# NVIDIA ($3B) and AMD ($1B) Back Meta’s $97.5B Offering to Power Muse AI Agent Infrastructure

Nvidia and AMD have committed a combined $4 billion into Meta's massive $97.5 billion equity offering—with Nvidia investing $3 billion and AMD investing $1 billion—cementing a multi-vendor silicon buildout designed specifically to run Meta's **Muse** personal and small-business AI agents at gigawatt scale. Rather than locking its AI roadmap to a single chipmaker, Meta is splitting the workload across Nvidia's Vera Rubin and BlueField-4 hardware and AMD's Helios rackscale systems and 6th-generation EPYC processors.

The scale of the financing reflects a fundamental shift in how consumer and enterprise AI infrastructure works. Answering a one-off chatbot prompt takes a fraction of a second of GPU time, but running an always-on personal AI agent like Meta's Muse—which lives inside its own dedicated virtual machine (VM) alongside a user's files, calendar, Slack, Zoom, and Dropbox tools—requires continuous CPU orchestration, low-latency memory, and hardware-enforced sandbox security for millions of simultaneous users.

On October 26, Meta expanded its planned equity offering to $97.5 billion, structured around $25 billion in common stock, $12.5 billion from 250 million convertible preferred depositary shares priced at $50 each, and up to $60 billion in non-voting Class C shares slated to begin trading in the first quarter of 2027. Within 24 hours, both Nvidia CEO Jensen Huang and AMD Chair and CEO Lisa Su announced expanded strategic hardware and software partnerships tied to the capital raise.

For enterprise architects, startup founders, and cloud buyers, Meta's dual-vendor blueprint reveals exactly how large-scale agentic workloads are being partitioned in 2026: high-density GPU racks (Nvidia Vera Rubin and custom AMD Instinct MI450/MI500 accelerators) handle frontier model training and heavy reasoning, while millions of high-core-count server CPUs (AMD 6th Gen EPYC and next-gen "Verano," alongside Nvidia 88-core Vera CPUs) host the isolated virtual machines where autonomous agents actually execute tasks.

## Fast Facts: Meta’s $97.5B Raise & The Nvidia + AMD Muse Compute Architecture

- **Capital Structure ($97.5B Total)**: Includes $25B in common stock, $12.5B in convertible preferred depositary shares ($50/share across 250M shares), and up to $60B in non-voting Class C shares (Q1 2027), backed by **$3B from Nvidia** and **$1B from AMD**, as reported by [Datacenter Dynamics](https://www.datacenterdynamics.com/en/news/nvidia-to-invest-3bn-in-metas-975bn-equity-offering-as-social-media-giant-plans-to-deploy-more-nvidia-hardware/) and the [AMD Newsroom](https://newsroom.amd.com/news/amd-meta-advance-strategic-partnership-power-metas-ai-infrastructure/).
- **Nvidia Hardware Footprint**: Expands Meta's deployment of **Vera Rubin** accelerators, **Vera CPUs**, **Spectrum-X Ethernet** networking, and **BlueField-4 DPUs**, plus collaboration on **OpenShell** and **Sentry** for hardware-level agent containment.
- **AMD Hardware Footprint**: First **AMD Helios** rackscale systems are now installed and running at Meta; **6th Gen AMD EPYC** (and upcoming **"Verano"**) CPUs power the dedicated VMs behind Meta's Muse agent, alongside custom **Instinct MI450**, **MI350X**, and future **MI500** GPUs.
- **Day-0 Open & Edge Model Support**: AMD delivered Day-0 ROCm and PyTorch support for Meta's **Muse Spark Open** models on Instinct GPUs and **Muse Glimmer** on Ryzen AI PCs and Radeon GPUs.
- **Consumer & SMB Muse Rollout**: Introduced in September 2026 via the pocket-sized **Charm** companion device for consumers and an SMB workspace tier integrating **Slack, Zoom, and Dropbox**.

