For a decade, AI scaling meant bigger GPUs and more data. Now, a new bottleneck is emerging: the humble CPU. As artificial intelligence systems evolve from pure language models into autonomous agents that take multiple steps, call tools, and interact with their environment, the central processing unit (CPU) has unexpectedly become part of the critical path. NVIDIA's newly announced Vera CPU represents a fundamental rethinking of data center processor design, shifting focus from maximizing cores per dollar to maximizing AI factory output per dollar.

Why CPUs Suddenly Matter in the Age of Agentic AI?

Each wave of artificial intelligence has introduced new scaling laws. Pretraining scaled intelligence through larger datasets, more parameters, and massively parallel GPU systems. Post-training scaled usefulness through instruction tuning and rebalancing GPUs for generative inference. Test-time scaling improved reasoning by giving models more generated tokens for thinking. Now, agentic AI and reinforcement learning scale actions. Models take more steps, call more tools, run more evaluations, and interact with execution environments to perform tasks.

The shift is subtle but profound. When an AI agent operates, it follows a cycle: a prompt kicks off generation on the GPU. The GPU generates parameters for a tool call to be executed on the CPU. The CPU executes that tool, producing results that feed back to the GPU to update weights during reinforcement learning or to generate the next prompt. As agents become more capable, they take more steps and run more checks, compounding CPU time and making it part of the critical path.

GPUs remain essential for model inference and training, but much of the execution surrounding model operations runs on CPUs. This includes sandboxed code execution, data retrieval, data processing, scheduling, and orchestration. For the past decade, the data center CPU market optimized around cloud economics, focusing on more cores and lower cost per core. But performance per core has not improved at the same rate, a problem compounded by the end of Moore's Law, which limits generation-on-generation performance improvements in CPUs.

How Does the Vera CPU Solve This Problem?

The NVIDIA Vera CPU is designed for the reality of modern agentic workloads, with fast per-core performance, high concurrency, and power-efficient memory bandwidth. The processor combines 88 NVIDIA Olympus cores with up to 1.2 terabytes per second of LPDDR5X memory bandwidth to keep cores fed through tool calls, sandboxed execution of native code and languages like Python or JavaScript, data retrieval, data processing, and orchestration.

The key requirement is fast per-core performance sustained at all times. Unlike cloud virtual machines where CPU sockets may sit idle, in AI factories the CPU sockets stay fully loaded, doing the work of many concurrent agents. Cores that remain fast under high system load reduce task completion time, delivering faster results while freeing up resources to serve the next request. For agents, this means lower latency across multistep requests. For reinforcement learning, this means more completed evaluations and more data from each training window, helping models reach a higher quality bar faster.

The NVIDIA Olympus core delivers up to 50% higher instructions per cycle (IPC) than NVIDIA Grace, incorporating advanced branch prediction and deep out-of-order instruction scheduling to sustain high throughput on branch-heavy, memory-sensitive agentic code. Olympus uses a neural branch predictor to reduce stalls in branch-heavy code, sustaining two taken branches per cycle with zero penalty. It also includes a 10-wide decode unit and a deep out-of-order engine designed to sustain high instructions per cycle.

Memory bandwidth is equally critical. Vera CPUs deliver up to 1.2 terabytes per second of LPDDR5X memory bandwidth, sustaining over 90% of peak memory bandwidth under load. The processor also offers 40% lower peak memory latency compared to x86 CPUs, ensuring Olympus cores are fed on time through retrieval, analytics, sandbox execution, and orchestration. Olympus adds a novel graph prefetcher built for indirect memory access patterns common in graph analytics and agent memory traversal. Combined with high memory per-core bandwidth, Vera CPUs deliver more than 3 times higher performance on graph traversal workloads compared with x86-based architectures.

