NVIDIA Readies 'Plan B' for China: GPUs Blocked, But Vera is a CPU and Prepares Orders

After the strong restrictions imposed by the United States on the sale of AI accelerators in China, NVIDIA is attempting to open new commercial opportunities through its first standalone CPU for data centers. The company has initiated contacts with various Chinese customers to propose Vera, a processor based on Arm architecture designed for agent-based artificial intelligence applications.

According to Reuters, orders can already be placed, and the first deliveries could start in August. The decision to focus on CPUs represents an alternative solution compared to the GPU market, which continues to face heavy commercial restrictions. NVIDIA itself admitted in recent months that its market share in China has practically vanished due to export controls.

Vera has already entered large-scale production, and the first units have been distributed to U.S. AI companies. The focus is now shifting towards the Chinese market, where several cloud and data center operators have expressed interest in the new platform.

NVIDIA Vera integrates 88 Arm Olympus cores, accompanied by 164 MB of L3 cache and 166 MB of L2 cache. The CPU supports up to 1.5 TB of LPDDR5X SOCAMM2 memory, while rack configurations can reach up to 400 TB of total unified memory. The platform also includes support for PCIe Gen6, CXL 3.1, and air or liquid cooling systems, with a TDP ranging from 250 to 450 watts (per unit).

One of the most interesting elements is represented by the Vera MGX racks, configurations that can host up to 256 Vera CPUs connected via high-speed NVLink. NVIDIA claims that such systems can deliver up to 90% higher performance in AI applications compared to traditional x86 solutions.

The interest from Chinese customers appears to be substantial. A large cloud company is said to have already planned the initial purchase of over 300 servers, each equipped with two Vera CPUs. The current goal is to finalize a validation phase before proceeding with any large-scale implementations.

The economic factor also highlights the ambitions of the project. Estimates indicate a price above $20,000 for each Vera CPU, while a complete rack of 256 processors would reach around $10 million. Despite the high costs, NVIDIA considers the Chinese market too important to forgo a direct presence.

The timing seems favorable. The AI sector is increasingly focusing resources on inference, an activity that requires a growing role for CPUs alongside specialized accelerators. In this scenario, NVIDIA enters direct competition with AMD and Intel, companies that have dominated the x86 server processor market for years.

The challenge involves not only historical manufacturers. More and more companies engaged in AI model development are designing proprietary processors to reduce dependence on external suppliers. NVIDIA, however, aims to become a key player in the CPU sector by 2026 and views Vera as a crucial piece of its expansion.

It remains to be seen what the definitive response from the Chinese market will be. Software compatibility, migration costs from existing platforms, and China’s growing focus on national technologies represent factors that could influence the adoption of the new CPU. For NVIDIA, however, Vera presents the opportunity to return to generating significant profit in a market that, undeniably, has become crucial for the future of artificial intelligence.