Nvidia to invest $3B in Meta as the two deepen work on Muse AI agent

Nvidia will invest $3 billion in Meta’s equity offering, the two companies announced on Tuesday, October 27, as they deepen their partnership on Muse, Meta’s personal AI agent, and the related Muse family of models. The stated goal of the expanded work is to improve inference and training performance for Muse, which the companies frame as a response to “popular demand for useful AI agents”.

The deal pairs a chip supplier’s cash with a customer’s hardware commitments: Meta says it intends to expand its use of Nvidia’s next-generation Vera Rubin systems and Vera CPUs, along with Nvidia networking. The agent-safety tooling in the announcement, however, is only at the “consider” stage.

What Nvidia and Meta committed to on October 27

The announcement came as a joint press release distributed through GlobeNewswire. Its main points:

• Equity investment. “In addition, NVIDIA will invest $3 billion in Meta’s equity offering,” the release says.
• Joint optimisation. The companies “intend to further deepen their collaboration by partnering on optimizations across workflows to maximize performance and improve total cost of ownership”. Total cost of ownership (TCO) is the full cost of running a system over its life, including hardware, power and operations.
• Hardware. Meta “intends to expand the deployment of NVIDIA infrastructure, including Vera Rubin and Vera CPUs for next-generation AI agents to deliver maximum performance per watt,” and will deploy Spectrum-X Ethernet networking and BlueField-4 to support network efficiency and throughput.
• Safety tools, not yet agreed. Meta and Nvidia “will also consider additional collaboration opportunities” on Nvidia’s OpenShell Reference Architecture, including OpenShell for safe deployment and Nvidia Sentry for monitoring agent behaviour.

Reuters summarised the arrangement as Muse models and agents, which already run on Nvidia infrastructure, working more closely with the chipmaker “to improve performance and reduce costs”. Seeking Alpha’s report, carried on TradingView, describes the same terms.

Item	What the sources say	Status
$3 billion investment	Into Meta’s equity offering	Announced; explainx.ai notes completion and closing conditions are referenced in forward-looking statements
Inference and training work on Muse	Expanded partnership	Announced
Vera Rubin and Vera CPUs	Meta “intends to” expand deployment	Stated intent
Spectrum-X Ethernet, BlueField-4	Meta will deploy for network efficiency and throughput	Stated plan
OpenShell and Sentry	“Will consider”	Not agreed
Muse benchmark figures	None published	Claim only

How the $3 billion fits Meta’s $97.5 billion raise

The investment sits inside a much larger financing. Earlier on the same day, Meta increased its equity offering to $97.5 billion, which Reuters describes as the largest fundraising in its history, as Meta accelerates investment in AI infrastructure.

EM360Tech breaks the offering into three parts: $25 billion of Class A common stock, upsized from $22.5 billion; 250 million preferred depositary shares priced at $50 each, around $12.5 billion; and, citing Stocktwits, an at-the-market share sale programme of about $60 billion. An at-the-market programme lets a company sell new shares gradually into the open market rather than all at once. EM360Tech also relays a Stocktwits report that Nvidia, AMD and Broadcom had agreed to buy up to $5 billion of the common shares between them.

Market commentary noted on 24/7 Wall St. attributes to CNBC’s Jim Cramer the view that Meta’s $25 billion equity raise was five times oversubscribed, but that the $60 billion at-the-market programme worries him, because he expects the selling to weigh on the stock.

The release itself is thin on deal mechanics. According to explainx.ai’s reading of it, it gives no share price, stake size or closing date.

What Muse is, and what Meta says it can do

Muse is central to Meta CEO Mark Zuckerberg’s plan to deliver “personal superintelligence” to the billions of people who use Meta’s services daily, Reuters reports, adding that the agent is known internally as Hatch. Modelled on the open-source AI agent OpenClaw, Muse can access a user’s email, calendar, payments, health, shopping and smart-home applications, and users can revoke access at any time.

EM360Tech reports that Meta launched Muse on 8 September and that, according to Meta, it runs on Muse Secure VM, a dedicated virtual machine with its own browser that holds both the agent and the user’s data. Meta also says a separate agent called Sentinel must approve anything Muse sends to the internet, and that Muse checks with the user before sensitive actions such as sending an email or making a purchase. On 12 October, Meta introduced Muse Wonder and Muse Wonder Full, which it calls its most capable models to date for agentic tasks.

The family also includes an open-weight model. explainx.ai summarises an October 15 Nvidia developer blog post on Muse Spark Open, describing an 816-billion-parameter mixture-of-experts model with 42 billion active parameters and a 1-million-token context window, with NVFP4 checkpoints for Blackwell Ultra. A mixture-of-experts model routes each token through only a subset of its parameters, which is why “active” parameters are far fewer than the total. The same post, per explainx.ai, shows a curve claiming over 50,000 tokens per second per GPU, with the caveat that it assumes an acceptance length of 6 and gives limited methodology.

