NVIDIA-Meta Partnership

The NVIDIA-Meta partnership is the commercial and engineering relationship between NVIDIA and Meta Platforms (Meta AI), under which Meta has built several generations of its AI research and production infrastructure on NVIDIA GPUs, networking and software. The relationship runs from NVIDIA V100 clusters in 2017 and the A100-based Research SuperCluster announced in 2022, through the H100 clusters used to train Llama 3, to a multiyear, multigenerational agreement announced on February 17, 2026 covering millions of Blackwell and Rubin GPUs.^[8][10][17] On October 26, 2026 the companies said they would expand their work on Meta's Muse models and agents, and NVIDIA said it would invest $3 billion in an equity offering Meta had launched that day.^[1][2] Meta also buys accelerators from AMD and designs its own MTIA chips, so NVIDIA is one of several suppliers rather than its only one.^[18][19]

Timeline

Research SuperCluster and the A100 era

NVIDIA said in January 2022 that Meta had twice chosen its technology as the basis for its research infrastructure. The first generation, built in 2017, used 22,000 NVIDIA V100 Tensor Core GPUs and handled 35,000 AI training jobs a day.^[8]

The second was the AI Research SuperCluster (RSC), announced on January 24, 2022. In its first phase RSC used 760 NVIDIA DGX A100 systems with a total of 6,080 A100 GPUs, linked on an NVIDIA Quantum 200 Gb/s InfiniBand network, for 1,895 petaflops of TF32 performance. NVIDIA said that once fully deployed, RSC was expected to be the largest customer installation of DGX A100 systems; Meta planned a second phase of 16,000 GPUs delivering 5 exaflops of mixed-precision AI performance. Penguin Computing was NVIDIA's delivery partner for the system, and NVIDIA reported that Meta's early benchmarks showed RSC training large NLP models 3x faster and running computer vision jobs 20x faster than the previous system.^[8] Meta later said RSC, with its 16,000 A100 GPUs, played an important role in developing the first LLaMA models and Llama 2.^[10]

Hopper era: Grand Teton and the 24,576-GPU clusters

At the Open Compute Project Global Summit on October 18, 2022, Meta announced Grand Teton, a GPU-based hardware platform that succeeded its Zion-EX design. NVIDIA said Meta was sharing Grand Teton as an open design for system builders, and Meta later said it had contributed the platform to the Open Compute Project.^[9][10] NVIDIA said Grand Teton used H100 Tensor Core GPUs and reported that Alexis Bjorlin, Meta's vice president of infrastructure hardware, thanked NVIDIA at the summit for its collaboration on the platform's design.^[9]

In an Instagram Reels post on January 18, 2024, Zuckerberg said Meta's computing infrastructure would include 350,000 H100 GPUs from NVIDIA by the end of 2024, and "almost 600k H100 equivalents of compute if you include other GPUs."^[11] Meta repeated the target on March 12, 2024, when it described two clusters of 24,576 H100 GPUs each, built on Grand Teton, Open Rack and PyTorch. One cluster used a RoCE (RDMA over converged Ethernet) fabric based on Arista 7800 switches with Wedge400 and Minipack2 rack switches; the other used NVIDIA Quantum-2 InfiniBand. Both interconnected 400 Gbps endpoints, which let Meta compare the two interconnects for large-scale training. Meta said it reached the expected performance on the large clusters partly by changing its network routing strategy together with changes to the NVIDIA Collective Communications Library, and that it used this cluster design for Llama 3 training.^[10]

Llama models on NVIDIA platforms

NVIDIA has repeatedly packaged Meta's open-weight Llama models for its own software stack. On April 18, 2024 it announced optimizations for Llama 3 across its platforms, offered the model at ai.nvidia.com as an NVIDIA NIM microservice, and reported that a single H200 GPU generated about 3,000 tokens per second on the 70-billion-parameter version in an initial test.^[12]

