes interconnected.
Thermal design impacts sustained performance. Power distribution influences density. Interconnection limits scalability. Storage determines whether models and agents can access data quickly enough. Liquid cooling moves from exotic option to core design element. ODM manufacturers like Wiwynn gain importance because they turn reference designs into deployable large-scale systems.
William Lin, President and CEO of Wiwynn, sums it up: “AI is no longer just a computing challenge, but a complete system engineering challenge.” This idea helps explain why Computex has become an increasingly important showcase for AI infrastructure. Many innovations are not just about new models or benchmark results—much of the real innovation is in how the hardware is packaged, powered, connected, and cooled.
For data center operators and cloud providers, these technologies have direct implications: legacy facilities may not be ready for modern AI density. Upgrading power, cooling, power distribution, and internal networking can be as costly as acquiring new servers. That’s why new rack-scale designs are about not only performance but also physical feasibility.
Wiwynn’s approach isn’t just a single product but a roadmap showing where AI data centers are headed: full racks, liquid cooling, HVDC power, integrated optics, and AI-specific storage. The resulting infrastructure is more complex, but also more aligned with the demands of upcoming models, agents, and applications.
Frequently Asked Questions
What did Wiwynn present at Computex 2026?
Wiwynn showcased rack-scale solutions for AI data centers, including NVIDIA Vera Rubin NVL72 platforms, AMD Helios, advanced liquid cooling, CPO optical interconnects, high-performance storage, and HVDC electrical distribution.
Why is liquid cooling important for AI?
Because AI accelerators and ASICs achieve very high power consumption. Liquid cooling effectively cools heat sources and supports higher density racks.
What is HVDC in data centers?
HVDC stands for high-voltage direct current power distribution. In this case, Wiwynn demonstrates an 800 VDC solution for high-density AI racks.
What role does optical interconnection play in AI?
Optical interconnection helps transfer large data volumes with higher bandwidth and lower power compared to copper, especially at high densities. It’s crucial for scaling high-density AI clusters.
via: prnewswire