NVIDIA and Meta expand Muse partnership as NVIDIA commits $3 billion to Meta's equity offering

1 hour ago

7 min read

NVIDIA and Meta have announced an expanded collaboration on Muse, Meta's AI assistant, and on the wider Muse family of models, and NVIDIA has said it will invest $3 billion in the equity offering that Meta launched on October 26 to finance its AI spending. The announcement, published in NVIDIA's newsroom on October 26, 2026, ties together two threads that until now ran in parallel: Meta's push to turn Muse into a frontier assistant and agent platform, and the equity raise that Meta is using to help finance its AI spending.

The technical side of the agreement is framed around inference and training performance. According to NVIDIA, the two companies will run joint optimizations across Meta's workflows to raise performance and lower total cost of ownership, and Meta intends to expand its deployment of NVIDIA infrastructure for next-generation agents, including Vera Rubin systems and Vera CPUs, with Spectrum-X Ethernet networking and BlueField-4 data processing units for network efficiency and throughput. The companies also said they will consider further work on NVIDIA's OpenShell reference architecture for safe agent deployment and on NVIDIA Sentry for monitoring agent behavior.

The financial side is short on detail. NVIDIA's release states the $3 billion figure and nothing else about price, share class or timing, while its forward-looking statements refer to the "anticipated completion" of the investment and "any applicable closing conditions". Meta's preliminary prospectus, filed with the SEC the same day, adds the missing context: NVIDIA is one of three chip suppliers named as cornerstone investors in the common-stock part of the offering.

··········

WHAT NVIDIA AND META ANNOUNCED.

The release describes the expansion as a response to demand for useful AI agents. Meta says that its Muse models and agents, which run on NVIDIA infrastructure, have posted frontier-level benchmark results and strong consumer download growth since launch. The release gives no download counts and no benchmark figures, so those claims remain Meta's characterization rather than measured results that readers can check.

Mark Zuckerberg, Meta's founder and CEO, summarized the company's position in a single line: "Muse is how we bring personal superintelligence to everyone." He added that Meta will keep working with NVIDIA on personal agents and frontier models. Jensen Huang, NVIDIA's founder and CEO, described AI agents as "a new computing paradigm", systems that understand goals, reason, use tools and complete work on a user's behalf, and said the partnership is meant to optimize the computing stack end to end so those agents are capable, efficient and secure.

The table below separates what the release presents as a commitment from what it presents as an intention or an area of study, because the wording differs from item to item.

........

Area	What the release says	Strength of the wording
Muse models	Joint optimization of inference and training across workflows	Planned collaboration
Compute	Expanded deployment of Vera Rubin and Vera CPUs	Meta "intends" to expand
Networking	Spectrum-X Ethernet and BlueField-4	Meta will deploy
Agent security	OpenShell reference architecture and Sentry monitoring	Further work to be considered
Investment	$3 billion in Meta's equity offering	Subject to completion and closing conditions
Timelines	None stated	Not disclosed

........

No timeline is given for any of the infrastructure items, and the release does not say how many systems Meta plans to add, where they will be installed or how the expansion relates to the existing NVIDIA capacity in Meta's data centers.

··········

HOW THE $3 BILLION FITS INTO META'S OFFERING.

Meta opened its offering on October 26 with two separate transactions that are not contingent on each other, as described in our earlier report on Meta's $35 billion stock offering and the planned $60 billion at-the-market program. The preliminary prospectus supplements, both marked as subject to completion, set out the structure at launch.

........

Component	Terms in the preliminary prospectus
Class A common stock	$22.5 billion base, plus a $3.375 billion underwriters' option
Settlement	Forward sale agreements, settlement expected on or before November 19, 2027
Mandatory convertible preferred	250,000,000 depositary shares at $50, plus a 37,500,000-share option
Preferred conversion	Mandatory conversion expected on August 15, 2029
Cornerstone investors	NVIDIA, AMD and Broadcom, up to $5.0 billion in aggregate
Use of proceeds	General corporate purposes

........

The common-stock prospectus states that NVIDIA, Advanced Micro Devices and Broadcom have "severally and not jointly" indicated an interest in purchasing up to an aggregate of $5.0 billion of Class A shares "at the same price and on the same terms and conditions as the other purchasers in this offering". The same document stresses that these indications are not binding: any of the three may buy more, fewer or no shares, and the underwriters may sell them more, fewer or none.

NVIDIA's announcement turns its part of that indication into a stated figure. Read together, the two documents suggest that NVIDIA accounts for $3 billion of the cornerstone interest, leaving up to $2 billion attributable to AMD and Broadcom combined, although neither company has published its own amount and the prospectus does not break the total down. This allocation is our reading of the two documents, not a figure either company has disclosed.

