For two decades, Nvidia established itself as the premier provider of graphics processing units (GPUs). Now, the technology giant is aiming to challenge the established dominance of Intel and AMD in the server CPU market. On May 31, 2026, during the GTC Taipei keynote, Nvidia formally introduced Vera, marking its first fully custom-designed data center processor. Nvidia branded the chip “the CPU for Agents,” designed to integrate with its new Rubin GPU within the Vera Rubin platform.
Technical Specifications and Architecture
The Vera chip incorporates 88 proprietary “Olympus” cores and is architecturally distinct from products licensed from Arm’s Neoverse line, unlike Nvidia’s previous Grace CPU. According to Nvidia’s product documentation, Vera is built to ensure instruction-set compatibility with Arm v9.2 while maintaining an entirely custom core design. The chip exposes 176 threads using a feature called Spatial Multithreading, or SMT-X.
In terms of performance metrics, the CPU supports up to 1.2 TB/s of memory bandwidth and up to 1.5 TB of LPDDR5X capacity per socket. It operates with a configurable Thermal Design Power (TDP) ranging from 250W to 450W, depending on how it is deployed. The Vera CPU is designed to work synergistically with the new Rubin GPU over sixth-generation NVLink, allowing the CPU and GPU to access a shared, coherent memory address space via NVLink-C2C. This capability is intended to eliminate the data transfer bottlenecks that Nvidia claims Vera is designed to resolve.
While Nvidia offers standalone Vera-only racks that can scale up to 256 liquid-cooled chips, the flagship deployment model is the Vera Rubin NVL72 rack, code-named Oberon. This system integrates 72 Rubin GPUs and 36 Vera CPUs into a single liquid-cooled cabinet. Each Rubin GPU contains 288 GB of HBM4 memory, while each Vera CPU holds 1.5 TB of LPDDR5X. When combined across a full rack, the system offers approximately 20.7 TB of HBM4 and roughly 54 TB of LPDDR5X, connected by the sixth-generation NVLink Switch. Nvidia asserts that this rack can deliver up to 3,600 petaflops of NVFP4 inference performance, though this figure originated from Nvidia’s internal materials.
Core Design and Performance Claims
The foundation of the Vera chip is the Olympus core, which represents Nvidia’s first ground-up CPU core design. Unlike the Grace CPU, which utilized off-the-shelf Arm Neoverse cores, Olympus retains Arm v9.2 compatibility for existing software while allowing Nvidia to design the fetch, decode, and branch-prediction logic internally. This custom tuning is optimized for the branch-heavy, low-arithmetic code generated by AI agent orchestration, rather than the dense mathematical computations typically handled by GPUs.
Nvidia’s approach to multithreading, SMT-X, differs from Intel’s Hyper-Threading and AMD’s SMT. Instead of time-slicing resources between two threads, SMT-X physically partitions a core’s resources, allowing two threads to run with genuine isolation. Nvidia argues this distinction is vital for large-scale AI environments running thousands of concurrent agent sandboxes on a single socket.
Independent performance testing provided a measured comparison. Phoronix ran a benchmark suite and determined that the 88-core Vera performed approximately 55% faster than Intel’s best single-socket Xeon 6980P and was ahead by 10% to 11% compared to AMD’s EPYC 9575F, based on a geometric mean across the test suite. However, the test was weighted toward specific workloads, including Python execution, code compilation, Java, and memory-stream tests, leaving general-purpose enterprise workloads untested.
Market Positioning and Financial Outlook
The Vera CPU enters a market segment traditionally divided between Intel and AMD. Data from Mercury Research, as cited by CNBC, showed that for the fourth quarter of 2025, server CPU shipment shares were: Intel at 60%, AMD at 24.3%, and Nvidia at 6.2%. The remaining share belongs to in-house Arm chips developed by companies like Amazon, Microsoft, and Google.
Nvidia’s Chief Financial Officer, Colette Kress, stated the financial potential of the new product line:
Vera CPU opens a brand-new $200 billion TAM for Nvidia, a market we have never addressed before.
Kress also noted the immediate revenue potential, stating:
We have visibility to nearly $20 billion in total CPU revenue this year, setting us up to become the world’s leading CPU supplier.
Morgan Stanley Research provided a bill-of-materials estimate, reporting that a fully configured NVL72 Vera Rubin rack cost approximately $7.8 million. This estimate highlights that memory is a major cost driver, accounting for 26% of the total rack cost, making the memory supply a critical factor in the system’s economics.
The Shift to Agentic Computing
The fundamental shift in the market is the rise of “agentic AI.” In previous years, the CPU was often treated as a secondary component, merely tasked with keeping the powerful GPUs fed. However, the complex processes involved in modern AI agents—such as planning, calling external tools, running code sandboxes, and integrating results—are predominantly handled by the CPU. This shift means that the CPU is becoming the primary bottleneck in scaling advanced AI workflows.
The Vera CPU’s design, particularly the Olympus core’s branch prediction and the SMT-X threading model, is specifically aimed at addressing this CPU-centric computational gap, rather than attempting to compete with a GPU’s raw mathematical throughput. The increased focus on system integration means that buyers must now weigh the flexibility of combining different CPU vendors (Intel or AMD) with Nvidia GPUs against the tightly coupled, integrated power of the Vera CPU-GPU stack.
Commercial availability for the Vera CPU is slated for fall 2026, distributed through system builders and cloud computing partners. While Nvidia has not disclosed official pricing for the chip itself, the initial interest has included major cloud providers such as Anthropic, OpenAI, and Oracle Cloud Infrastructure, with OCI specifically planning to deploy hundreds of thousands of Vera-based units starting this year. Furthermore, the New York Stock Exchange is reportedly evaluating the chip for high-throughput market data processing, indicating the core’s appeal extends beyond AI workloads.
