AI talking to AI, where reasoning becomes action
At COMPUTEX 2026 in Taipei, ASRock Rack introduced server hardware built around a quiet but consequential shift in artificial intelligence: the move from systems that answer questions to systems that act autonomously. Powered by NVIDIA's purpose-built Vera CPU, the new 2UXGM-VERA2 platform is designed for agentic AI — where machines reason, decide, and execute without waiting for human instruction at each step. This moment reflects a broader reckoning in computing, where the humble CPU, long overshadowed by the GPU in AI conversations, reasserts itself as the critical layer between thought and action.
- The AI industry's center of gravity is shifting from making models smarter to making them autonomous — and that demands entirely new hardware foundations.
- NVIDIA's Vera CPU delivers a claimed 50% speed advantage over conventional processors for agentic and reinforcement learning workloads, signaling that the old infrastructure was never built for this moment.
- Liquid cooling has moved from optional to essential, with ASRock Rack displaying multiple thermal architectures capable of sustaining the heat output of dense, continuous AI execution at rack scale.
- ASRock Rack is placing a full-stack bet — from industrial edge robotics to cloud inference clusters — that the agentic era requires not just new chips but entirely reimagined server ecosystems.
- The company's COMPUTEX showcase positions it as an early infrastructure partner for AI factories where autonomous systems must operate at speed, scale, and without interruption.
At COMPUTEX 2026 in Taipei, ASRock Rack unveiled the 2UXGM-VERA2, a server system built around NVIDIA's Vera CPU — a processor designed not for generating text or images, but for something more consequential: enabling AI systems to reason and act without human intervention at each step.
The distinction reflects a genuine inflection point in the industry. For years, progress meant larger models and better predictions. Now the focus is shifting toward agentic AI — systems where AI coordinates with AI, where reasoning becomes action, where a model doesn't just answer but executes sequences of tasks. That shift places new demands on the CPU. NVIDIA's Vera was purpose-built for this, combining custom Olympus cores, high-bandwidth LPDDR5X memory, and a Scalable Coherency Fabric to deliver a claimed 50% speed advantage over conventional infrastructure on agentic and reinforcement learning workloads.
ASRock Rack president Weishi Sa framed it directly: the industry is moving from response generation to autonomous action execution, and the Vera CPU provides the layer that makes that transition real. The company is building an entire server lineup around the processor, from standalone CPU systems to larger configurations pairing Vera with NVIDIA's HGX Vera Rubin NVL8 GPU clusters.
The COMPUTEX booth illustrated what running these systems at scale actually requires. Liquid cooling featured prominently — double-rack-width distribution units for AI factory environments, fully liquid-cooled 2U and 5U systems managing both CPU and GPU thermals simultaneously, and in-row coolant distribution for the densest deployments. The portfolio extended beyond the data center as well, with systems targeting enterprise visual computing, inference cloud environments, and industrial edge applications including autonomous robotics and medical devices.
Founded in 2013, ASRock Rack has spent over a decade building server hardware for cloud computing. The Vera announcement represents the company's conviction that the next wave of AI will demand not just new GPUs, but new CPUs, new cooling strategies, and new architectures spanning the full distance from edge to cloud.
At COMPUTEX 2026 in Taipei, ASRock Rack unveiled a new class of server hardware built around a fundamental shift in how artificial intelligence systems work. The company's centerpiece was the 2UXGM-VERA2, a system powered by NVIDIA's Vera CPU—a processor designed not for the familiar task of generating text or images, but for something more autonomous: enabling AI systems to reason through problems and execute decisions without human intervention at each step.
The distinction matters because the AI industry itself is changing shape. For years, the focus was on making models bigger and better at prediction. Now the focus is moving toward what researchers call agentic AI—systems where AI talks to AI, where reasoning becomes action, where a model doesn't just answer a question but carries out a sequence of tasks. That shift puts enormous pressure on the CPU, the traditional workhorse of general computing. NVIDIA's Vera CPU was purpose-built for this moment, combining custom-designed Olympus cores with high-bandwidth LPDDR5X memory and NVIDIA's Scalable Coherency Fabric. The result, according to the company's specifications, is a 50 percent speed advantage over conventional CPU infrastructure when running complex agentic and reinforcement learning workloads.
Weishi Sa, president of ASRock Rack, framed the moment plainly: the industry is moving from systems that generate responses to systems that execute autonomous actions. The Vera CPU, he said, provides the execution layer that turns model reasoning into action at the speed required by modern AI factories. ASRock Rack is planning an entire lineup of servers around this processor, ranging from standalone CPU systems to larger scale-up configurations that pair Vera with NVIDIA's HGX Vera Rubin NVL8 GPU clusters.
The company's booth at COMPUTEX showcased the infrastructure required to run these systems at scale. Liquid cooling has become essential for high-density AI deployments, and ASRock Rack displayed several approaches: the NVIDIA Vera Rubin NVL72 cooled with a double-rack-width liquid-to-air distribution unit for AI factory environments, and fully liquid-cooled systems like the 2U16X-GNR2/DLC and 5U16X-GNR2/DLC that combine liquid cooling for both CPU and GPU while maintaining airflow for other components. The company also demonstrated rack-scale liquid-cooling solutions with in-row coolant distribution units designed for the densest deployments.
Beyond the data center, ASRock Rack showed a broader portfolio. The 6UXGM-GNR2/DLC supports up to eight liquid-cooled NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs for enterprise AI and visual computing. The 4UXGM-GNR2 CX8 targets inference cloud environments. And the 2UXGI-Thor, built on NVIDIA's IGX Thor platform, bridges to industrial edge applications—real-time sensor processing and functional safety for autonomous robotics and medical devices.
ASRock Rack, founded in 2013, has spent over a decade building server hardware for cloud computing. The company operates under three principles: creativity, consideration, and cost-effectiveness. The Vera announcement represents a bet that the next wave of AI computing—the move from response generation to autonomous action—will require not just new GPUs but new CPUs, new cooling strategies, and new server architectures from edge to cloud. The company invited visitors to its booth at Taipei Nangang Exhibition Center, Hall 2, to see live demonstrations of agentic AI systems running on this new infrastructure.
Notable Quotes
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
The Hearth Conversation Another angle on the story
Why does the CPU suddenly matter so much for AI? Weren't GPUs doing all the heavy lifting?
GPUs excel at parallel matrix math—the core of training and inference. But agentic AI means the system is making decisions, branching logic, coordinating multiple steps. That's CPU work. When AI talks to AI, you need fast sequential execution, not just raw throughput.
So this Vera CPU is just a faster processor?
It's purpose-built for this specific problem. Custom cores, high-bandwidth memory, a coherency fabric designed for AI workloads. It's not a general-purpose chip trying to do everything. It's optimized for one thing: making autonomous AI systems run 50 percent faster.
Why is liquid cooling suddenly everywhere in these systems?
Density. You're packing more compute into the same space, and air cooling hits a wall. Liquid lets you run hotter, tighter, more efficiently. For AI factories running 24/7, the cooling strategy is as important as the chips themselves.
Is this just for massive data centers, or does it trickle down?
ASRock is showing everything from edge devices like the IGX Thor for robotics and medical imaging, all the way up to rack-scale deployments. The architecture is the same principle at every level—purpose-built for AI execution, not retrofitted from older designs.
What does this mean for companies building AI systems right now?
It means the infrastructure is finally catching up to what the software is trying to do. For years, people were running agentic AI on hardware designed for something else. Now there's hardware that actually fits the workload.