Nvidia’s first Vera CPU systems are now in customers’ hands. The company hand-delivered the servers to OpenAI, Anthropic, SpaceXAI, Oracle and AWS, capping a week in which AWS also received its first Vera Rubin GPU unit.
The AWS handoff came with a much larger commitment. Nvidia said the two companies, partners for 16 years, agreed to add 2 million GPUs to their collaboration and to build Vera CPU-based infrastructure for the cloud giant.
Agents spend much of their time orchestrating tools and fetching context between model calls, and Nvidia built Vera for exactly that job. The chip combines 88 in-house cores with memory bandwidth of 1.2TB/s, and per-core throughput on agent-style work runs up to 1.8x higher than earlier designs, the company says.
The deliveries push the chip from announcement into production, with all five partners now running early systems. Buck, who made the rounds, has argued that AI factories need processors that keep up with agents’ busywork, not just the GPUs that answer queries.
