Nvidia has spent decades building its dominance on Graphics Processing Units (GPUs), powering everything from gaming to artificial intelligence. But, with the launch of the Vera CPU, the company is now making a more direct push into territory long dominated by Intel and AMD.
Unveiled at GTC Taipei, Vera is not just another server processor. Nvidia is positioning it as the first CPU designed specifically for AI agents, a category of systems that is moving beyond answering queries to executing tasks, running code, and interacting with systems autonomously.
“AI agents will be the largest users of computing,” said Jensen Huang, founder and chief executive of Nvidia. “Vera is the first CPU designed for that future.”
Why Nvidia is building its own CPU
While Nvidia is best known for GPUs, this is not its first attempt at Central Processing Units (CPUs). The company has been working toward this moment for over a decade, from early Tegra chips to more recent data centre processors like Grace.
Grace, however, relied on off-the-shelf Arm core designs. Vera changes that.
With Vera, Nvidia has designed its own custom CPU core, called Olympus, marking its return to fully in-house CPU architecture for the first time in years.
This shift is significant. By moving away from licensed designs, Nvidia gains tighter control over performance, efficiency, and how the CPU interacts with its GPUs and software stack.
More importantly, it allows the company to tailor the processor specifically for AI workloads, rather than general-purpose computing.
What makes Vera different from traditional CPUs
At a high level, Vera is still a CPU. It handles general-purpose computing tasks, runs operating systems, and supports workloads that are not suited for GPUs.
But the way it is designed reflects a different priority.
Unlike traditional CPUs built for a wide range of applications, Vera is optimised for what Nvidia calls “agentic workloads”. These include tasks such as code execution, orchestration logic, data processing, and reinforcement learning, all of which sit alongside GPU-driven AI models in modern systems.
According to Nvidia, the chip can deliver up to 1.8 times faster task completion compared to x86 processors across these workloads.
This performance gain is not just about raw speed, but about how AI systems actually operate today.
Modern AI systems do not rely solely on GPUs. While GPUs handle model training and inference, CPUs manage the surrounding tasks, including running code, handling inputs, coordinating processes, and evaluating outputs. As AI agents become more complex, this CPU-side work is growing rapidly.
Vera is designed to handle that layer.
Built for the “AI factory”
Nvidia increasingly describes data centres as “AI factories”, where models generate tokens, process data, and execute tasks at scale. In this model, the role of the CPU changes. Instead of being a general-purpose processor, it becomes a coordinator, handling orchestration, memory movement, and execution environments that support AI systems.
Vera reflects this shift. The chip features 88 custom Olympus cores, supports spatial multithreading, and delivers up to 1.2TB per second of memory bandwidth, significantly higher than traditional server CPUs.
It is also tightly integrated with Nvidia’s GPU ecosystem through high-bandwidth interconnects, allowing faster communication between CPU and GPU.
This is critical because many AI workloads are limited not by compute, but by how quickly data can move between components.
A different approach to CPU design
Architecturally, Vera also takes a different route compared to competing server chips. While many modern CPUs split compute cores across multiple chiplets, Nvidia keeps all 88 cores within a single compute die, surrounded by separate chiplets for memory and I/O.
This allows the processor to behave more like a single large CPU, avoiding the complexity of partitioning workloads across multiple compute units.
This design prioritises low-latency communication between cores, which is particularly important for AI workloads that rely on tightly coordinated execution, but it comes at the cost of reduced flexibility compared to chiplet-based designs.
In addition, the company has designed the CPU to anticipate application behaviour and process instructions more efficiently, reducing the time agents spend waiting on CPU-bound operations.
More than just a GPU companion
One of the more notable aspects of Vera is that Nvidia is not positioning it solely as a companion to its GPUs. While it will power systems such as the upcoming Vera Rubin platforms, the company is also offering it as a standalone CPU for servers, marking its most direct attempt yet to compete in the broader data centre processor market.
This is a shift from earlier strategies, where Nvidia CPUs primarily existed to support GPU workloads.
Now, the company is attempting to become a full-stack computing provider, offering CPUs, GPUs, networking, and software as a unified platform.
Where RTX Spark fits in
The launch of Vera follows Nvidia’s introduction of RTX Spark, a superchip designed for Windows PCs that combines a CPU and GPU into a single platform for local AI workloads.
The RTX Spark brings together a Grace CPU and Blackwell GPU with unified memory, enabling AI models and agents to run directly on personal devices.
Together, RTX Spark and Vera point to a broader strategy.
On one end, Nvidia is building hardware for personal AI computing. On the other, it is building infrastructure for large-scale AI systems. In both cases, the goal is the same: to support AI agents as they move from cloud-based tools to systems that operate continuously across devices.
What this means for consumers
Vera itself is not a consumer product. It will power data centres, cloud infrastructure, and enterprise AI systems.
But its impact will eventually reach users. As AI agents become more capable, the systems powering them need to handle more than just model inference. They need to execute tasks, manage workflows, and operate across environments in real time.
That requires stronger CPU performance alongside GPUs.
In practical terms, this could lead to:
-
Faster execution of AI-driven tasks across services
-
More complex workflows being handled automatically by AI systems
-
Reduced latency in cloud-based AI operations
While these improvements happen in data centres, they directly affect how AI systems behave from a user perspective, making them more responsive, capable, and reliable in real-world use.
Big-tech names adopting Vera
According to Nvidia, customers exploring the Vera CPU include finance leader NYSE, global AI labs Anthropic, OpenAI and SpaceXAI, and hyperscalers ByteDance, CoreWeave, Lambda, Nebius, Nscale and Oracle Cloud Infrastructure (OCI). The company said that Vera is also being integrated into AI infrastructure from world-leading system manufacturers such as Dell Technologies, HPE, Lenovo and Supermicro, along with Taiwan system builders.
