Nvidia pitches Vera CPU to China as US export curbs stall GPU sales

Nvidia has told Chinese clients that its new Vera central processor for AI data centers could be available as soon as August and that orders can be placed now, three sources familiar with the matter said, as the company pivots to a product category that faces fewer US export restrictions than its graphics processors.

"Chinese clients have shown interest in Vera because it sits in a different regulatory category than our GPUs," a person familiar with Nvidia's sales strategy said. The company declined to comment officially.

Vera is Nvidia's first standalone CPU built for agentic AI — systems that perform tasks autonomously. The Arm-based chip packs 88 Olympus cores with 176 threads and 1.2 terabytes per second of memory bandwidth, delivering a 1.5 times improvement in instructions per cycle over its predecessor Grace, according to Nvidia's published specifications. A single Vera processor costs "well north" of $20,000 before bulk discounts, and a fully configured rack of 256 chips runs to about $10 million depending on memory configuration, according to research firm SemiAnalysis.

The sales push comes as Nvidia's China market share has "effectively fallen to zero," Chief Executive Officer Jensen Huang said in October, hurt by US export controls on advanced chips and Beijing's push for domestic alternatives. The company generated $4.6 billion in H20 revenue from China in the comparable quarter last year, but its latest outlook excludes China data center compute revenue after licensing uncertainty blocked a clean path for GPU sales. Washington has licensed about 10 Chinese firms to buy the H200 graphics processor, but not a single delivery has been made, as Chinese officials have withheld approval to nurture domestic suppliers.

The Vera bet

Nvidia expects $20 billion in revenue from Vera chip sales by the end of its fiscal year in January, Huang said at the chip's unveiling in March. One major Chinese cloud company plans to place an order for more than 300 servers, each containing two Vera CPUs, one of the sources said. The company plans to deploy the systems for testing first and decide whether to place official orders based on the results.

Leading cloud firms including Alibaba and ByteDance are collaborating with Nvidia to deploy Vera, the company said at launch, though it did not confirm whether ordering had begun. Chinese clients plan to initially deploy Vera chips only in their overseas data centers for testing, one source said.

Current US export controls are written around accelerator performance, memory bandwidth, and related thresholds that apply to graphics processing units. A purpose-built CPU running orchestration and agent coordination workloads does not map as cleanly onto those parameters, giving Nvidia a regulatory opening that may not stay open indefinitely.

Competitive crossfire

Vera puts Nvidia in direct competition with Intel and AMD, which have dominated the processor market with the x86 architecture for decades. Intel has notified Chinese customers of server CPU delivery lead times of up to six months, Reuters reported in February. AMD last month flagged that the global CPU market is "tight," with demand outpacing forecasts.

AMD is pushing back hard. The company published performance projections claiming its upcoming Epyc Venice platform — built on Zen 6 architecture with up to 256 cores and 512 threads on TSMC's 2nm process — will deliver 3.3 times Vera's per-rack performance. AMD also projects a 70% overall improvement in performance and efficiency over its current Turin generation. The benchmarks were built using methodology drawn from Phoronix's early Vera testing, which was conducted at Nvidia's headquarters under conditions Nvidia controlled.

Whether initial Chinese interest translates into large-scale adoption remains uncertain, partly due to software ecosystem compatibility issues and the constraints of migrating workloads built around domestic AI chips, one source said. Nvidia shares trade at about 35 times forward earnings. The company's ability to convert Vera orders into recognized revenue will depend on how quickly hyperscalers standardize on tightly connected Nvidia systems rather than buying CPUs and GPUs on separate tracks.

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