AI talking to AI has placed CPU execution on the critical path
At COMPUTEX 2026 in Taipei, ASRock Rack has unveiled a new generation of server infrastructure built around NVIDIA's Vera CPU, a processor designed not for the AI of yesterday—where humans prompt and machines respond—but for a coming era in which AI systems direct one another and act autonomously at scale. The announcement marks a meaningful inflection point: the industry is shifting its foundational architecture from model scaling toward agentic execution, where speed, thermal density, and coherent memory fabric become the new measures of readiness. ASRock Rack's broad lineup, stretching from edge robotics to liquid-cooled AI factories, suggests that this transition is no longer theoretical—it is being built into hardware today.
- The AI industry is undergoing a structural shift away from human-to-AI interaction toward AI-to-AI autonomy, and the hardware infrastructure must now catch up to that ambition.
- NVIDIA's Vera CPU—combining custom Olympus cores, LPDDR5X memory, and Scalable Coherency Fabric—completes agentic and reinforcement learning workloads 50% faster than conventional CPU systems, raising the stakes for competitors.
- ASRock Rack is not betting on a single product but on an entire ecosystem, from liquid-cooled AI factory systems handling extreme thermal density to edge platforms enabling real-time physical AI in robotics and medical environments.
- Liquid cooling runs throughout the lineup as a quiet but urgent signal: the compute density required to run thousands of simultaneous agentic AI instances generates heat that air cooling alone can no longer manage.
- Live agentic AI demonstrations on the show floor transform abstract specifications into visible, functioning infrastructure—an attempt to close the gap between research promise and production confidence.
At COMPUTEX 2026 in Taipei, ASRock Rack is showcasing the 2UXGM-VERA2, a server system powered by NVIDIA's Vera CPU and built for what the industry is calling the age of agentic AI. This is not a modest upgrade. The shift being described is from AI systems that respond to human queries toward AI systems that communicate with one another and execute autonomous decisions at enterprise scale.
The Vera CPU achieves this through a combination of custom NVIDIA Olympus cores, high-bandwidth LPDDR5X memory, and NVIDIA's Scalable Coherency Fabric—delivering agentic and reinforcement learning workloads 50 percent faster than traditional CPU infrastructure. ASRock Rack president Weishi Sa described the moment as a frontier where systems move beyond generating content to executing autonomous action at the speed modern AI factories demand.
The company's lineup spans the full range of deployment scenarios. At the high end sits the NVIDIA Vera Rubin NVL72, a massive liquid-cooled system designed for next-generation AI factory environments. For enterprise GPU workloads, the 6UXGM-GNR2/DLC supports up to eight liquid-cooled NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. The 4UXGM-GNR2 CX8 targets inference cloud environments where latency and throughput are paramount.
At the industrial edge, the 2UXGI-Thor—built on NVIDIA's IGX Thor platform—is designed for Physical AI: real-time sensor processing and functional safety for autonomous robotics and medical applications where hardware must perceive, reason, and act in the physical world without human intervention. This is not a data center abstraction. It is hardware living at the boundary between digital reasoning and physical consequence.
Liquid cooling appears throughout the lineup, reflecting how seriously the industry is treating thermal density as compute stacks grow denser. ASRock Rack, founded in 2013 as a server-focused spinoff, is using COMPUTEX 2026 to position itself at the center of this infrastructure wave—with live agentic AI running on the hardware at the booth, inviting visitors to see autonomous execution not as a specification on paper, but as something already happening.
At COMPUTEX 2026 in Taipei, ASRock Rack is rolling out a new class of server hardware built around a fundamental shift in how artificial intelligence actually works. The company's booth at the Taipei Nangang Exhibition Center is showcasing the 2UXGM-VERA2, a system powered by NVIDIA's Vera CPU—a processor engineered specifically for what the industry is calling the age of agentic AI. This is not incremental. The move represents a pivot away from the familiar model of humans querying AI systems toward a future where AI systems communicate with and direct other AI systems, making autonomous decisions at scale.
