Nvidia Plans $3 Billion Meta Investment as Muse AI Partnership Expands

A $3 billion investment ties Muse more closely to Nvidia

Nvidia plans to invest $3 billion in Meta's equity offering as the companies expand their partnership on Muse, the personal artificial intelligence agent designed to work across email, calendars, payments, shopping and other applications. The agreement combines financing with a wider effort to make Meta's AI models faster and less expensive to operate.

The companies' joint partnership announcement, dated October 27, 2026, says Meta intends to expand its deployment of Nvidia Vera Rubin systems and Vera central processing units. Meta will also deploy Spectrum-X Ethernet networking and BlueField-4 technology to support the movement of data through its computing infrastructure. Nvidia's newsroom version was dated October 26.

Meta increased its planned equity fundraising to $97.5 billion on October 27, its largest fundraising effort to date. Its initial plan, announced the previous day, called for $95 billion to support investment in AI infrastructure.

Muse sits at the consumer end of that spending. Meta wants an assistant that can understand a person's circumstances and complete tasks across several services, rather than simply provide an answer in a chat window. Nvidia supplies much of the computing infrastructure already used by Muse models and agents.

The expanded relationship therefore connects three parts of the same business: the money to build computing capacity, the hardware that runs AI, and the software that people use. For Meta, the immediate challenge is serving growing demand while controlling the cost of each task an agent performs.

How Meta plans to raise the money

Meta's initial $95 billion plan had two parts: $35 billion through underwritten public equity offerings and another $60 billion through an at-the-market offering program. The subsequent increase brought the proposed total to $97.5 billion.

An underwritten offering involves financial institutions arranging the sale of shares under agreed terms. An at-the-market program allows a company to sell shares into the market over time at prevailing prices. Both bring money into the company by selling ownership stakes. Existing shareholders can end up owning a smaller percentage of the business as more shares are issued.

Nvidia, Advanced Micro Devices and Broadcom had expressed interest in purchasing up to $5 billion of Meta shares. Nvidia's announced $3 billion investment equals 60% of that potential combined amount. Purchase commitments from AMD and Broadcom were not specified.

Nvidia describes its own transaction as a proposed investment subject to completion and applicable closing conditions. The companies gave no closing date or detailed timetable for the expanded hardware deployments.

What Muse is supposed to do for users

Muse, known internally as Hatch, is central to Meta chief executive Mark Zuckerberg's ambition to deliver more personal assistance to the billions of people using the company's services. Built using the open source agent OpenClaw as a model, Muse can access connected applications with a user's permission. Users can revoke that access at any time.

An AI agent goes beyond generating text. It can work through steps toward a goal, use software tools, examine results and decide what to do next. Connecting email, calendars, shopping and payment services gives such software more ways to complete a request, while making the scope of its permissions more consequential.

Zuckerberg, Meta's founder and chief executive, described the purpose of Muse in the announcement:

Muse is how we bring personal superintelligence to everyone. We're excited to keep working with NVIDIA to deliver the best personal agents and frontier models for billions of people.

A recent launch already has an international audience

Meta introduced Muse in North America in September. Its Muse launch in Japan on October 23 brought the service to its first market outside North America, following more than 10 million downloads in the month after its debut.

Japanese users gained access through iOS, Android and the web, with many basic functions free and subscriptions starting at 2,799 yen a month. Meta also pursued connections with Mercari, household finance service Money Forward ME and ticket seller Ticket Pia.

The Japan rollout gives the infrastructure agreement a concrete setting. An assistant that continues working after its app closes needs computing resources in the cloud. Requests involving several applications can require repeated model decisions, tool calls and checks before a task is finished.

Meta's Japan announcement said Muse asks for approval before important actions such as sending email or making purchases. Users can choose connected services and review completed and planned operations. A Toronto user, Matt Robb, had earlier said Muse arranged a Facebook Marketplace sale and shared his home address without his knowledge. He later explained that he had selected continuing permission for messages rather than approval for a single use.

For people using the service, the practical issue is how much authority an approval grants and how long it lasts. That question becomes more pressing when an agent can keep acting while its owner is away from the screen.

Why training and running models require different work

Nvidia and Meta say their collaboration will improve both training and inference performance. Training is the process through which a model learns patterns from data. Inference is what happens when a trained model processes a new request and produces an answer or chooses an action.

Muse uses Nvidia infrastructure already. The new work will focus on tuning the computing system and the steps involved in running models, with the stated aims of maximizing performance and reducing total cost of ownership. That cost includes operating a system, not just buying its equipment.

Nvidia says Muse models and agents have achieved performance comparable with leading systems on benchmark tests and attracted rapidly growing downloads. Its partnership announcement gives no named benchmark results, download totals or measured cost reductions.

Meta has also been expanding the models behind its AI products. The Muse Wonder model launch on October 12 introduced Muse Wonder and Muse Wonder Full through developer products and subscriptions. The Muse Spark Open release on October 15 gave developers a model they could download and operate themselves, subject to Meta's license conditions.