The maker’s claims versus independent evidence

The release states that Meta’s Muse models and agents “have delivered frontier-level benchmark performance and surging consumer downloads since launch”. No benchmark, download count or model name accompanies that sentence; explainx.ai calls it a company claim that cannot be checked from the release. The sources provide no independent leaderboard result for any Muse model.

The executives’ quotes are about direction rather than numbers. Nvidia founder and CEO Jensen Huang said:

“AI agents represent a new computing paradigm — software that understands our goals, reasons, uses tools and gets work done on our behalf.”

He added that the companies are expanding the partnership “to optimize the computing infrastructure end to end for capable, efficient and secure AI agents”. Meta founder and CEO Mark Zuckerberg said:

“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.”

One detail in the release is inconsistent. Its summary ties Sentry to “structured, unstructured and graph data processing,” while the body describes Sentry as a tool for monitoring agent behaviour. explainx.ai flags the same split.

Why a chip supplier is buying into its customer

The release does not explain the rationale for the investment. Coverage offers interpretations, clearly labelled as such.

24/7 Wall St. argues the deal converts “what had been a straightforward customer relationship into a direct ownership stake,” giving Nvidia a financial interest in Meta’s success while Meta depends on Nvidia hardware. Because Vera Rubin, Vera CPU, Spectrum-X and BlueField-4 are Nvidia’s next-generation data-centre products, the outlet says the agreement “effectively pre-loads demand for hardware that in some cases has not yet reached full commercial scale”.

explainx.ai, in what it labels its own reading rather than a company statement, writes that Nvidia selling chips to Meta and then putting $3 billion into Meta’s equity offering “looks circular, and some readers will say so”. In that reading, a stake ties a large buyer closer at a time when hyperscalers are also building custom silicon, while Meta gains capital, a deeper engineering relationship and a stated focus on lower TCO for a service used by billions.

The deal extends an existing relationship. explainx.ai, citing Unite.AI, says Meta announced a multi-year infrastructure partnership with Nvidia in February 2026 that included Nvidia Confidential Computing for WhatsApp private messaging and Spectrum-X Ethernet. EM360Tech places the deal in a wider pattern of Nvidia “writing bigger cheques across the AI market”.

What it means for builders, and the open risks

For developers working with Muse models, explainx.ai’s assessment is that the announcement changes “very little” today. Muse Spark Open already has Nvidia-optimised checkpoints and works with TensorRT-LLM, SGLang and vLLM, according to the October 15 Nvidia post it summarises. The new commitment promises more joint optimisation across training and inference, but with no performance or cost figures attached.

Points the sources raise for anyone building on or alongside agents like Muse:

• Agent access is a real operational risk. EM360Tech notes that agents with access to email and payments carry real operational risk, citing an Axios report that a Meta employee went viral after an OpenClaw agent deleted files from her Mac once she gave it access to her email.
• Meta admits Muse can err. Meta has acknowledged that Muse can still make mistakes, while saying the agent has reached the minimum safety threshold for release.
• Data questions remain. Meta has faced scrutiny over its data collection practices before, and EM360Tech says Muse’s reach into health and payments data “will keep that question live”.
• Safety tooling is not committed. The Nvidia tools in the announcement target this problem but “remain at the consideration stage,” and Meta has not said when, or whether, OpenShell or Sentry will ship with Muse.
• Hardware dependence. explainx.ai says co-optimising Muse for Vera Rubin makes Nvidia hardware “the path of least resistance” for the model family, in its own analysis. 24/7 Wall St. similarly concludes that AI capability is increasingly tied to specific next-generation infrastructure across networking, compute and accelerators.

Nvidia has been working on agent security beyond this deal: EM360Tech reports it launched the Open Secure AI Alliance in July after the Hugging Face breach and said it plans to contribute open tools to the software that controls AI agents.

What to watch next

• Closing of the investment. The release references completion and closing conditions in its forward-looking statements, per explainx.ai, so the final terms and timing are still to come.
• A decision on OpenShell and Sentry. Whether Meta adopts Nvidia’s deployment and monitoring tools for Muse is the clearest open question in the announcement.
• Numbers behind the claims. Any published benchmark results or download figures would allow the “frontier-level” claim to be checked.
• Meta’s at-the-market selling. How the roughly $60 billion programme is executed is a point of concern in market commentary.
• Nvidia’s next results. 24/7 Wall St. reports Nvidia has set November 17 for its third-quarter results, covering the quarter ended October 25.