The forward-sale structure matters for anyone trying to connect the investment to Meta's spending. Meta's prospectus says the company does not expect to receive proceeds when the common-stock offering closes, only when the forward sale agreements are settled, which is expected on or before November 19, 2027. Any cash attributable to NVIDIA's purchase therefore reaches Meta's balance sheet on the forward-settlement schedule rather than immediately, and Meta describes the intended use of that cash only as general corporate purposes.

··········

WHY THE CHIP SUPPLIERS ARE BUYING IN.

The three cornerstone investors are all chip companies that sell into AI data centers. NVIDIA supplies the GPUs, CPUs and networking named in the partnership, AMD sells competing accelerators, and Broadcom sells custom silicon and networking. Their participation turns part of the equity raise into a transaction with companies that compete for a share of the capital spending that hyperscalers such as Meta are committing to AI infrastructure.

That spending is large. In its third-quarter 2026 results, released on October 21, Meta reported capital expenditures of $34.91 billion for the quarter, including principal payments on finance leases, and narrowed its full-year 2026 guidance to $135-145 billion from a prior range of $130-145 billion. The same release reported revenue of $64.60 billion, up 26% year over year, and a headcount of 66,775, down 15%. Meta has not given capital expenditure guidance for 2027.

The NVIDIA release does not describe the investment as conditional on purchases, and nothing in either document links the $3 billion to a hardware order. The pattern of a supplier buying equity in a large customer while that customer expands its deployment of the supplier's systems is, however, the context in which investors will read the agreement, and it is the main reason the financial terms deserve the same attention as the technical ones.

··········

WHAT THIS MEANS FOR MUSE USERS AND DEVELOPERS.

For people who use Meta's assistant, nothing changes immediately. The release announces no new Muse features, no new plans or prices and no new regions, and it gives no date for any improvement that the joint optimization work might produce. The practical effect, if the plans are carried out, would show up later as more capacity and lower serving costs for Meta's models, which in turn affect how fast responses arrive and how widely the most expensive models can be offered.

The models that would benefit are already in use. Meta introduced Muse Wonder and Muse Wonder Full on October 12, describing Muse Wonder as available to developers through the Meta Model API and Muse Code, Muse Wonder Full as available to Muse Code subscribers, and both as relying on features Meta calls Active Memory and Active Focus. Meta has said Muse Wonder will soon power experiences in Muse, starting with subscribers.

Developers who run Meta's open models on their own hardware have a more concrete reference point. Meta's Muse Spark Open, a coding-focused mixture-of-experts model with 816B total parameters and 42B active per token, was trained on NVIDIA GB300 NVL72 systems, according to an NVIDIA developer post published on October 15. That post says Meta produced NVFP4 checkpoints optimized for Blackwell Ultra, that the model is served through SGLang, vLLM and TensorRT-LLM, and that NVIDIA Dynamo provides deployment recipes with disaggregated prefill and decode. NVIDIA reported up to 50,000 tokens per second per GPU and up to 600 tokens per second per user in a test on four Blackwell GPUs with speculative decoding at an acceptance length of 6, figures that apply to that configuration and not to production Muse traffic.

The security items are the part most relevant to teams building agents. NVIDIA has described Sentry as a hardware-isolated framework for monitoring autonomous agents, and the release places it alongside OpenShell as an area for further work with Meta. Because the wording is "consider", developers should not expect Muse-specific tooling from this announcement until either company publishes something more specific.

··········

WHAT IS ESTABLISHED AND WHAT REMAINS OPEN.

The established facts are limited but clear. NVIDIA has said it will invest $3 billion in Meta's equity offering, subject to completion and closing conditions; Meta intends to expand its use of Vera Rubin systems and Vera CPUs and will deploy Spectrum-X networking and BlueField-4; the two companies will optimize Muse inference and training together; and Meta's prospectus names NVIDIA, AMD and Broadcom as cornerstone investors with non-binding indications of up to $5.0 billion in aggregate for the Class A common stock.

Much is still undisclosed. Neither company has published the share class, price or settlement terms of NVIDIA's purchase beyond what the prospectus says about the cornerstone investors in general, nor the individual amounts for AMD and Broadcom. The infrastructure plans carry no volumes, dates or locations, and Meta's statements about Muse benchmark results and download growth come without figures. Until Meta's final offering documents and its next disclosures on capital spending are available, the agreement is best read as a commitment of capital and a statement of direction rather than a detailed deployment plan.

··········

·····

DATA STUDIOS

·····

datastudios.org