Nvidia and Meta announced on October 27, 2026 that they will deepen their partnership on Muse and the Muse family of models, and that Nvidia will invest $3 billion in Meta's equity offering. The details come from Nvidia's press release on GlobeNewswire, and the release is short on numbers but specific about what is committed and what is only being considered.
This post breaks that line between the two, adds the history of how the companies got here, and explains what it does and does not change for people who build on Muse.
TL;DR: what is confirmed
| Item | Status per the release |
|---|---|
| Expanded work on Muse and the Muse family to improve inference and training performance | Announced |
| Joint optimization across workflows to improve total cost of ownership | Stated intent |
| Meta expands deployment of Vera Rubin and Vera CPUs | Stated intent ("Meta intends to") |
| Meta deploys Spectrum-X Ethernet and BlueField-4 | Stated plan |
| Nvidia invests $3 billion in Meta's equity offering | Announced; completion and closing conditions referenced in forward-looking statements |
| Collaboration on OpenShell and Sentry | Only "will consider" |
| Benchmark figures for Muse | None published |
What did the release actually say?
Nvidia says the two companies "will partner to expand their work on Muse and the Muse family of models to improve inference and training performance, reflecting popular demand for useful AI agents." It adds that Meta's Muse models and agents, which use Nvidia infrastructure, "have delivered frontier-level benchmark performance and surging consumer downloads since launch." No benchmark, download count or model name accompanies that sentence, so it is a company claim we cannot check from this source.
On infrastructure, 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 for network efficiency and throughput. Then one line: "In addition, NVIDIA will invest $3 billion in Meta's equity offering." The release does not give a share price, a stake size or a closing date.
Jensen Huang framed it around agents: "AI agents represent a new computing paradigm," and said the companies will "optimize the computing infrastructure end to end for capable, efficient and secure AI agents." Mark Zuckerberg said "Muse is how we bring personal superintelligence to everyone."
OpenShell and Sentry: under consideration only
The release says Meta and Nvidia "will also consider additional collaboration opportunities" on the NVIDIA OpenShell Reference Architecture, "including NVIDIA OpenShell for safe deployment and NVIDIA Sentry for monitoring agent behavior." The news summary at the top also ties Sentry to structured, unstructured and graph data processing.
The key word is consider. Nothing here says Meta will ship agents inside OpenShell or monitor them with Sentry. It is still a notable signal: both companies are pairing agent growth with deployment safety and monitoring, which matches where the industry is heading. Our look at agent security guardrails covers why runtime monitoring matters once agents take real actions.
How did Nvidia and Meta get here?
This is an extension of a relationship that has been building all year.
- February 2026. Meta announced a multi-year infrastructure partnership with Nvidia, including Nvidia Confidential Computing for WhatsApp private messaging and Spectrum-X Ethernet across its infrastructure, as summarized by Unite.AI.
- October 15, 2026. Nvidia's developer blog published Scale Agentic Coding Workflows with Meta's Muse Spark Open on NVIDIA GB300 NVL72. It describes an 816B-parameter mixture-of-experts model with 42B active parameters and a 1M-token context, NVFP4 checkpoints for Blackwell Ultra, and a Pareto curve claiming over 50K tokens per second per GPU, with the caveat that the curve assumes an acceptance length of 6 and the post gives limited methodology. We covered the model in our Muse Spark Open guide.
- Mid-October. Nvidia's draft-model work for the same model is in our post on DFlash 2 and speculative decoding for Muse Spark Open.
- October 27, 2026. The expanded Muse collaboration and the proposed $3B investment.
Meta's own capacity plans are the other half of the story. We explained the scale of its build-out in the Prometheus supercluster explainer, and the rollout of the Vera Rubin generation across the industry in Nvidia's Vera Rubin NVL72 claims.
Why would Nvidia invest in a customer?
Nvidia selling chips to Meta and then putting $3 billion into Meta's equity offering looks circular, and some readers will say so. The release does not explain the rationale, so what follows is our reading, not a company statement.
