BusinessMarkets, money & strategy
Nvidia to Invest $3 Billion in Meta's Equity Offering as the Companies Expand Their Muse AI Partnership
Nvidia will buy $3 billion of stock in Meta's first $35 billion equity tranche, part of a plan to raise up to $95 billion for AI infrastructure, while Meta expands deployment of Nvidia's Vera Rubin systems for its Muse models.
From the WSJ · Technology article of October 27, 2026 and cited sources: Nvidia is committing $3 billion to Meta's stock sale. [1] [3] [4] [5]
SOURCES[1] Nvidia to Invest $3 Billion in Meta’s Equity Offering as… wsj.com[3] CNBC: Nvidia to invest $3 billion in Meta cnbc.com[4] NVIDIA Newsroom: NVIDIA and Meta to Expand Collaboration… nvidianews.nvidia.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- WSJ · Technology article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 27, 2026 article published
- Covered
- Oct 27 edition
- Evidence
- 12 sources across 11 sites
- 1Nvidia to Invest $3 Billion in Meta’s Equity Offering as…wsj.com
- 2The Guardian: Meta plans $95bn stock offeringtheguardian.com
- 3CNBC: Nvidia to invest $3 billion in Metacnbc.com
- 4NVIDIA Newsroom: NVIDIA and Meta to Expand Collaboration…nvidianews.nvidia.com
- 5SiliconANGLE: Meta to sell up to $95B in stocksiliconangle.com
- 6TechSpot: Meta plans $95 billion stock offeringtechspot.com
Reach: not measured yet
Nvidia is committing $3 billion to Meta's stock sale. Meta plans to raise up to $95 billion, starting with about $35 billion now: $22.5 billion of Class A shares sold through forward agreements and $12.5 billion of mandatory convertible preferred shares. A further $60 billion is planned for 2027. [1] [3] [4] [5] [9]
The wider partnership commits Meta to more Nvidia Vera Rubin systems, Vera CPUs, Spectrum-X networking and BlueField-4. The two companies will jointly tune training and inference cost for Muse, and they may use Nvidia's OpenShell and Sentry tools to deploy and monitor AI agents. [4]
Meta's largest chip supplier is directly funding part of a capital raise that will help pay for its chips. Implication: vendor financing in AI is deepening, and Meta could become more tied to Nvidia's hardware and agent tooling.
Sources are gathered and the dossier follows.
Executive brief
Nvidia, Meta's biggest chip supplier, will put $3 billion into Meta's new stock sale. Meta needs the money to buy, among other things, more Nvidia hardware (NVIDIA Newsroom). The stock sale is part of a plan to raise up to $95 billion. About $35 billion launches now and a planned $60 billion follows in 2027 (SiliconANGLE). Meta expects to spend $135–145 billion on capital projects in 2026 (TechSpot). Alongside the investment, the two companies are widening their work on Meta's Muse AI models and agents, which run on Nvidia's newest chips.
What changed and event timeline
Meta borrows heavily
Meta sold $30 billion of bonds and arranged $29 billion of financing for its Hyperion data center in Louisiana, led by Pimco and Blue Owl. Its stock sale a year later adds a different kind of funding ().
Muse Spark launches
This was the first model from Meta Superintelligence Labs. Unlike Meta's earlier Llama models, its weights were not released publicly ().
Meta announces the $95 billion plan
The first $35 billion is $22.5 billion of Class A stock sold through forward agreements plus $12.5 billion of mandatory convertible preferred shares. Morgan Stanley and Goldman Sachs are running the sale ().
Nvidia and Meta expand their partnership
Nvidia commits $3 billion to the offering. Meta plans to deploy more Nvidia Vera Rubin systems, Vera CPUs, Spectrum-X networking and BlueField-4 ().
Capabilities and access
- Muse Spark: handles text and images natively, reasons before answering, uses tools, and can coordinate several agents.
- It is available on meta.ai and in the Meta AI app, with a private API preview for selected users (The Decoder).
- Newer model: SiliconANGLE names a later model, "Muse Wonder Full" (SiliconANGLE). No version details or model card for it appear in the reviewed sources.
Read the full section
- Muse Spark: handles text and images natively, reasons before answering, uses tools, and can coordinate several agents. A "Contemplating" mode runs agents in parallel (The Decoder).
- Access: closed weights. It is available on meta.ai and in the Meta AI app, with a private API preview for selected users (The Decoder). It is also built into WhatsApp, Instagram, Facebook, Messenger and Ray-Ban glasses (DataCamp).
- Newer model: SiliconANGLE names a later model, "Muse Wonder Full" (SiliconANGLE). No version details or model card for it appear in the reviewed sources.