## Silicon Workload Split: How Meta Divides Muse AI Agent Compute Between Nvidia and AMD

Instead of treating AI compute as a monolithic GPU cluster, Meta's engineering teams separate training, inference, VM hosting, and security monitoring across specialized silicon layers.

| Infrastructure Layer | Nvidia Stack Deployed at Meta | AMD Stack Deployed at Meta | Practical Role in Running Muse AI Agents |
|---|---|---|---|
| **Frontier Training & Heavy Reasoning** | **Vera Rubin** GPU architectures | **AMD Helios** rackscale systems + **Instinct MI350X / Custom MI450** | Trains frontier Muse models and serves deep multi-step reasoning turns |
| **Dedicated Agent VM Hosting (CPU)** | **Nvidia Vera CPU** (88 Olympus cores) | **6th Gen AMD EPYC** (millions deployed) & future **"Verano"** CPUs | Runs the isolated per-user Virtual Machines hosting each user's Muse agent and data |
| **Cluster Networking Fabric** | **Spectrum-X Ethernet** | **AMD Pensando** networking | Low-latency east-west traffic between agent sandboxes and inference pools |
| **Agent Security & Containment** | **BlueField-4 DPUs**, **OpenShell**, & **Sentry** | Hardware-isolated VM memory encryption on EPYC | Enforces sandbox isolation when Muse interacts with Slack, Zoom, or Dropbox |
| **Developer & Local Edge Ecosystem** | CUDA / TensorRT optimization | **ROCm / PyTorch / Triton** + Day-0 **Muse Glimmer** on Ryzen AI PCs | Lets developers run **Muse Spark Open** in the cloud and **Muse Glimmer** on laptops |

## Why Every Personal and SMB AI Agent Needs Its Own Virtual Machine

The most important technical detail in both the Nvidia and AMD announcements is how Meta architects **Muse**. Unlike a standard web chatbot that only exists while a browser tab is open, Muse is built as a proactive agent that works continuously toward a user's goals—whether through the consumer **Charm** hardware device or inside a small business's **Slack, Zoom, and Dropbox** workspace.

### 1. The Rise of CPU-Heavy Agent Sandboxes

To let an AI agent read files, organize Dropbox folders, draft Slack replies, and join Zoom workflows without leaking data between tenants, Meta runs each Muse environment inside a **dedicated Virtual Machine (VM)** that co-locates the agent runtime with the user's private data. That architectural choice explains why AMD highlighted that its **6th Gen EPYC processors** are the compute engine powering Muse's dedicated environments: while GPUs generate the tokens, CPUs run the millions of persistent Linux sandboxes, file systems, and tool connectors where agents spend 80% of their wall-clock time.

### 2. Multi-Vendor Leverage & Open-Weight Day-0 Readiness

By pairing Nvidia's Vera Rubin and OpenShell security stack with AMD's Helios racks, custom MI450 GPUs, and ROCm software, Meta prevents vendor lock-in while ensuring that its open-weight releases—**Muse Spark Open** for data centers and **Muse Glimmer** for local PCs—run natively on both CUDA and AMD ROCm/Ryzen AI hardware on launch day.

## Practical Takeaways for Enterprise Infrastructure & AI Teams

1. **Budget for CPU Sandbox Density, Not Just GPU FLOPS**: If you are deploying autonomous agents that run bash commands, browser sessions, or MCP connectors (like Slack and Dropbox), your infrastructure bottleneck will quickly shift from GPU token generation to isolated CPU/RAM sandbox capacity.
2. **Test Muse Spark Open and Muse Glimmer on Non-Nvidia Hardware**: Because AMD and Meta co-engineered Day-0 PyTorch, Triton, and ROCm support, teams can run **Muse Spark Open** on cost-effective AMD Instinct cloud instances and **Muse Glimmer** locally on Ryzen AI laptops.
3. **Isolate Agent State Per User or Per Tenant**: Follow Meta's VM-per-agent design pattern—never let a multi-tenant agent process share a single filesystem or credential cache across different customers or departments.