[DigitalToday reporter Chi-gyu Hwang] Nvidia has released detailed specifications for its AI data centre CPU, Vera, and said it supplied the chip in June to OpenAI, Anthropic and SpaceX.

Nvidia is expanding its push into the AI server market by designing CPUs as well as GPUs.

Vera is Nvidia's first server CPU designed from the core up by the company.

CNBC reported on Monday, local time, that Nvidia said Vera was designed to reduce bottlenecks in AI agent environments and offers 50 percent higher AI agent processing performance than Intel and AMD x86 chips.

The design focuses on single-core speed rather than core count. Hanna Kutang (한나 쿠탕), a Vera product marketer at Nvidia, explained that it was designed to boost GPU utilisation through per-core speed, high memory bandwidth and low latency.

Nvidia is expanding sales of complete rack-level systems beyond selling standalone chips. Vera will be supplied both as a standalone product and combined with GPUs. A liquid-cooled rack can bundle 256 Vera chips, and a configuration that installs 2 Vera chips in a single server is also offered. It will also be released as the 'Vera Rubin' system combined with GPUs.

Nvidia said early AI servers at the time of ChatGPT's 2022 launch used a structure that connected up to 8 GPUs to a single CPU. But as more AI agents run in the background with less human involvement, the importance of CPUs that feed data and handle tasks has grown again, it explained.

Adoption of Vera is still at an early stage. Nvidia disclosed only Oracle among major cloud service providers on its partner list. OpenAI plans to deploy Vera at scale starting this quarter. Vera's power consumption ranges from 250 watts to 450 watts and it supports up to 1.5 terabytes of low-power memory per chip.