NVIDIA has begun taking orders from customers in China for Vera, its first standalone data center CPU, with deliveries expected to start as early as August. Vera packs 88 Arm cores and is tuned for running fleets of AI agents, marking a push to broaden a business that has long centered on GPUs. With US export controls freezing GPU sales to China, the move also points to a strategy of using a CPU, which is less likely to fall under those restrictions, as a way back in.

What kind of processor is Vera

Vera is a data center CPU whose details NVIDIA unveiled at GTC in March 2026. Until now, the company's server products revolved around GPUs, with CPUs added as supporting parts. Vera changes that relationship and becomes the first product NVIDIA can sell as a CPU in its own right. It is already in full production.

Inside are 88 of NVIDIA's own "Olympus" cores, compatible with the Armv9.2 instruction set. A mechanism that lets each core handle two streams of work at once brings the total to 176 threads. Each chip carries up to 1.5 TB of LPDDR5X memory and 1.2 TB per second of memory bandwidth, a design built to move large volumes of data quickly.

On the performance side, a defining trait is that all 88 cores are treated as a single group rather than being split apart. In many-core x86 CPUs, cores tend to be divided into several clusters, and exchanging data across those clusters can introduce latency. Vera removes that partitioning and smooths out core-to-core communication. NVIDIA says instructions per cycle (IPC) rose 1.5 times over the previous generation, translating into roughly 50 percent higher real-world performance.

Why it is aimed at agentic AI

Vera targets the job of running agentic AI that carries out tasks on its own. Where a chat-style AI returns a single answer to a single prompt, an agent breaks a goal into many steps and repeats work such as compiling code, querying databases, and browsing the web. That kind of flow increases the share of work handled by the CPU, so a powerful CPU matters alongside the GPU.

NVIDIA also prepared a dedicated rack that holds 256 Vera chips with liquid cooling. Compared with a previous-generation CPU configuration, it claims up to 6 times the CPU throughput and 2 times the performance on agentic AI workloads. As the center of gravity in AI shifts from training models to running trained ones, or "inference," the design is positioned to capture the rising CPU load that inference brings. CPU demand is tight, and Intel is reported to have quoted Chinese buyers waits of six months.

Price and the scale of the business

Vera is not a cheap product. A single chip is said to cost well north of 20,000 USD (about 3.2 million yen) before discounts, and a single rack holding 256 chips runs to roughly 10 million USD (about 1.6 billion yen). NVIDIA expects 20 billion USD (about 3.2 trillion yen) in Vera revenue by the end of January 2027.

※1 USD = 160 JPY

Early users include AI developers such as Anthropic and OpenAI. The CPU market has long been held by Intel and AMD with x86 chips, so NVIDIA, already dominant in GPUs, is now stepping squarely into the market for the processor itself.

A "side door" into the China market

What stands out about this round of orders is the China business context. NVIDIA's China operations have fallen sharply under US export controls, and shipments of the H200, its second most powerful AI chip, are said to remain halted. Because the controls mainly target high-end GPUs and are less likely to cover general-purpose CPUs, there is room for Vera to slip through. Even so, whether an advanced CPU like Vera would require an export license remains unclear.

The Chinese side is moving cautiously. One major cloud provider is reported to have ordered more than 300 servers, each carrying two Vera chips, intending to start with test deployments. For now the plan is to run them in overseas data centers rather than inside China, and full adoption hinges on software compatibility and the cost of migrating off domestic chips. China is pushing hard for semiconductor self-reliance, and even as it nurtures its own alternatives, Alibaba and ByteDance have been named as Vera collaborators, creating a tangled picture.

Summary

Vera is NVIDIA's first standalone data center CPU, built around 88 Arm cores and tuned for agentic AI. It is a product that extends the company's GPU-built strength into the CPU market, and at the same time a move to use a CPU, which regulators are less likely to reach, as a foothold while GPU sales to China stay frozen. Whether Chinese customers commit to full-scale adoption or keep it to trials remains to be seen. The August shipping milestone looks set to be the first real test.