NVIDIA and Meta deepen Muse collaboration on training and inference

NVIDIA and Meta said they will expand their collaboration on training and inference for Meta’s Muse model family. Joint workflow optimization is the concrete workstream; deployment details and claimed performance gains remain unquantified.

The work aims to improve performance and reduce total cost of ownership across Muse workflows. NVIDIA and Meta disclosed no benchmarks, cost baselines, engineering milestones, contract duration or financial value for the technical collaboration. For now, those benefits are company targets rather than demonstrated results.

Meta intends to expand its use of NVIDIA infrastructure for next-generation AI agents, including Vera Rubin and Vera CPUs. It also plans to deploy Spectrum-X Ethernet networking and BlueField-4. The announcement gave no start date or deployment scale and did not confirm that any of those deployments were complete.

The companies are also considering NVIDIA OpenShell for agent deployment and NVIDIA Sentry for monitoring agent behavior. OpenShell is a reference architecture for deploying AI agents with security controls. Both technologies were presented as possible additional work, not committed deployments.

The October announcement builds on a multi-year infrastructure partnership announced in February covering AI training, inference, networking and model optimization. The Muse family includes Muse Wonder models. Meta separately said Muse Wonder was available through the Meta Model API and Muse Code, while use in more Muse experiences remained planned, beginning with subscribers.

NVIDIA’s announcement also said it would invest $3 billion in Meta’s equity offering. A Meta offering disclosure from October 26 described NVIDIA, AMD and Broadcom as having nonbinding indications of interest for as much as $5 billion in aggregate. The documents reviewed do not establish that NVIDIA’s proposed investment was binding or completed.