Nvidia Pitches Vera CPU to China With August Delivery Target

Nvidia has begun telling cloud companies in China that they can place orders for its new Vera processor, with deliveries possible as soon as August, Reuters reported on Friday, June 12, citing three sources familiar with the matter.

The pitch is the company's most concrete attempt yet to climb back into a market that US export controls have largely closed to it. Chief Executive Jensen Huang has put the problem bluntly: Nvidia's market share in China has effectively fallen to zero.

Early interest looks real, if cautious. One major Chinese cloud provider plans to order more than 300 servers, each fitted with two Vera chips, to test the hardware before committing to larger deployments, according to the Reuters reporting. Alibaba and ByteDance were among the cloud providers Nvidia named as early Vera partners when it unveiled the chip in March; neither has commented on the new sales push.

What Vera actually is

Vera is not a graphics processor. It is a general-purpose CPU built on Arm architecture with 88 cores, manufactured by TSMC on its 3-nanometer process, and aimed at what the industry calls agentic AI: software that chains together many model calls, tool lookups and decisions. That kind of workload leans harder on conventional processor cores than model training does, which is why CPU demand is tightening just as the industry's center of gravity shifts from training to inference. SK hynix supplies memory across the product family, and Nvidia says the chip is in full production.

Within Nvidia's lineup, Vera is one half of the Vera Rubin platform, where it normally sits beside Rubin GPUs. The China offer deliberately splits the pair. The CPU travels light, regulation-wise. The GPU would not.

It is not a cheap part. Vera processors sell at premium data-center prices, and a fully configured 256-chip rack represents a multimillion-dollar purchase, according to the reports.

Why a CPU can go where GPUs cannot

Washington's export rules target the accelerators used to train and run advanced AI models; general-purpose CPUs face far fewer restrictions. That distinction is the entire strategy here, and recent history explains the urgency. The US government approved around 10 Chinese companies to buy Nvidia's H200, a less advanced GPU, yet none of those chips has been delivered, with Beijing discouraging the purchases to give domestic chipmakers room. Nvidia told investors last month it has recorded no revenue at all from H200 sales to China.

The CPU market Nvidia is walking into, by contrast, is undersupplied. Intel warned Chinese customers in February of server CPU lead times of up to six months, Reuters has reported, and Nikkei reported in March that average waits had stretched from one to two weeks to as long as eight to 12 weeks. AMD said last month that demand keeps exceeding its expectations.

That scarcity is one reason analysts keep raising their forecasts. Bank of America on June 11 lifted its estimate of the 2030 server CPU market to more than 70 billion, from 25 billion, implying 37 percent compound annual growth between 2025 and 2030. The bank's analyst Vivek Arya called agentic AI a powerful demand accelerant that expands the CPU opportunity and lifts both x86 incumbents and ARM challengers.

Huang has framed the ambition in bigger terms. On the company's May earnings call he described the total CPU opportunity as a market worth hundreds of billions of dollars and explicitly included China in that figure, and Nvidia has said it already has significant visibility into CPU revenue this year. Separately, Huang has described China as one of the largest potential annual markets for the company.

Analysts who heard Huang's Computex keynote and follow-up questions this month read the CPU line as deliberate strategy rather than improvisation. Deutsche Bank's team described the company as pursuing a holistic approach to AI, with the newest strategic target area being CPUs to further the company's leadership in GPUs. In plain terms: sell the whole rack, not just the accelerator, and where the accelerator is banned, sell the rest of the rack anyway.

The catches: software, then Washington

Two things could still stall the plan. The first is technical. Data centers in China, like most everywhere, are built around x86 processors from Intel and AMD. Vera is an Arm chip, and Reuters' sources flagged ecosystem and compatibility hurdles as a likely brake on large-scale adoption. The 300-server pilot order reads as exactly what it is: a stress test before anyone commits serious money to rebuilding their software stack around a new architecture.

The second is regulatory. Vera's lighter-touch export classification is a snapshot, not a guarantee, and the H200 episode already showed that approval on paper does not equal chips on racks. Whether US authorities revisit the chip's status as it gains traction in Chinese data centers is the question hanging over the whole effort. Beijing is a variable too: the same industrial policy that froze H200 purchases to protect domestic chipmakers could, in principle, be turned on a foreign CPU that starts winning sockets from local alternatives.

Nvidia shares were little changed on Friday and are up about 10 percent this year. The next fixed point on the calendar is June 24, when the company holds its annual stockholder meeting and is likely to face direct questions about the China plan, and when Qualcomm is expected to make its own AI CPU announcements. If August deliveries happen on schedule, the second half of the year becomes a live experiment: can a 0,000 CPU do what a generation of export-controlled GPUs could not, and get Nvidia paid in China again?