Nvidia unveils Vera CPU with 2x performance gain over x86 rivals

Nvidia's custom Arm-based Vera CPU marks the company's most direct challenge yet to AMD and Intel's dominance in data center processors.

Nvidia Corp. is pushing into the $30 billion data center CPU market with its custom Vera processor, claiming 2x the performance of x86 chips from Advanced Micro Devices Inc. and Intel Corp. on agentic AI workloads.

"CPUs are becoming the bottleneck in agentic AI systems," Ian Buck, vice president of accelerated computing at Nvidia, said. "We designed Vera to eliminate that bottleneck."

Vera uses Nvidia's own Olympus microarchitecture on a monolithic die — a departure from the chiplet designs used by AMD's EPYC and Intel's Xeon lines. The chip delivers up to 1.2 terabytes per second of memory bandwidth through LPDDR5X, roughly three times more bandwidth per core than competing server platforms, according to Nvidia. The company also claims 40% lower memory latency under load and five times greater bandwidth-per-watt efficiency.

The CPU expansion opens a new front in Nvidia's rivalry with AMD and Intel, which together control more than 90% of the data center CPU market. Nvidia shares trade at about 35x forward earnings, reflecting expectations that its AI platform — spanning GPUs, CPUs, networking and software — can sustain growth beyond the current accelerator cycle.

Built for the Agent Loop

Nvidia designed Vera specifically for what it calls the "agent loop" — the rapid back-and-forth between GPU inference and CPU-driven tasks such as code execution, database lookups and tool orchestration. Unlike traditional cloud CPUs optimized for throughput and core density, Vera prioritizes single-threaded execution and predictable memory latency, the company said.

The Olympus core features a wide 10-way decode front end designed to sustain high single-threaded performance under heavy workloads. Nvidia's second-generation Scalable Coherent Fabric links the CPU's IP blocks on a single die, delivering roughly three times greater core-to-core bandwidth than multi-chiplet designs. The monolithic approach eliminates the latency penalties associated with stitching together multiple chiplets, a trade-off that AMD and Intel have accepted to scale core counts.

Early Benchmarks and Customer Traction

Nvidia shared benchmark data showing Vera delivering up to 2x faster agentic sandbox startup and job completion compared with current x86 platforms. In streaming data processing workloads developed with Redpanda and Hewlett Packard Enterprise Co., Vera showed as much as 6x higher performance. Los Alamos National Laboratory reported up to 7x faster scientific computing performance versus an Intel Sapphire Rapids-based supercomputer. Nvidia did not disclose the test conditions for all comparisons, and independent benchmarks will be needed to verify the claims.

The Vera Rubin NVL72 system — pairing 36 Vera CPUs with 72 Rubin GPUs — is scheduled to ship in the second half of 2026, with Microsoft Corp. and Oracle Corp. among early customers. OpenAI already has one Vera Rubin rack in use, Nvidia executives said during a tour of the company's data center lab in Silicon Valley. Vera chips were delivered to OpenAI, Anthropic and SpaceX in June, according to Nvidia representatives.

For investors, the question is whether Vera can gain meaningful share in a market dominated by AMD and Intel's entrenched x86 server platforms. Nvidia's advantage lies in vertical integration: Vera is co-optimized with its GPUs, networking and CUDA software, potentially offering lower total cost of ownership for AI workloads. If agentic AI scales as expected, Vera could capture a slice of the CPU market that IDC projects will exceed $35 billion annually by 2028.

This article is for informational purposes only and does not constitute investment advice.