The Agentic Shift: ASRock Rack's Liquid-Cooled AI Future

SAN JOSE, Calif. – October 12, 2026 – The engines powering the global economy are undergoing a structural rewrite. For the past three years, the technology sector has been singularly focused on generative AI—models that take a prompt and return an output. But as the 2026 OCP Global Summit kicks off today in San Jose, the hardware on display signals a decisive pivot toward a fundamentally different computing paradigm: agentic AI.

Agentic AI systems do not simply answer questions; they operate autonomously. They spin up code sandboxes, execute multi-step reasoning, run continuous reinforcement-learning environments, and manage complex data pipelines without human intervention. This shift requires a radical reimagining of data center architecture, moving away from traditional x86 bottlenecks and air-cooled servers toward purpose-built, liquid-cooled silicon ecosystems.

At the forefront of this physical transformation is ASRock Rack Inc., which today unveiled its expanded AI server portfolio at booth D37. Featuring the NVIDIA Vera CPU and the NVIDIA HGX Rubin NVL8 platform, the company’s latest hardware highlights exactly how structural changes in AI software are dictating the physics of tomorrow's data centers.

Engineering for Agentic AI: The End of Static Inference

To understand the hardware ASRock Rack is showcasing, one must first understand the specific computational demands of agentic AI. Traditional CPUs struggle with the irregular, control-heavy, and latency-sensitive workflows required by autonomous agents. When an AI agent must continuously evaluate its own outputs, query external databases, and adjust its parameters in real-time, the memory bandwidth of standard DDR5 platforms becomes a fatal chokepoint.

ASRock Rack is addressing this directly with the introduction of the 2UXGM-VERA2, a modular system engineered specifically for autonomous execution. The server is built around dual NVIDIA Vera CPUs, a processor purpose-built for agentic AI and reinforcement learning. Featuring 88 custom Olympus cores on a monolithic die, the Vera CPU is designed to deliver high single-thread performance to ensure software environments and evaluation loops complete rapidly.

Crucially, each Vera CPU supports up to 1.5 terabytes of LPDDR5X memory, delivering up to 1.2 terabytes per second of aggregate memory bandwidth. This allows the 2UXGM-VERA2 to manage thousands of concurrent agent sandboxes without the die-hop latency that plagues traditional chiplet-based designs.

Alongside the CPU-heavy workflows, agentic AI requires massive-scale reasoning models that demand unprecedented GPU interconnectivity. To that end, ASRock Rack is displaying the 2U16X-GNR2/DLC Rubin. This platform incorporates the NVIDIA HGX Rubin NVL8 architecture, integrating eight NVIDIA Rubin GPUs via sixth-generation NVLink 6 interconnects. The result is 28.8 terabytes per second of NVLink bandwidth within the server and 176 terabytes per second of HBM4 memory bandwidth, eliminating the data bottlenecks that throttle multi-step generation and long-context reasoning.

The Thermal Reality: Liquid Cooling as a Baseline

As compute density scales to meet the demands of agentic AI, the laws of thermodynamics are forcing a hard pivot in data center design. The industry is currently transitioning to the NVIDIA Rubin Ultra generation, where power consumption per rack is projected to reach approximately 225 kilowatts—a massive leap from the roughly 150-kilowatt envelope of previous GB300 systems. At these densities, traditional air cooling is not just inefficient; it is physically obsolete.

ASRock Rack's OCP showcase makes it clear that Direct-to-Chip Liquid Cooling (DLC) is now the baseline for enterprise AI infrastructure. The 2U16X-GNR2/DLC Rubin is a fully liquid-cooled platform, integrating an IT Gear Connector for a 54V DC busbar to align with modern, high-efficiency power architectures. The system extends liquid cooling beyond the GPUs and CPUs, featuring hot-swap liquid-cooled E1.S NVMe drive bays and a liquid-cooled NVIDIA BlueField-4 DPU.

Similarly, ASRock Rack introduced the 6UXGM-GNR2 CX8/DLC, based on the NVIDIA MGX architecture. This system supports liquid-cooled NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, targeting heavy enterprise AI, visual computing, and inference deployments that require sustained performance without thermal throttling.

"Our industry is rapidly evolving to support the shift towards agentic AI, driving demand for infrastructure that can not only manage the increasing thermal requirements but also deliver substantial compute density and memory bandwidth," said Weishi Sa, President of ASRock Rack. "By integrating NVIDIA accelerated computing into our modular server designs, we are helping organizations build scalable, high-performance infrastructure engineered to handle the rigorous computational and thermal requirements of agentic AI and the workloads of tomorrow."

The Open Rack V3 Balancing Act: Standardized Frames, Diverse Silicon

While NVIDIA’s silicon dominates the AI narrative, hyperscale cloud providers and enterprise operators are acutely aware of the risks of single-vendor lock-in. They demand the flexibility to mix and match compute architectures while maintaining a standardized physical footprint. This tension is driving the widespread adoption of the Open Compute Project’s Open Rack V3 (ORV3) standard.

ASRock Rack’s presence at OCP demonstrates a sophisticated multi-architecture strategy designed to operate within these standardized ORV3 frames. Beyond the NVIDIA-centric systems, the company is showcasing a wide array of alternative compute nodes. This includes the liquid-cooled 2OU4N-VENICE2/DLC and the air-cooled 2OU2N-SP8, both powered by next-generation AMD EPYC 9006 Series Processors (codenamed Venice).

For operators deeply embedded in the Intel ecosystem, ASRock Rack is presenting solutions like the 2OU3N3G-GNR, built on Intel Granite Rapids architecture. Furthermore, the inclusion of the ARM-powered 2OU2N-ARM server highlights a commitment to power-efficient, cloud-native compute environments.

By offering NVIDIA, AMD, Intel, and ARM solutions within the same ORV3 modular ecosystem, ASRock Rack is providing data center facility managers with the ultimate procurement flexibility. Operators can deploy liquid-cooled HGX Rubin systems for their most demanding agentic AI workloads, while utilizing air-cooled AMD or ARM nodes for general-purpose cloud hosting—all within the same rack infrastructure, sharing the same 54V DC power busbars and blind-mate liquid cooling manifolds.

Scaling the Next Generation of AI Factories

The hardware currently on the floor at the San Jose Convention Center is not conceptual; it represents the immediate future of global compute deployment. With NVIDIA Vera CPUs and Rubin GPUs entering full production in the second half of 2026, major cloud service providers and AI factories are already beginning their rollout phases.

ASRock Rack is currently offering NVIDIA GB300 NVL72 systems to meet immediate rack-scale demands. However, the company is already previewing the next-generation NVIDIA Vera Rubin NVL72 platform, signaling its near-future availability. These rack-scale systems, which combine 72 Rubin GPUs and 36 Vera CPUs, represent the ultimate manifestation of the agentic AI infrastructure shift, delivering up to 30 times higher throughput per megawatt on agentic workloads compared to previous generations.

The transition occurring at the intersection of energy, industry, and digital infrastructure is monumental. As AI models evolve from passive tools into active, autonomous agents, the engines powering our world must be rebuilt from the silicon up. Through its embrace of modular liquid cooling, diverse compute architectures, and next-generation NVIDIA silicon, ASRock Rack is providing the foundational hardware required to rewrite the rules of global technological competition for decades to come.