For the Llama 3.1 release on July 23, 2024, NVIDIA launched its AI Foundry service for building custom Llama 3.1 models, along with Llama 3.1 NIM microservices. NVIDIA said Llama 3.1 was trained on more than 16,000 H100 GPUs, and that the two companies had produced a distillation recipe for building smaller Llama 3.1 models. Zuckerberg was quoted in NVIDIA's release: "The new Llama 3.1 models are a super-important step for open source AI."^[13] Meta's own account gives the same scale: Llama 3.1 405B was trained on more than 16,000 H100 GPUs, the first model in the Llama series trained at that scale.^[14]

NVIDIA later used Llama as a base for its own models. The Llama Nemotron reasoning family, announced on March 18, 2025, was "built on Llama models" and post-trained by NVIDIA.^[15]

Blackwell, Spectrum-X and the February 2026 agreement

At the October 2024 OCP Global Summit, Meta introduced Catalina, a high-power rack "based on the NVIDIA Blackwell platform full rack-scale solution" and built to support the GB200 Grace Blackwell Superchip.^[14] A year later, on October 13, 2025, NVIDIA announced that Meta would integrate NVIDIA Spectrum Ethernet switches into the Facebook Open Switching System (FBOSS), its software for managing network switches at scale. Gaya Nagarajan, Meta's vice president of networking engineering, said the integration was going into the Minipack3N switch and FBOSS.^[16]

On February 17, 2026, NVIDIA announced "a multiyear, multigenerational strategic partnership with Meta spanning on-premises, cloud and AI infrastructure."^[17] According to NVIDIA's release:

• Meta would build hyperscale data centers for both training and inference, with "large-scale deployment of NVIDIA CPUs and millions of NVIDIA Blackwell and Rubin GPUs."
• Meta would continue deploying Arm-based NVIDIA Grace CPUs for data center production applications, which NVIDIA called "the first large-scale NVIDIA Grace-only deployment."
• The companies were collaborating on NVIDIA Vera CPUs, "with the potential for large-scale deployment in 2027."
• Meta would deploy GB300-based systems in a unified architecture spanning its own data centers and NVIDIA Cloud Partner deployments, and had adopted the Spectrum-X Ethernet platform across its infrastructure footprint.
• Meta had adopted NVIDIA Confidential Computing for WhatsApp private processing, and the companies were working to extend it to other Meta products.^[17]

Zuckerberg said Meta would "build leading-edge clusters using their Vera Rubin platform to deliver personal superintelligence to everyone in the world." The release did not state a dollar value for the agreement.^[17]

October 2026 expansion: Muse and agent infrastructure

Meta launched Muse, its personal AI agent, in September 2026.^[5][7] In the offering materials it filed on October 26, 2026, Meta said it had onboarded more than 10 million Muse users in the United States and Canada within the early weeks of launch, and that it had also delivered Muse Wonder, its first frontier model.^[2]

The same day, NVIDIA and Meta announced they would "expand their work on Muse and the Muse family of models to improve inference and training performance."^[1] The press release, dated October 26, 2026, said:

• The companies would partner on optimizations across workflows to maximize performance and improve total cost of ownership.
• Meta and NVIDIA would "consider additional collaboration opportunities" on the NVIDIA OpenShell Reference Architecture, including NVIDIA OpenShell for safe deployment and NVIDIA Sentry for monitoring agent behavior. This was framed as consideration, not a deployment commitment.
• Meta intends to expand its deployment of NVIDIA infrastructure, including Vera Rubin and Vera CPUs, and will deploy Spectrum-X Ethernet and BlueField-4.^[1]

NVIDIA describes Sentry as a security framework, part of the OpenShell reference architecture announced at GTC Berlin, that runs on BlueField-4 data processing units and monitors AI agents and enforces access policies from a hardware-isolated trust domain, independent of host software.^[20] The press release states, without giving figures, that Muse models and agents "have delivered frontier-level benchmark performance and surging consumer downloads since launch."^[1]

Jensen Huang said: "NVIDIA and Meta are expanding our partnership to optimize the computing infrastructure end to end for capable, efficient and secure AI agents, so Muse can empower billions of people to achieve more and pursue greater ambitions." 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."^[1]

NVIDIA's investment in Meta's equity offering

The October 26, 2026 press release states: "In addition, NVIDIA will invest $3 billion in Meta's equity offering."^[1] Its forward-looking statements refer to "NVIDIA's proposed investment in Meta, including its anticipated completion and any applicable closing conditions." The release gives no share count, price or ownership percentage.^[1]

