CoreWeave Deploys NVIDIA Vera CPU for AI Agent Workloads, Rack-Scale Vera Rubin Systems Enter Availability
Specialized AI processors expand compute capacity for agentic AI, while power and infrastructure challenges drive long-term investment strategies.
The story
CoreWeave announced on September 30, 2026, the integration of NVIDIA's Vera CPU into its cloud platform, marking one of the first major commercial rollouts of a processor designed specifically for AI agents. The company detailed a rack-scale system that packs 128 Vera CPUs and 11,264 cores into a single cabinet, supporting over 11,000 concurrent agent environments.
This move highlights a shift in AI infrastructure focus beyond general-purpose GPUs to specialized silicon for tasks like code sandboxes and reinforcement learning. Additionally, CoreWeave confirmed that its NVIDIA Vera Rubin NVL72 cluster, a combined rack-scale system with 72 Rubin GPUs and 36 Vera CPUs, has entered limited availability, with Cognition named as its initial production customer. This deployment underscores the increasing demand for tailored compute solutions to power the next generation of AI applications.
Silicon
| Chip | Maker | What | For Whom |
|---|---|---|---|
| Vera CPU | NVIDIA | CPU designed for AI agents, deployed in rack-scale systems with 11,264 cores. | CoreWeave cloud platform for customers running agent-heavy, GPU-coupled workloads. |
| Iris | Meta | In-house AI chip entering production to reduce reliance on external GPUs. | Powering Meta's Facebook and Instagram AI platforms. |
| Ryzen AI Max+ APU | AMD (in Acemagic F9A) | Provides 131 TOPS of AI compute with Radeon 8065S iGPU. | Acemagic F9A workstation mini PCs for AI workloads. |
| RTX Spark N1X | NVIDIA (in ASUS ProArt 16) | Arm SoC for creator laptops, featuring 6,144 CUDA cores. | ASUS ProArt 16 creator laptops, with an October 8, 2026 launch in China. |
The build-out
| Project | Who | Scale | Where |
|---|---|---|---|
| AI infrastructure capacity expansion | CoreWeave, NVIDIA | Over 5 gigawatts of shared AI infrastructure capacity planned. | CoreWeave's cloud platform. |
| AI computing infrastructure expansion | Meta | Plans to double capacity from 7 GW in 2026 to 14 GW in 2027. | Global data centers. |
| AI chip rental agreements | Meta, Firmus | Adds to Meta's computing capacity. | Firmus data centers in Southeast Asia and Melbourne, Australia. |
| Colossus / Colossus II data centers | SpaceX, Anthropic | Projected to rise from 2.1 gigawatts in 2026 to 6 gigawatts by end of 2027. | Terrestrial data centers. |
| Next-generation AI factory infrastructure | Compal Electronics | Showcasing advanced power and cooling for megawatt-scale compute density. | OCP Global Summit 2026, San Jose. |
Supply & policy signals
TSMC's leading-edge wafer capacity is sold out through the first half of 2028.
Implication: Indicates sustained high demand for advanced AI chips and potential constraints for new orders.
TSMC capital expenditure is expected to grow approximately 15% annually through 2029.
Implication: Reflects strong AI demand visibility, but tool suppliers are struggling to keep pace with increased capital expenditure.
Memory availability (HBM, DRAM, flash) remains the biggest system-level risk.
Implication: Shortages are forcing trade-offs in next-generation system design, highlighting a critical bottleneck beyond compute.
AI data center leases are stretching to 15- to 20-year take-or-pay contracts.
Implication: Power grid access has become the primary constraint, driving long-term commitments for sites with existing electrical capacity.
Anthropic committed $518 billion in non-cancelable AI infrastructure investment over the next decade.
Implication: Signals long-term demand across the AI value chain, including semiconductors and memory, despite market volatility.
What we'll be watching
- ASUS ProArt 16 laptop, powered by NVIDIA RTX Spark N1X, is slated for launch on October 8, 2026, in China.
- SoftBank's AI Data Center GPU Cloud service will begin selling access to its GB200 NVL72 infrastructure in October 2026.
- Billings for SpaceX's new $13.3 billion/year AI compute hosting agreement are set to commence on December 1, 2026.
- Meta aims to double its overall computing power to 14 gigawatts by 2027, following the September production start of its Iris AI chip.
- TSMC's 2nm production is targeted to reach 200,000 wafers per month in 2027, as its advanced nodes continue to ramp up.
- Monitoring how memory shortages continue to impact next-generation system design and pricing across the AI hardware ecosystem.
Reporting + analyst voices: grounded via Google Search at publish time.