The Agentic AI Pivot: ASUS Reveals Next-Gen Datacenter Strategy at OCP

SAN JOSE, Calif. – October 12, 2026 – For the last three years, the artificial intelligence narrative has been dominated by a single metric: how fast a massive cluster of GPUs can train a large language model. But as AI transitions from passive chatbots to autonomous agents capable of executing complex workflows, the underlying infrastructure is facing a profound architectural reckoning.

At the Open Compute Project (OCP) Global Summit 2026, ASUS unveiled a comprehensive datacenter portfolio that underscores this industry-wide pivot. Headlining the announcement is the XA P2N-E2, a 2U NVIDIA MGX architecture-based server powered by NVIDIA Vera CPUs and designed specifically for agentic AI and reinforcement learning workloads. This release, alongside a broader multi-silicon strategy featuring advanced Intel and AMD platforms, signals a critical inflection point in how enterprises will build the next generation of AI factories.

The Agentic Pivot: Why AI's Next Wave Needs CPUs

The emergence of agentic AI—systems that can autonomously plan, write code, interact with external APIs, and evaluate their own outputs—has exposed a new bottleneck in the modern datacenter. These execution-layer tasks are fundamentally different from the parallel matrix multiplication that GPUs handle so well. They require sequential processing, complex conditional logic, and robust sandboxing for security.

The XA P2N-E2 addresses this shifting dynamic by combining up to two dual-slot NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs with the high single-thread performance of NVIDIA Vera CPUs.

"We are seeing a shift where the model inference itself is only a fraction of the total compute time," noted one enterprise AI infrastructure architect attending the summit. "When an agent writes a Python script to analyze a financial dataset or orchestrate a supply chain, that code needs a secure, high-speed CPU environment to execute. If your infrastructure is just a sea of GPUs, your agents will stall waiting on basic compute tasks."

By optimizing for high-bandwidth, energy-efficient memory, the new hardware ensures that AI agents remain responsive during demanding evaluation cycles. This dedicated focus on the execution layer effectively bridges the gap between thought and action in autonomous systems, a necessary evolution as businesses look to deploy AI that actively improves operational lives rather than just answering queries.

The Tri-Silicon Datacenter: Balancing Ecosystems

While the NVIDIA-powered systems drew significant attention, the broader story from the summit floor is the strategic necessity of a multi-vendor datacenter. The Taiwanese hardware manufacturer dedicated substantial floor space to new platforms powered by Intel and AMD, reflecting a reality where enterprise CIOs are fiercely guarding against vendor lock-in while seeking the best silicon for specific workloads.

The showcase introduced the RS700-E12-RS4U and RS720-E12-RS12U platforms, powered by Intel Xeon 6 processors. Built on the Open Compute Project's DC-MHS (Data Center-Modular Hardware System) standard, these platforms emphasize tool-less serviceability and modular deployment. This modularity is crucial for datacenters that must rapidly adapt their physical footprints to accommodate virtualization and data-intensive orchestration tasks.

On the AMD front, the company expanded its architectural choices with the RS700A-E14B-R16U, driven by 6th Gen AMD EPYC 9006-series processors, codenamed 'Venice'. Featuring PCIe 6.0 connectivity, this platform is aimed squarely at compute-heavy infrastructure requirements. For accelerated AI outside the NVIDIA ecosystem, the ESC8000A-E13P supports AMD Instinct MI350P PCIe accelerators, while the massive 10U XA AM3A-E13 integrates eight AMD Instinct MI355X accelerators.

This tri-silicon approach allows organizations to align their compute architecture with specific demands. Whether deploying Intel for high-efficiency virtualization, AMD for enterprise resilience and alternative acceleration, or NVIDIA for cutting-edge agentic execution, flexibility has become the ultimate datacenter currency.

Liquid or Air: Navigating Datacenter Realities

The physical reality of housing these next-generation systems presents perhaps the most daunting challenge for the 2026 landscape. The thermal design power of modern AI silicon is pushing legacy datacenters to their breaking points, creating a stark divide between what is technologically possible and what is structurally feasible.

The company's showcase highlighted this tension perfectly. At the extreme high end, the XA VR721-E3 AI POD integrates 36 NVIDIA Vera CPUs and 72 NVIDIA Rubin GPUs through the NVIDIA Vera Rubin NVL72 platform. This rack-scale system relies on a fully liquid-cooled architecture to manage the immense heat generated by large-scale AI training and inference.

However, retrofitting legacy enterprise facilities for direct-to-chip or immersion cooling is incredibly capital-intensive and disruptive.

"You cannot simply drop a 100-kilowatt liquid-cooled rack into a facility built in 2015 without tearing the walls down," explained a lead datacenter facility operator at the summit. "We need high-density solutions that can still utilize our existing chilled air infrastructure while we phase in liquid loops over the next five years."

To address this transition, hybrid and advanced air-cooled solutions were prominently featured. The XR R1I12B-M1BHFS combines direct-to-chip liquid cooling for its NVIDIA HGX Rubin NVL8 baseboard with traditional air cooling for its Intel Xeon 6 processors. Meanwhile, the AMD-powered XA AM3A-E13 utilizes advanced Thermal Radar 3.0 technology to optimize air-cooling efficiency across both CPU and GPU components, allowing for high-density AI deployments without requiring immediate, dedicated chilled-water plumbing.

Building the Next-Generation AI Factory

The evolution of the datacenter is no longer just about raw compute; it is about orchestrating an entire ecosystem. High-speed networking, scalable context memory, and responsive data infrastructure are now co-equal pillars of the modern AI factory.

To this end, the summit booth featured integrations of NVIDIA Spectrum-X Spectrum-6 Ethernet switching for high-speed networking and the NVIDIA CMX Context Memory Storage Platform. The latter is particularly crucial for agentic systems, which require massive, rapidly accessible context windows to remember past interactions and tool outputs during multi-step, complex tasks.

Furthermore, the inclusion of NVIDIA Groq 3 LPX interactive AI inference accelerators demonstrates a commitment to ultra-fast token generation, a necessity for keeping autonomous agents responsive in real-time applications where latency translates directly to lost value.

Extending beyond the hardware stack, the integration of the NVIDIA Omniverse DSX Blueprint for AI Factory Digital Twins allows facility managers to physically simulate and optimize their datacenter operations before a single server is deployed. By mapping out thermal flows and compute workloads in a high-fidelity simulation, operators can mitigate the risks associated with deploying high-density infrastructure.

This full-stack, modular architecture represents a maturation of the AI industry. We are moving past the brute-force era of simply stacking accelerators, entering a nuanced phase where compute density, thermal management, and specialized silicon must be perfectly balanced to unlock the true potential of autonomous AI.