For Nvidia, an equity stake ties a large buyer closer at a moment when hyperscalers are also building custom silicon. Co-optimizing Muse for Vera Rubin makes Nvidia hardware the path of least resistance for one of the largest open-weight and consumer-facing model families. For Meta, the investment brings capital and a deeper engineering relationship, and the stated goal of better total cost of ownership matters for a company serving billions of users.
The scale also needs perspective. The release gives no figure for the full offering, so the $3 billion cannot be compared to Meta's total capital spending or to Nvidia's revenue from Meta from this source alone. Be careful with commentary that fills in those numbers without citing a filing.
What does it change for developers?
Today, very little. Muse Spark Open already has Nvidia-optimized checkpoints and works with TensorRT-LLM, SGLang and vLLM, according to Nvidia's October 15 post. The announcement promises more joint optimization across training and inference, which could mean faster kernels, better quantization and cheaper serving on Nvidia hardware over time. It does not announce a new model, a new price, an API change or a benchmark.
If you run or plan to run Muse Spark Open, the sensible move is to watch the model's deployment docs and Nvidia's recipes for updates rather than change anything now. If you are choosing hardware, note that the stated direction is toward Vera Rubin and Vera CPUs, not toward other accelerators. Our post on Apple and Meta bringing Muse Spark Open to Mac Studio shows the other side of the picture: Meta wants Muse running on several platforms, and this release does not make Nvidia exclusive.
What are Vera CPUs, Spectrum-X and BlueField-4 doing in this deal?
The release names three pieces beyond the GPUs, and each addresses a different bottleneck in running agents at scale.
Vera CPUs. Agents spend a lot of time outside the model: calling tools, running code, parsing results and juggling state. That work runs on CPUs, and the release explicitly says Vera CPUs are for "next-generation AI agents." The point is that agentic workloads are not only GPU-bound, so CPU capacity and the link between CPU and GPU start to matter. Our coverage of AMD's Helion CPU rack for agentic AI shows the same shift on the other side of the market.
Spectrum-X Ethernet. Training and serving a large mixture-of-experts model moves a huge amount of data between GPUs. Better networking improves utilization, which is where total cost of ownership is won or lost. The release describes the goal as "network efficiency and throughput" and gives no measured gain.
BlueField-4. This data-processing unit offloads networking, storage and security tasks from the main processors. In an agent setting it is also a natural place to enforce isolation, which loosely connects to the OpenShell talks, though the release does not say they are linked.
None of the three comes with a performance number, a unit count or a delivery date, which is why "intends to expand" is the honest summary. Treat any headline that quotes a gigawatt figure or a GPU count for this announcement as coming from somewhere other than Nvidia's release.
Claimed versus verified
- "Frontier-level benchmark performance." Meta and Nvidia's claim. No scores in the release. For measured comparisons, see our Muse Wonder comparison.
- "Surging consumer downloads." No figures in the release. Our earlier look at Muse's ad spend and retention numbers shows how third-party estimates differ from company framing.
- "Maximum performance per watt." A stated goal, with no measured result.
- The $3B investment. Announced by Nvidia. The release's forward-looking statements refer to anticipated completion and closing conditions, so wait for filings to confirm terms and timing.
What to watch next
- A Meta or Nvidia regulatory filing that gives the terms of the equity offering and the closing date.
- Any benchmark or serving-cost data from the joint Muse optimization work.
- Whether Meta ships agents on OpenShell or adopts Sentry.
- Timing of Vera Rubin and Vera CPU deployment at Meta, which the release states only as intent.
- How rivals respond, since a tighter Nvidia and Meta tie shapes where open-weight models are tuned first.
Related reading
- Meta Muse Spark Open: 816B open-weight agentic coding model
- Nvidia DFlash 2 draft model for Muse Spark Open
- Meta Prometheus supercluster: what is confirmed
- Nvidia Vera Rubin NVL72: claimed versus verified
- Apple and Meta: Muse Spark Open on Mac Studio
- Meta Muse ad budget and retention claims
- Official: Nvidia's press release
Details are accurate as of October 27, 2026 and rest on Nvidia's press release; the investment and deployment plans may change before closing.