Nvidia Launches Vera CPU: First Direct Challenge to Intel and AMD in Datacenter Market
Key Takeaways
- ▸Vera marks Nvidia's first direct CPU challenge to Intel and AMD, backed by commitments from eight major cloud providers and hyperscalers
- ▸Custom Armv9.2 Olympus cores built on TSMC 3nm with monolithic compute die design deliver higher bandwidth and lower latency than competing architectures
- ▸Purpose-engineered for AI infrastructure: GPU management and AI agent hosting, addressing two distinct workloads beyond traditional compute
Summary
Nvidia has officially entered the CPU market as a direct competitor to Intel and AMD with the launch of Vera, a custom-built processor featuring 88 cores, up to 1.5 TB of LPDDR5X memory, and 1.8 TB/s of NVLink connectivity. The chip represents Nvidia's most ambitious push beyond GPUs, with major cloud providers including Meta, Alibaba, ByteDance, Oracle, CoreWeave, Lambda, Nebius, and NScale already committing to deployments. Vera's architecture diverges significantly from traditional x86 designs, featuring a monolithic compute die built on TSMC's 3nm process with custom Armv9.2 cores—dubbed Olympus—and disaggregated memory and I/O chiplets. The CPU is purpose-built for two key workloads: serving as an AI head node to manage GPU orchestration in Vera Rubin systems, and hosting AI agents that run inference workloads unsuitable for GPU acceleration. The dual-socket Vera CPU Superchip configuration delivers 176 total cores and 3 TB of memory with 1.8 TB/s of bidirectional bandwidth between chips, enabling hyperscalers to build massive AI inference clusters with up to 128 superchips (22,528 cores) in a single liquid-cooled rack.
- Dual-socket Superchip configuration with 1.8 TB/s NVLink-C2C connectivity enables massive-scale AI inference clusters with 3 TB memory per superchip pair
Editorial Opinion
Vera signals a fundamental shift in Nvidia's strategy—moving beyond GPU dominance to own the entire AI infrastructure stack. While technically impressive, the success of Vera will ultimately depend on whether AI workloads truly benefit from custom silicon over established CPU architectures; Nvidia's ability to convert whitepaper specifications into real-world performance gains, especially for the contested AI agent hosting use case, remains to be proven in production deployments.
