Nvidia ships Vera CPUs to OpenAI and other AI customers

Nvidia has released specifications, benchmarks and architectural details for Vera, its first self-developed data center CPU, as it moves into a server CPU market long dominated by Intel and AMD. Company representatives said Vera chips were delivered in June to OpenAI, Anthropic and SpaceX, with OpenAI planning to deploy them at scale starting this quarter. Ian Buck said systems based on Vera Rubin technology are in full production and being deployed across major customers, with CoreWeave, Google Cloud, Microsoft Azure, Meta and Dell also named among adopters.

Vera uses Nvidia’s custom Olympus core rather than an off-the-shelf Arm design. Nvidia says the chip emphasizes single-core frequency, memory bandwidth and low latency to keep GPUs highly utilized during AI agent workloads, and claims 50% higher performance in AI agent tasks compared to x86 chips used by Intel and AMD. The CPU can be sold individually, in dual-Vera servers, in liquid-cooled racks of 256 Vera chips, or as Vera Rubin systems paired with GPUs. Its power consumption ranges between 250 and 450 watts, and a single chip supports up to 1.5TB of memory.

Competition remains difficult. Gartner analyst Kevin Knox said AMD is currently the one to beat in enterprise AI server CPUs, with about a 33% share of the server CPU market, while Intel holds 66.8%. Nvidia claims the entire server CPU market could eventually reach ? billion, while Bernstein estimated the mature server CPU market in 2025 at only about ? billion, underscoring Nvidia’s bet that AI agents will expand demand rather than simply replace existing deployments.