The offering

Meta announced the offering on October 26, 2026 as part of a financing plan to support its investment in AI infrastructure. As launched, the plan consisted of about $35 billion of underwritten public offerings (about $22.5 billion of Class A common stock and about $12.5 billion of depositary shares representing preferred stock), plus a planned at-the-market (ATM) program of about $60 billion of Class C common stock.^[2] Meta priced both offerings the same day and increased the common stock offering to $25.0 billion.^[3]

Class C common stock is to be established under Meta's amended and restated certificate of incorporation, filed in Delaware on October 21, 2026, and has no voting rights except as provided in that certificate or as required by law.^[2][4] Investing.com reported that the upsized underwritten offerings would raise up to $37.5 billion.^[6]

Cornerstone investors

Meta's launch announcement and its preliminary prospectus supplement for the common stock name NVIDIA, AMD and Broadcom as "Cornerstone Investors" that had, severally and not jointly, "indicated an interest in purchasing up to an aggregate of $5.0 billion of shares of Class A Common Stock" at the same price and on the same terms as other buyers. The filings say these indications are not binding agreements, that any of the three may buy more, fewer or no shares, and that underwriters earn the same discount on their shares as on others.^[2][4] The filings do not break the $5.0 billion down by investor.

NVIDIA's $3 billion figure comes from its own release.^[1] AMD announced on October 27, 2026 that it would invest $1 billion in Meta's equity offering, in a release that also said AMD Helios racks were installed and being tested at Meta and that AMD EPYC processors support the computing environments behind Muse.^[7] CNBC reported the two commitments together as $4 billion and said that on October 27 AMD shares were up 3%, NVIDIA was flat and Meta was down 1%.^[5]

Context

Meta's offering materials said it expected $135 billion to $145 billion of capital expenditures in 2026, had generated $97 billion of operating cash flow in the first three quarters of 2026, and held about $82 billion of cash, cash equivalents and marketable securities at the end of the third quarter. Meta described the equity offerings as a complement to operating cash flow, investment-grade debt and strategic partnerships.^[2] CNBC noted that NVIDIA has invested in OpenAI, Anthropic, Intel and xAI, and cited Dealroom data showing NVIDIA in 121 investment rounds in 2026 as of October 27, up from 13 in 2022.^[5] Other NVIDIA investments in its customers and suppliers are collected in the List of AI compute deals.

Other suppliers and in-house silicon

Meta describes a portfolio approach to sourcing AI silicon.^[19] On February 24, 2026, a week after the NVIDIA announcement, AMD and Meta announced a multi-year agreement to deploy up to 6 gigawatts of AMD Instinct GPUs, starting with a custom GPU based on the MI450 architecture in Helios racks, with shipments for the first gigawatt expected to begin in the second half of 2026. As part of that deal AMD issued Meta a performance-based warrant for up to 160 million AMD shares, vesting with GPU shipment milestones.^[18] The direction of the equity in the two arrangements differs: AMD granted Meta a warrant on AMD stock, while NVIDIA committed to buy Meta stock.^[18][1]

Meta also designs the Meta Training and Inference Accelerator. In March 2026 Meta said it deploys hundreds of thousands of MTIA chips for inference across organic content and ads, that the MTIA 300 was already in production for ranking and recommendations training, and that it was developing four new generations of MTIA chips within two years while "sourcing silicon from a range of industry leaders."^[19] Meta's cloud-capacity contracts with providers such as CoreWeave are listed in the same list of compute deals.

