NVIDIA launches Rubin superchips for agentic AI infrastructure

On January 5, 2026, NVIDIA unveiled the Rubin (Vera Rubin) platform at CES. It introduced six Rubin superchips and an AI-native storage architecture for gigascale inference and long-context multi-agent workloads. The announcement appeared in NVIDIA’s newsroom and was independently reported by TechCrunch, The Verge, and WIRED. (NVIDIA; TechCrunch; The Verge; WIRED)

Rubin pairs Rubin-class accelerators with an AI-native storage layer. The layer preserves and serves very long model context and large agent state at low latency. That design targets multi-agent and chain-of-thought workloads. CEO Jensen Huang told WIRED the chips are entering “full production.” The platform targets cloud and on-prem deployments that need sustained, high-throughput inference and extended context windows. (NVIDIA; WIRED)

Rubin aims to reduce engineering overhead by integrating storage, model context, and accelerators for agentic applications. Cloud and enterprise teams are likely to test Rubin nodes for multi-agent coordination, long memory, and gigascale throughput. Early procurement and partner integrations will reveal real-world performance and costs.

Why It Matters

• Rubin integrates accelerators with AI-native storage to preserve much longer model context without custom data plumbing—useful for multi-agent workflows and long-recall agents.
• Teams running gigascale inference can evaluate Rubin nodes to reduce end-to-end latency and sustain high-throughput chained-agent tasks.
• Cloud and on-prem infrastructure teams should validate context preservation, storage I/O, and provisioning costs in pilots before production rollouts.