NVIDIA built a CPU just for AI agents

For the longest time, the AI race felt like a pure GPU numbers game. But as modern models evolve into agents that write Python, execute code in sandboxes, and query databases, a new bottleneck has emerged: the GPUs spend way too much time waiting around for slow CPUs.

At GTC Taipei, NVIDIA took direct aim at that problem by officially launching Vera, its first CPU built specifically for AI agents, and confirmed the chip is already in full production.

An 88-core engine built for sandbox execution

In modern AI factories, the economics are shifting from cores per dollar to tokens per dollar. An agent doesn't just predict the next word anymore; it orchestrates workflows, spins up micro-environments, and iterates on code.

Vera tackles this with 88 custom Olympus cores, spatial multithreading, and an LPDDR5X memory subsystem that pushes up to 1.2TB/s of bandwidth. NVIDIA claims it finishes tasks 1.8x faster than traditional x86 CPUs. Open-source benchmark suite Phoronix confirmed that Vera delivered the fastest overall performance across agent-critical workloads like code compilation, Python, Java, and database processing.

NVIDIA's previous Grace CPU already shipped nearly 2.5 million units, but Vera is expanding the footprint. It powers standalone CPU servers, serves as the host processor for the Vera Rubin superchip with 1.8TB/s coherent NVLink-C2C bandwidth, and runs inside the new Vera BlueField-4 STX storage platform.

The adoption list is basically a who's who of AI infrastructure. OpenAI, Anthropic, and SpaceXAI are planning to adopt it, alongside hyperscalers like Oracle Cloud Infrastructure, ByteDance, and CoreWeave. Even the New York Stock Exchange, which processes over 1.1 trillion messages daily, is testing Vera with HPE and Redpanda to power its financial infrastructure. Systems from Dell, HPE, Lenovo, and Supermicro will begin rolling out this fall.

Why it matters

This looks like one of the most calculated architectural plays NVIDIA has made in years. Agentic AI is brutal on general-purpose hardware because hundreds of sandbox calls and verification loops happen before a single coherent answer is returned. By combining custom Olympus silicon with massive 1.2TB/s memory bandwidth, Vera turns the host processor into a purpose-built conduit that keeps high-priced GPUs fully fed.

For the industry, this is an existential threat to Intel and AMD in the data center. Up until now, data center operators bought NVIDIA for GPUs and stuck with x86 for host servers. With major OEMs like Dell and HPE now selling standalone Vera server configurations, it establishes the first widely supported non-x86 standard compute tier in mainstream enterprise racks.

If you are an individual developer or running a modest startup, there is no need to jump on standalone hardware right now. These dense liquid-cooled racks and dual-socket enterprise boxes are engineered for hyperscale AI factories. We believe the real win for most builders will come later this fall, when cloud providers like OCI integrate Vera into their reasoning clusters, likely dropping the latency and cost of long-running, multi-step agent queries.