At the Open Compute Project (OCP) Global Summit in San Jose, Wiwynn Corporation announced a suite of next‑generation AI infrastructure components, including a liquid‑cooled NVIDIA Vera Rubin rack, an AMD‑based Helios system, novel diamond‑based microchannel coolers, and a co‑packaged optics (CPO) rack solution. The showcase targets the growing demand for higher compute density and lower energy cost in large‑scale generative AI and machine learning workloads.
New AI rack platforms for dense training workloads
Wiwynn’s two flagship rack designs address the divergent needs of hyperscale cloud providers. The first, built around NVIDIA’s Vera Rubin NVL72 reference, integrates 72 Rubin GPUs and 36 Vera CPUs on a single chassis. According to Wiwynn, the platform uses 6th‑generation NVLink to deliver up to 1.2 petaflops of mixed‑precision performance while keeping power draw under 120 kW per rack. The second platform, branded “Helios,” follows the Open Rack Wide (ORW) specification and pairs AMD Instinct MI450 GPUs with 6th‑gen EPYC CPUs, Pensando networking ASICs, and ROCm software stack. Wiwynn positions Helios as a more open‑source‑friendly alternative that can be customized for specific cloud‑service‑provider (CSP) requirements.
Thermal engineering beyond copper
Cooling has become the primary bottleneck as AI accelerators push thermal design power (TDP) beyond 500 W per chip. Wiwynn demonstrated three experimental approaches:
- Diamond‑based microchannel cooling: By bonding a synthetic‑diamond composite directly to silicon, the team claims a thermal conductivity of 900 W/m·K, roughly three times that of copper, shortening heat‑transfer paths at the die level.
- 6 kW double‑sided cold plate with cavity PCBA: The design incorporates vertical power delivery (VPD) to reduce IR losses by up to 80 % and improve coolant flow uniformity, delivering a 15 % thermal performance gain over conventional cold plates.
- Lightweight inner‑manifold and tubing: Using high‑strength aluminum alloys and polymer‑based manifolds, the solution reduces cooling‑system mass by 30 %, easing rack installation and shipping logistics.
Wiwynn collaborated with Corintis on a topology‑optimized microchannel pattern that matches chip power maps, a technique that could become standard as ASICs grow more heterogeneous.
Co‑packaged optics for rack‑scale bandwidth
To address the “memory wall” that limits data movement between accelerators, Wiwynn unveiled a rack‑level CPO solution. The system integrates optical transceivers directly onto the motherboard, reducing electrical‑to‑optical conversion steps. Partnering with Ayar Labs, Corning, GUC, and Senko, the demo featured live telemetry of optical engine health, dynamic wavelength allocation, and a modular patch‑panel architecture that supports multi‑vendor fiber management. Wiwynn argues that this approach can sustain 400 Gb/s per link without the latency penalties of traditional copper backplanes.
Why these advances matter now
Generative AI models such as GPT‑4‑style LLMs routinely require clusters of thousands of GPUs, and the cost of power and cooling can exceed 30 % of total operating expense. Gartner forecasts the worldwide AI software software market to reach $126.5 billion in 2026, while IDC projects AI‑related infrastructure spending to surpass $150 billion the same year. Innovations that shrink the power‑to‑performance ratio directly improve TCO for CSPs and enable more aggressive scaling of model size.
Competitive context and ecosystem impact
Major cloud operators have already begun deploying Nvidia H100‑based racks and AMD MI200 systems. Wiwynn’s claim of being among the first “pilot build partners” for the Vera Rubin platform suggests early access to NVIDIA’s silicon roadmap, a potential differentiator against rivals like Supermicro and Quanta. However, the company’s thermal prototypes remain at the demonstration stage; mass‑production readiness, reliability data, and cost per rack have not been disclosed.
Practical applications and potential constraints
Enterprises seeking on‑prem AI clusters for confidential workloads could benefit from the integrated cooling and optics, especially where data‑center floor space is limited. The diamond‑based cooling approach may also appeal to edge deployments with strict acoustic and thermal envelopes. On the downside, the specialized materials and custom cold‑plate designs could raise upfront CAPEX and require new maintenance expertise, potentially slowing adoption among operators that favor commodity‑grade solutions.
Implications for AI software stacks
The Helios platform’s reliance on AMD’s ROCm stack aligns with a broader industry push toward open‑source drivers and toolchains, potentially easing integration with frameworks like PyTorch and TensorFlow that have added ROCm support. Conversely, the NVIDIA Vera Rubin rack will likely continue the dominance of CUDA‑based ecosystems, reinforcing the split between CUDA‑centric and ROCm‑centric development pipelines. Developers may need to abstract hardware dependencies to maximize portability across Wiwynn’s dual‑vendor offerings.
Future outlook
Wiwynn plans to extend its CPO technology into a 2‑U form factor and explore hybrid liquid‑air cooling for mid‑range AI workloads. If the company can translate its prototype performance into reliable, cost‑effective production, it could influence the next wave of AI infrastructure that prioritizes density, energy efficiency, and seamless high‑speed connectivity.
Market Landscape
The AI infrastructure market is accelerating as generative models demand ever‑larger compute clusters. IDC predicts AI‑related spending will exceed $150 billion in 2026, driven largely by hyperscale data centers expanding GPU density. Companies such as NVIDIA, AMD, and Intel are competing on both silicon performance and system‑level integration. While NVIDIA dominates the high‑end GPU segment, AMD’s open‑source ROCm strategy is gaining traction among cost‑sensitive operators. Co‑packaged optics, pioneered by Facebook’s Open Compute Project and now adopted by vendors like Cisco and Dell, aim to overcome the bandwidth ceiling of traditional backplanes. Wiwynn’s entry into this space reflects a broader industry shift toward rack‑scale optical interconnects, which IDC expects to capture 15 % of total AI data‑center bandwidth by 2027.
Top Insights
- Wiwynn’s AI racks combine liquid cooling and co‑packaged optics to cut power‑to‑performance costs, a key factor as AI workloads hit 500 W per accelerator.
- The diamond‑based microchannel cooler promises three‑fold higher thermal conductivity than copper, but its manufacturing scalability remains unproven.
- Helios offers an open‑source‑friendly AMD stack, potentially lowering software licensing fees for cloud providers that favor ROCm over CUDA.
- Co‑packaged optics could sustain 400 Gb/s links, addressing the bandwidth bottleneck that hinders multi‑GPU scaling in large LLM training.
- Adoption risk centers on the cost and service complexity of custom cooling solutions versus the industry’s preference for commodity hardware.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