The Vera CPU itself combines custom-designed NVIDIA Olympus cores with high-bandwidth LPDDR5X memory and NVIDIA's Scalable Coherency Fabric. In benchmarks, the architecture completes complex agentic and reinforcement learning workloads 50 percent faster than traditional CPU infrastructure. Weishi Sa, president of ASRock Rack, framed the moment plainly: the industry is entering a frontier where systems move beyond generating text or images to executing autonomous actions. The Vera processor, he said, provides the execution layer that turns model reasoning into autonomous action at the speed required by modern AI factories.
ASRock Rack is not launching a single product. The company is unveiling a comprehensive lineup spanning the full spectrum of deployment scenarios. At the high end, the exhibit features the NVIDIA Vera Rubin NVL72, a massive system cooled by a double-rack-width liquid-to-air coolant distribution unit designed for next-generation AI factory environments where density and thermal management are existential concerns. For enterprises needing GPU acceleration, ASRock Rack is showing the 6UXGM-GNR2/DLC, which supports up to eight liquid-cooled NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs for enterprise AI and visual computing workloads. The 4UXGM-GNR2 CX8 targets inference cloud environments where latency and throughput matter most.
The company is also bridging toward the industrial edge. The 2UXGI-Thor, built on NVIDIA's IGX Thor platform, is designed for what the industry calls Physical AI—real-time sensor processing and functional safety for autonomous robotics and medical applications where a system must perceive, reason, and act in the physical world without human intervention. This is not a server sitting in a data center rack. This is hardware that lives at the boundary between digital reasoning and physical consequence.
Liquid cooling appears throughout the lineup, a signal of how seriously the industry is taking thermal density. The 2U16X-GNR2/DLC and 5U16X-GNR2/DLC both combine liquid cooling for CPU and GPU with sufficient airflow for flexible integration of other peripherals. ASRock Rack is also demonstrating rack-scale liquid-cooling solutions with in-row coolant distribution units for high-density deployments—the kind of infrastructure required when you're stacking enough compute to run thousands of agentic AI instances simultaneously.
ASRock Rack itself was established in 2013 as a spinoff focused on cloud computing server hardware. The company has inherited ASRock's design philosophy—creativity, consideration, cost-effectiveness—and applied it to the server market. What's happening at COMPUTEX 2026 is the company's attempt to position itself at the center of the next wave of AI infrastructure, the moment when agentic AI moves from research demonstrations to production deployments at scale. The booth demonstrations include live agentic AI running on the hardware, a chance for visitors to see the systems not as abstract specifications but as functioning infrastructure executing autonomous tasks in real time.
Citas Notables
The AI industry is entering a new frontier where systems move beyond generating responses to executing autonomous actions. NVIDIA Vera is the CPU for the age of AI, providing the high-speed execution layer that turns model reasoning into autonomous action at the speed of the AI factory.— Weishi Sa, President of ASRock Rack
La Conversación del Hearth Otra perspectiva de la historia
Why does the shift from human-to-AI to AI-to-AI communication require a completely different kind of CPU?
Because the bottleneck changes. When a human asks a question and waits for an answer, latency is measured in seconds. When an AI system is orchestrating dozens of other AI systems making decisions in parallel, you need CPU execution that can handle rapid context switching, coherent memory access, and high-speed reasoning loops. The Vera is built for that workload, not for the old pattern.
The 50 percent speed improvement—is that real, or is that marketing math?
It's measured against traditional CPU infrastructure running the same agentic workloads. Whether it's real depends on whether you believe the benchmarks, which I do. But the more important point is that it's purpose-built. A general-purpose CPU wasn't designed for this. Vera was.
Why is liquid cooling suddenly everywhere in this lineup?
Density. If you're running agentic AI at scale, you need to pack a lot of compute into a small space. Air cooling hits a wall. Liquid cooling lets you push further without the system throttling itself or burning out. It's not optional anymore.
The 2UXGI-Thor for robotics and medical devices—that feels like a different market entirely.
It is and it isn't. The underlying principle is the same: a system that can perceive, reason, and act autonomously in real time. In a data center, that's orchestrating other AI systems. In a robot or a medical device, it's controlling physical actuators. Same CPU architecture, different application.
What happens if this doesn't take off? What if agentic AI stays niche?
Then ASRock Rack has bet heavily on a future that doesn't materialize. But the company is betting that agentic AI is not a research curiosity—it's the direction the entire industry is moving. If they're right, they're positioned early. If they're wrong, they've built beautiful hardware for a market that doesn't exist yet.