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13NVIDIA, "NVIDIA and Meta to Expand Collaboration on Muse and Muse Family of Models Amid Robust Growth," press release, October 26, 2026. nvidianews.nvidia.com/...models-amid-robust-growth
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Meta Platforms, "Meta Announces Launch of Offering of Depositary Shares Representing Series A Mandatory Convertible Preferred Stock and Offering of Class A Common Stock in Connection with Forward Sale Agreements," free writing prospectus (SEC Form FWP), October 26, 2026. sec.gov/...d116268dfwp
3. ^1 ^2 ^3 ^4 ^5 ^6Meta Platforms, Pricing term sheet for concurrent offerings of Class A common stock and Series A depositary shares, free writing prospectus (SEC Form FWP), dated October 26, 2026, filed October 27, 2026. sec.gov/...d382225dfwp
4. ^1 ^2 ^3Meta Platforms, Preliminary prospectus supplement, $22,500,000,000 Class A Common Stock (SEC Form 424B5), October 26, 2026. sec.gov/...d94886d424b5
5. ^1 ^2 ^3CNBC, "Nvidia, AMD to invest $4 billion in Meta as tech giants deepen AI partnership," October 27, 2026. cnbc.com/...nvidia-invest-meta
6. ^Investing.com (via Yahoo Finance), "Nvidia to commit $3 bln to Meta share offering as part of bigger AI partnership," October 27, 2026. finance.yahoo.com/...a-invest-3-bln-meta-064611851
7. ^1 ^2AMD, "AMD and Meta Advance Strategic Partnership to Power Meta's Superintelligence Infrastructure," press release, October 27, 2026. ir.amd.com/...tas-superintelligence-infrastructure
8. ^1 ^2 ^3 ^4 ^5 ^6Charlie Boyle, "Meta Works with NVIDIA to Build Massive AI Research Supercomputer," NVIDIA Blog, January 24, 2022. blogs.nvidia.com/...meta-ai-supercomputer-dgx
9. ^1 ^2 ^3Cliff Edwards, "Meta's Grand Teton Brings NVIDIA Hopper to Its Data Centers," NVIDIA Blog, October 18, 2022. blogs.nvidia.com/...meta-grand-teton
10. ^1 ^2 ^3 ^4 ^5Meta, "Building Meta's GenAI Infrastructure," Engineering at Meta, March 12, 2024. engineering.fb.com/...g-metas-genai-infrastructure
11. ^1 ^2CNBC, "Mark Zuckerberg indicates Meta is spending billions of dollars on Nvidia AI chips," January 18, 2024. cnbc.com/...s-spending-billions-on-nvidia-ai-chips
12. ^1 ^2Ankit Patel, "Wide Open: NVIDIA Accelerates Inference on Meta Llama 3," NVIDIA Blog, April 18, 2024. blogs.nvidia.com/...-llama3-inference-acceleration
13. ^1 ^2NVIDIA, "NVIDIA AI Foundry Builds Custom Llama 3.1 Generative AI Models for the World's Enterprises," press release, July 23, 2024. nvidianews.nvidia.com/...m-llama-generative-models
14. ^1 ^2 ^3Dan Rabinovitsj and Omar Baldonado, "Meta's open AI hardware vision," Engineering at Meta, October 15, 2024. engineering.fb.com/...metas-open-ai-hardware-vision
15. ^1 ^2NVIDIA, "NVIDIA Launches Family of Open Reasoning AI Models for Developers and Enterprises to Build Agentic AI Platforms," press release, March 18, 2025. nvidianews.nvidia.com/...uild-agentic-ai-platforms
16. ^1 ^2NVIDIA, "NVIDIA Spectrum-X Ethernet Switches Speed Up Networks for Meta and Oracle," press release, October 13, 2025. nvidianews.nvidia.com/...works-for-meta-and-oracle
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7NVIDIA, "Meta Builds AI Infrastructure With NVIDIA," press release, February 17, 2026. nvidianews.nvidia.com/...nfrastructure-with-nvidia
18. ^1 ^2 ^3AMD, "AMD, Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs," AMD Newsroom, February 24, 2026. newsroom.amd.com/...e-expanded-strategic-partnersh
19. ^1 ^2 ^3Meta, "Expanding Meta's Custom Silicon to Power Our AI Workloads," Meta Newsroom, March 11, 2026. about.fb.com/...-silicon-to-power-our-ai-workloads
20. ^NVIDIA, "Inside NVIDIA Sentry: Secure AI Agents at Runtime," NVIDIA Technical Blog, October 21, 2026. developer.nvidia.com/...ecure-ai-agents-at-runtime