Those releases create several kinds of computing demand: developing models, serving requests through Meta's products and supporting tools used by developers. The Nvidia agreement covers the Muse model family as well as the personal agent.

The hardware plan extends beyond AI chips

Meta intends to expand Vera Rubin and Vera CPU deployments for its next generation of agents, with performance per watt as a stated goal. That measure describes how much work a system completes for the electricity it uses, an important operating consideration when AI runs across large data centers.

Central processing units, or CPUs, handle general computing work. In Nvidia's agent architecture, Vera CPUs support the software that coordinates an agent and its tools, while Vera Rubin infrastructure performs model computation.

Spectrum-X Ethernet networking and BlueField-4 address other parts of the system. Networking moves information among machines. BlueField-4 data processing units handle specialized infrastructure functions and can run Nvidia's Sentry security software. Faster model computation has to be supported by systems that move data and manage activity efficiently.

The agreement builds on a partnership announced February 17, 2026, covering large deployments of Nvidia technology for Meta's AI infrastructure and core business. At that time, Meta also said it had adopted Nvidia Confidential Computing for WhatsApp and Spectrum-X networking across its infrastructure.

Nvidia founder and chief executive Jensen Huang described the change he expects agents to bring to everyday computing:

This changes how we use computers, from operating applications to working with intelligent agents.

OpenShell would put boundaries around agent actions

Nvidia and Meta are considering further collaboration on OpenShell for secure deployment and Sentry for monitoring agent behavior. This part of the Muse agreement remains under consideration, while Meta has stated plans to expand the computing and networking deployments.

Nvidia introduced its OpenShell reference architecture on October 21. The architecture combines computing equipment and software to isolate agents, observe their activity and enforce rules about what they can do. A runtime is the environment in which software executes; OpenShell is designed to give agents a controlled environment while they work.

Nvidia says Sentry, running on BlueField-4, can quarantine and stop an agent within milliseconds if it attempts to move outside its software boundary. Monitoring takes place outside the agent's own execution environment, with enforcement supported by the hardware.

Meta is listed among model builders whose agents can run on OpenShell infrastructure. The October 27 partnership announcement goes further by identifying possible joint work specifically around deployment and monitoring for Muse.

For an assistant connected to payments, health services or household devices, access controls operate at more than one level. A user chooses which services the assistant may use. Infrastructure controls define where its software may run and what actions the computing environment will permit.

Nvidia is investing across the AI industry

Nvidia's proposed Meta investment adds to stakes in companies including OpenAI, Anthropic, Intel and xAI. Dealroom data puts Nvidia's participation at 121 investment rounds so far in 2026, compared with 13 in 2022. That is more than nine times the earlier figure.

Those figures count investment rounds rather than dollars invested. They show how much more frequently Nvidia is participating in financing as companies expand their AI businesses.

The Meta agreement gives Nvidia two connected roles: supplying computing infrastructure and becoming an investor in a major customer. Meta, meanwhile, is seeking capital while deepening its technical work with a supplier already embedded in its data centers.

Shares in both companies were largely unchanged in trading before the market opened on October 27. Their stated business goals center on what comes next: more capable agents, better computing efficiency and the ability to serve billions of people.

How the announcements fit together

Meta's recent product releases and Nvidia's infrastructure announcements have arrived in quick succession. The main dates show how the consumer service, underlying models and financing plans connect:

• February 17, 2026: Meta announced a broad partnership with Nvidia covering AI infrastructure, networking and confidential computing.
• September 2026: Meta introduced the Muse personal agent in North America.
• October 12: Meta launched Muse Wonder and Muse Wonder Full.
• October 15: Meta released Muse Spark Open for developers to download and operate under its license.
• October 21: Nvidia launched the OpenShell reference architecture.
• October 23: Muse expanded to Japan.
• October 26: Meta announced an initial $95 billion equity fundraising plan.
• October 27: The joint release detailed the expanded Muse partnership and Nvidia's proposed $3 billion investment; Meta's fundraising plan reached $97.5 billion.

Meta now has to turn that combination into a service people continue to use and pay for. The companies have identified the main technical priorities: improve model performance, reduce operating costs, expand computing capacity and develop controls for agents that act across other services.

Key Points

• Nvidia plans to invest $3 billion in Meta's equity offering, subject to completion conditions.
• Meta increased its planned equity fundraising from $95 billion to $97.5 billion.
• The companies will deepen work on Muse models and agents to improve training, inference and operating efficiency.
• Meta plans expanded Vera Rubin and Vera CPU deployments, alongside Spectrum-X networking and BlueField-4.
• Further collaboration on OpenShell deployment controls and Sentry monitoring remains under consideration.
• Muse users can grant and revoke access to connected applications, making the scope of permissions central to how the agent operates.