Steps to Understand the Vera CPU's Design Advantages

• Core Architecture: The NVIDIA Olympus core delivers 50% higher instructions per cycle than previous generations, using neural branch prediction and deep out-of-order scheduling to maintain throughput on complex, branch-heavy code typical of AI agents and scripting engines.
• Memory Subsystem: Vera CPUs provide up to 1.2 terabytes per second of LPDDR5X memory bandwidth with 40% lower latency than x86 alternatives, plus a specialized graph prefetcher for indirect memory access patterns in agent memory traversal.
• System Efficiency: The processor pairs its architecture with high-bandwidth LPDDR5X memory to reduce memory power to under 30 watts, compared with over 100 watts for traditional DDR5 configurations, with a configurable 250 to 450 watt thermal design power range.
• Fabric Design: The NVIDIA Scalable Coherency Fabric connects all cores and a unified cache across a monolithic mesh, delivering predictable latency and 50% faster core-to-core data movement compared with CPUs that fragment compute across dies.

Together, these components enable the Vera CPU to deliver more than 1.8 times higher sandbox performance across agentic workloads under full load compared with competing processors.

What Does This Mean for AI Factories?

The shift from cores per dollar to tokens per dollar represents a fundamental change in how data centers should be designed for AI. AI factories need CPUs with high core counts to run thousands of concurrent agents, reinforcement learning environments, sandboxes, and services. They need high per-core performance because each agentic step is gated by sequential execution. And they need energy-efficient memory bandwidth to keep data moving without turning CPU infrastructure into a bottleneck.

The NVIDIA Scalable Coherency Fabric connects all cores and a unified cache across a monolithic mesh, delivering predictable latency and 50% faster core-to-core data movement compared with CPUs that fragment compute across dies. For reinforcement learning and agentic AI, that predictability helps keep evaluation loops sustained under full load.

Beyond performance, agentic AI places increasing pressure on infrastructure efficiency. As AI factories scale to thousands of CPUs, memory power can become a major contributor to platform power, cooling demand, and operating cost. The Vera CPU pairs its architecture with high-bandwidth LPDDR5X memory to reduce memory power compared with traditional DDR server designs. With a configurable 250 to 450 watt thermal design power range, the Vera CPU reduces combined CPU and memory subsystem power while delivering the bandwidth needed for agentic inference and reinforcement learning environments. For AI factories, this translates into better performance per watt, lower operating costs, and more efficient use of power and cooling infrastructure.

Meanwhile, NVIDIA is also expanding the ecosystem of tools that agents can use. The company released a major collection of open source physical AI skills and tools spanning NVIDIA Omniverse, Cosmos, Alpamayo, and Metropolis for robotics, autonomous vehicles, vision AI, and industrial digital twins. These physical AI skills turn complex physical AI training, evaluation, and deployment workflows into repeatable, optimized, and agent-executable instructions.

Industry leaders including Agile Robots, Cadence, Dassault Systèmes, Delta Electronics, Foxconn, Pegatron, PTC, Siemens, Synopsys, and TSMC are already using NVIDIA physical AI tools to accelerate physical AI development. In electronic manufacturing, Pegatron reduced model training and deployment time by 67% using synthetic data generated from the Defect Image Generation skill. Delta Electronics generated synthetic defect data and used the skill to catch excess soldering on metal busbars, improving detection rate by 17%. Foxconn, working with DeepHow, used the skill to improve manufacturing efficiency by catching errors early, boosting first pass yield by about 3%.

"AI agents are revolutionizing software development, and that shift is now coming to physical AI, extending into the systems that will transform transportation, manufacturing, healthcare and robotics," said Jensen Huang, founder and CEO of NVIDIA.

Jensen Huang, Founder and CEO at NVIDIA

The era of agentic AI requires a shift in CPU design, from maximizing cores per dollar to maximizing AI factory output. The Vera CPU represents the first major processor redesign specifically for this new paradigm, signaling that the bottleneck in AI infrastructure is shifting away from raw GPU compute and toward the orchestration, tool execution, and data movement that happens on CPUs.