# NVIDIA says it will invest $3 billion as the companies deepen their Muse work

NVIDIA says it will invest $3 billion in Meta's Class A equity offering and both companies named model, systems and agent-safety work.[1] Several parts remain intentions rather than delivered products.

NVIDIA announced a planned $3 billion investment in Meta's offering as both companies expanded work on Muse models, NVIDIA systems and possible agent controls.[1] Buyers still need performance, capacity, security and closing evidence.

## What NVIDIA and Meta announced

[NVIDIA's joint announcement [1]](https://nvidianews.nvidia.com/news/nvidia-and-meta-to-expand-collaboration-on-muse-and-muse-family-of-models-amid-robust-growth) says the companies will expand work on Meta's Muse assistant and model family. Their engineers plan to optimize training, inference and workflows, with performance and total cost of ownership as stated goals.

Meta intends to deploy more NVIDIA Vera Rubin systems and Vera CPUs, plus Spectrum-X Ethernet networking and BlueField-4 infrastructure.[1] The release provides no order quantities, delivery schedule, installed capacity or measured cost reduction.

[Reuters also reported [2]](https://www.reuters.com/technology/nvidia-meta-partner-muse-agent-nvidia-invests-3-billion-2026-10-27) the expanded Muse partnership and NVIDIA's planned $3 billion investment on October 27. Reuters described Muse as central to Meta's personal-agent strategy.

## The $3 billion changes one part of Meta's financing story

NVIDIA says it will invest $3 billion in Meta's Class A equity offering.[1] The release's forward-looking statement calls the investment proposed and refers to anticipated completion and closing conditions, so readers should still look for completion and final allocation in offering disclosures.

[Meta's October 26 offering announcement [3]](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Announces-Launch-of-Offering-of-Depositary-Shares-Representing-Series-A-Mandatory-Convertible-Preferred-Stock-and-Offering-of-Class-A-Common-Stock-in-Connection-with-Forward-Sale-Agreements/default.aspx) had named NVIDIA, AMD and Broadcom as potential cornerstone investors for up to $5 billion in aggregate. Meta explicitly said those indications were non-binding.[3]

NVIDIA's named $3 billion plan provides firmer evidence for one supplier's participation, but it does not confirm what AMD or Broadcom will buy. It also does not show that Meta has received the money or deployed infrastructure bought with the proceeds.

## Why the engineering partnership matters

A company that develops models and a company that supplies accelerators, CPUs, networking and data-processing hardware can tune several bottlenecks together. For Muse, the teams can work on kernels, serving, interconnect traffic, memory movement and workload scheduling rather than treating the model and cluster as separate products.

The announcement names better performance and lower total cost of ownership as goals.[1] It does not publish a benchmark, baseline, test setup or target date. Teams evaluating Muse should ask for measured cost per accepted task and p95 latency on their own traffic.

The investment also ties a major supplier financially to a major customer. Customers should distinguish shared commercial incentives from independent evidence about model quality, safety or infrastructure economics.

## OpenShell and Sentry are possible work, not a finished Muse control plane

NVIDIA and Meta say they will consider additional collaboration on OpenShell for safer deployment and Sentry for monitoring agent behavior and data processing.[1] The wording does not say that Meta has deployed either system for Muse.

A useful agent control must enforce policy outside the model, record trusted telemetry and interrupt an unsafe action quickly enough to matter. A product name alone cannot show that those controls work across Meta's connectors, permissions and user-facing tools.

Teams should test denied tool calls, indirect prompt injection, credential access, network egress and emergency quarantine. They should also check whether operators can reconstruct who authorized an action and what data the agent used.

## What builders and buyers should verify next

- Check Meta's final offering documents for NVIDIA's completed allocation and any stated closing conditions.
- Ask which Muse workloads use jointly optimized software and request a reproducible before-and-after benchmark.
- Treat Vera Rubin deployment plans as future capacity until Meta reports installed, powered and available systems.
- Measure accepted-task cost, latency, retries and tool failures rather than relying on aggregate model benchmarks.
- Confirm whether OpenShell or Sentry protects the actual Muse path, and document which actions each control can block.
- Run prompt-injection, permission-denial, credential and network-egress tests with production-equivalent connectors.
- Keep a second model or infrastructure route until the partnership produces measurable service improvements.

## What remains uncertain

- NVIDIA and Meta did not publish the investment's closing date, final share allocation or special commercial terms.[1]
- The companies did not state hardware quantities, delivery dates, regions or usable capacity tied to the partnership.[1]
- No public benchmark supports the promised training, inference or total-cost improvements yet.[1]
- The announcement says the companies will consider OpenShell and Sentry work, so deployment and coverage remain unconfirmed.[1]
- NVIDIA's investment does not confirm whether AMD or Broadcom will participate in Meta's offering.[3]
- Supplier and customer statements do not provide independent evidence that Muse demand, quality or economics will meet their projections.

## Frequently asked questions

- Did NVIDIA invest $3 billion in Meta? NVIDIA announced a planned $3 billion investment in Meta's Class A equity offering. Its release calls the investment proposed and subject to completion and closing conditions.[1]
- Is this separate from Meta's $95 billion financing plan? No. The planned $3 billion belongs to the launched Class A offering within Meta's financing plan, while the joint model and infrastructure work is a separate commercial partnership.[1][3]
- What products are involved? The release names Muse models and agents, Vera Rubin, Vera CPUs, Spectrum-X, BlueField-4, OpenShell and Sentry.[1]
- Has Meta deployed OpenShell for Muse? The release does not say so. It says the companies will consider additional collaboration.[1]
- Does the deal prove lower inference costs? No. The companies state that lower total cost is a goal but provide no benchmark.[1]
- What is the first useful customer check? Ask for a measured Muse result on your workload, including accepted-task cost, latency and the exact infrastructure used.

## Sources

- [[1] NVIDIA: expanded Meta and Muse collaboration](https://nvidianews.nvidia.com/news/nvidia-and-meta-to-expand-collaboration-on-muse-and-muse-family-of-models-amid-robust-growth)
- [[2] Reuters: NVIDIA and Meta deepen AI ties](https://www.reuters.com/technology/nvidia-meta-partner-muse-agent-nvidia-invests-3-billion-2026-10-27)
- [[3] Meta: equity offering and financing plan](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Announces-Launch-of-Offering-of-Depositary-Shares-Representing-Series-A-Mandatory-Convertible-Preferred-Stock-and-Offering-of-Class-A-Common-Stock-in-Connection-with-Forward-Sale-Agreements/default.aspx)