Technical analysis for researchers and developers
- Training: Meta rebuilt pretraining from scratch over nine months, changing the architecture, optimizer and data curation, then applied reinforcement learning.
- What the partnership covers: joint tuning of training and inference speed and cost on Vera Rubin.
- Reproducibility: with closed weights, outside researchers can only test Muse through the gated API.
Read the full section
- Training: Meta rebuilt pretraining from scratch over nine months, changing the architecture, optimizer and data curation, then applied reinforcement learning. Meta reports "thought compression," meaning shorter reasoning, when training penalized response time (The Decoder).
- What the partnership covers: joint tuning of training and inference speed and cost on Vera Rubin. The companies may also use Nvidia's OpenShell reference architecture to deploy agents safely and Nvidia Sentry to monitor agent behavior (NVIDIA Newsroom).
- Reproducibility: with closed weights, outside researchers can only test Muse through the gated API.
Claims and evidence
- Artificial Analysis Intelligence Index: an independent tester scored Muse Spark 52.
- Humanity's Last Exam: 50.2 with extended thinking. Meta reported this score itself; it is not independent (The Decoder).
- "Frontier-level" benchmarks and surging downloads: this is Nvidia's description in its press release, with no figures given (NVIDIA Newsroom).
Read the full section
- Artificial Analysis Intelligence Index: an independent tester scored Muse Spark 52. That puts it behind Gemini 3.1 Pro (57), GPT-5.4 (57) and Claude Opus 4.6 (53) (DataCamp).
- Humanity's Last Exam: 50.2 with extended thinking. Meta reported this score itself; it is not independent (The Decoder).
- "Frontier-level" benchmarks and surging downloads: this is Nvidia's description in its press release, with no figures given (NVIDIA Newsroom).
- Muse Wonder Full beating "GPT-6 Astra and Claude Opus 5.5": reported by SiliconANGLE. No independent evaluation was found to back it (SiliconANGLE).
Context and prior work
- Muse marks Meta's move away from open-weights Llama models toward closed frontier models.
- He joined Meta through a $14.3 billion investment deal (DataCamp).
- The equity raise comes on top of the bonds and joint ventures above.
Read the full section
- Muse marks Meta's move away from open-weights Llama models toward closed frontier models.
- Alexandr Wang leads Meta Superintelligence Labs. He joined Meta through a $14.3 billion investment deal (DataCamp).
- The equity raise comes on top of the bonds and joint ventures above. Meta has also disclosed $279 billion of future data-center lease commitments (Tech Funding News).
Limitations, safety and contested findings
- How much the chipmakers want to buy is unclear. TechSpot describes $5 billion combined (TechSpot). Either way, those indications were nonbinding (TipRanks).
- Safety: Muse Spark showed "evaluation awareness," meaning it flagged test scenarios as possible alignment traps.
- Market risk: Ken Griffin says AI is "almost certainly" a bubble.
Read the full section
- How much the chipmakers want to buy is unclear. Tech Funding News reports Nvidia, AMD and Broadcom each showed interest in up to $5 billion of stock (Tech Funding News). TechSpot describes $5 billion combined (TechSpot). Either way, those indications were nonbinding (TipRanks). Nvidia's own figure is $3 billion.
- Safety: Muse Spark showed "evaluation awareness," meaning it flagged test scenarios as possible alignment traps. It is also weaker at long agentic tasks and coding (The Decoder).
- Market risk: Ken Griffin says AI is "almost certainly" a bubble. Elon Musk predicts a wave of failures and mergers among smaller compute providers (TechSpot).
Business and practitioner implications
- Money flowing in a circle: a supplier is funding its customer's purchases.
- Dilution and control: the new nonvoting Class C shares let Meta raise money without weakening Zuckerberg's control.
- Market reaction: Meta rose 1.6% in premarket trading (TechSpot). Nvidia, AMD and Broadcom gained about 2–3% (TipRanks).
Read the full section
- Money flowing in a circle: a supplier is funding its customer's purchases. That echoes earlier vendor financing in AI and raises questions about how real the demand is (Tech Funding News).
- Dilution and control: the new nonvoting Class C shares let Meta raise money without weakening Zuckerberg's control. He holds about 60.6% of the votes with roughly 13.4% of the economic stake (Tech Funding News).
- Market reaction: Meta rose 1.6% in premarket trading (TechSpot). Nvidia, AMD and Broadcom gained about 2–3% (TipRanks).
- For developers: Muse access is still gated. Agent stacks built on OpenShell or Sentry would tie applications more closely to Nvidia.
The source trail.
Sources (12)
Nvidia to Invest $3 Billion in Meta’s Equity Offering as They Expand AI Partnership
Feed excerpt only.
wsj.com