Nvidia’s Vera Is a CPU “Built for Agents, Not Humans” — 88 Custom Cores Aimed at Intel and AMD’s Last Stronghold

“This CPU is built for agents,” Jensen Huang told the COMPUTEX 2026 audience. “All the CPUs of the past we built for humans.” With that, Nvidia introduced Vera — its first central processor purpose-built for agentic AI — and planted a flag in the one corner of the data center it had always ceded to Intel and AMD: the CPU.

The premise is that agents change what a CPU is for. When software does not just answer a prompt but executes multi-step tasks, the CPU lands squarely on the critical path — running Python runtimes, sandboxed code execution, tool calls, data retrieval, orchestration logic and analytics pipelines while the GPUs wait on it. Nvidia’s framing is that the design goal flips from “cores per dollar” to “tokens per dollar”: every millisecond the CPU saves on a tool call is accelerator time that doesn’t go to waste.

Vera’s answer is 88 custom Olympus cores — Nvidia’s own Arm-based core design — paired with Spatial Multithreading that partitions core resources into 176 threads for predictable throughput under load. Nvidia says Olympus delivers up to 50% higher instructions-per-cycle than its Grace core, using a neural branch predictor that sustains two taken branches per cycle, a 10-wide decode unit and roughly 40% lower peak memory latency than x86. Feeding it is an LPDDR5X memory subsystem pushing up to 1.2 TB/s of bandwidth while drawing under 30 watts, versus the 100-plus watts of comparable DDR5 setups.

The performance claims are aimed straight at the incumbents. Nvidia says Vera delivers more than 1.8x the sandbox throughput of x86 across agentic workloads under full load — code compilation, analysis and Python execution — and runs about 63% faster than its 72-core Grace predecessor on a geomean of tests. Early independent benchmarks backed the direction of travel: Phoronix reported the 88-core Vera outpacing both AMD EPYC and Intel Xeon parts in its debut runs. Inside a Vera Rubin system, Vera connects to the Rubin GPUs over NVLink-C2C at up to 1.8 TB/s of coherent bandwidth, so the chip is both the orchestrator and the GPUs’ closest neighbor.

Vera ships two ways: as the host CPU in the flagship NVL72 Vera Rubin racks, and standalone in CPU-only server racks — dense liquid-cooled configurations for large-scale agentic AI and flexible two-socket air-cooled boxes for enterprise and cloud. Nvidia named a heavyweight list of early evaluators, including Anthropic, OpenAI, SpaceX, NYSE, ByteDance, CoreWeave and Oracle Cloud Infrastructure, with systems from Dell, HPE, Lenovo and Supermicro plus Taiwan builders due starting this fall.

Strategically, Vera is the most aggressive part of Nvidia’s COMPUTEX blitz. Having already won the accelerator, Nvidia can now sell the entire agentic rack — GPU, CPU and the NVLink fabric between them — and bill itself as the only vendor whose silicon is designed end to end for how agents actually run. For Intel and AMD, the x86 server CPU has been the last defensible tier of the AI data center; an Arm-based Vera that beats their flagships on the workloads everyone is racing toward turns that stronghold into contested ground.