By NVIDIALearning Objectives
- Understand what Vera CPU is and why NVIDIA built a CPU after two decades of GPU-only AI hardware
- Identify the agentic workloads Vera is designed to accelerate
- Place Vera in the broader NVIDIA platform alongside Blackwell GPUs and Vera Rubin
What Is NVIDIA Vera CPU?
Vera is NVIDIA's first CPU purpose-built for agentic AI workloads — the first time the company has shipped a general-purpose CPU as a named product line rather than as an Arm-licensed peripheral on its Grace Hopper boards. NVIDIA introduced Vera at Computex 2026 in March and began shipping to partner labs in May 2026, with CEO Jensen Huang hand-delivering the first units to top customers.
The strategic framing is what makes Vera distinctive. NVIDIA positions Vera as the host CPU for autonomous-acting agents — software that behaves like users, opening browsers, navigating apps, calling tools, and orchestrating multi-step tasks. That workload pattern looks more like classic desktop computing than batch GPU inference, and CEO Jensen Huang has positioned Vera as a "brand new $200 billion total addressable market" that NVIDIA has never previously addressed.
💡Key Concept
Why an agentic CPU is different from a server CPU: Inference GPUs accelerate matrix multiplication for the model itself, but agents spend a large fraction of their wall-clock time doing things that look more like operating-system work — running headless browsers, parsing JSON tool responses, reading and writing files, calling APIs, managing memory across long-running sessions. Each of those is a classic CPU workload. Vera is designed to do all of that next to the GPU rather than on a generic server-class CPU bottleneck.
✅Tip
Visit NVIDIA Vera: blogs.nvidia.com/blog/vera-cpu-delivery — enterprise sales via NVIDIA's standard data-center channel
Pricing Tiers
- Standalone CPU SKU
- Channel partners + direct sales
- Volume discounts
- Vera CPU paired with Rubin GPUs
- Full-rack reference design
- Available to hyperscalers and frontier labs
- Drop-in CPU complement to GB200 deployments
- For agentic workloads layered on existing GPU clusters
NVIDIA does not publish list prices for data-center silicon — Vera SKUs are sold through enterprise channels with terms tied to volume, support level, and bundled rack-scale deployments. Standalone Vera sales hit $20 billion year-to-date as of NVIDIA's Q1 fiscal 2027 earnings.
Core Features
Purpose-Built for Agentic AI
Vera's defining feature is the workload it targets — autonomous agents that string together model inference, tool calls, browser actions, and external API calls into long-running sessions. The chip is architected to host the orchestrating loop next to the GPU rather than across a network or a generic x86 CPU bottleneck, lowering tail latency for the kind of work where the model spends as much time waiting on tool responses as it spends generating tokens.
Rack-Scale Reference Design — Vera Rubin
Vera ships in two configurations. Standalone Vera drops into hyperscaler and frontier-lab data centers as a complement to existing Blackwell GPU deployments. Vera Rubin is the rack-scale reference design that pairs Vera CPUs with the Rubin generation of GPUs (Blackwell's successor in NVIDIA's data-center roadmap) — a tightly-integrated CPU-plus-GPU rack optimized for agentic workloads end to end.
NVLink and CUDA Integration
Vera is integrated with NVIDIA's broader software stack: CUDA for GPU compute, NVLink and NVSwitch interconnects for high-bandwidth CPU-to-GPU memory access, and the NVIDIA AI Enterprise runtime for production deployment. The integration is the durable moat — Vera benefits from the same software ecosystem (CUDA libraries, container images, scheduling tooling) that has anchored NVIDIA's GPU dominance for the past decade.
Compute Density Positioning
Huang positioned Vera at Computex 2026 as part of a broader claim that NVIDIA's silicon roadmap will outpace the combined output of all alternative chip vendors through 2028. Whether or not that holds against AMD Instinct, Cerebras, Groq, and the Tesla/SpaceX/xAI Terafab joint venture, Vera materially expands the surface area NVIDIA is competing on — from GPU acceleration alone to CPU plus GPU plus interconnect plus software runtime.
Strengths
- Purpose-built for agents: First CPU architected specifically for orchestrating long-running agent sessions rather than as a generic server-class CPU
- NVIDIA software ecosystem: Inherits the CUDA, NVLink, and NVIDIA AI Enterprise stack that anchors most production AI deployments
- Rack-scale option (Vera Rubin): Reference design pairs Vera with NVIDIA's next-generation GPU line for end-to-end agentic workloads
- Massive standalone demand: $20 billion in standalone year-to-date sales, signaling that enterprise buyers see immediate use for the CPU as a complement to existing GPU footprints
- Drop-in compatibility: Designed to complement existing Blackwell-based deployments rather than requiring a forklift upgrade
Limitations & Considerations
- Enterprise-only distribution: Vera is not sold through retail or developer channels; access requires a data-center procurement relationship
- Architectural lock-in: Vera's value is highest when paired with NVIDIA GPUs and NVIDIA's software runtime — labs running on AMD Instinct or other accelerators get less of the integration benefit
- Early market signal: "Agentic AI workloads" as a category is still being defined; Vera's $200 billion total addressable market framing is forward-looking, not a current spend
- Concentration risk: Vera deepens NVIDIA's role as the single dominant supplier across CPU, GPU, networking, AND software — sharpening the broader question of how concentrated AI infrastructure has become around one vendor
NVIDIA's Position in the Agentic AI Stack
Vera arrives as NVIDIA's Q1 fiscal 2027 earnings posted record revenue of $81.6 billion — a 20% sequential jump — with data-center revenue at a record $75.2 billion and forward guidance of $91 billion next quarter. The earnings disclosure also surfaced $43 billion in non-marketable startup equity (nearly double the $22 billion balance from January), including a $30 billion commitment to OpenAI described by Huang as "quite significant" capacity expansion. Combined with the company's existing CUDA software moat, Vera positions NVIDIA to anchor agentic AI workloads end-to-end — hardware (CPU plus GPU), software (CUDA, AI Enterprise), and structural capital across the buyer side of the market.
Best Use Cases
| Workload | Why Vera |
|---|---|
| Autonomous browser agents | Host CPU for browser orchestration and tool-call loops alongside model inference |
| Multi-step tool calling | Low-latency CPU work for parsing tool responses and managing agent state |
| Long-running agent sessions | CPU-side memory management for sessions that persist across hours or days |
| Hybrid CPU plus GPU inference | Drop-in complement to existing Blackwell GPU clusters for agentic workloads |
| Vera Rubin rack-scale deployments | Full-stack CPU plus GPU rack design from NVIDIA |
When to choose alternatives:
- Pure GPU training workloads → NVIDIA H100/H200/Blackwell or AMD Instinct
- Inference-only workloads at low cost-per-token → Cerebras Inference or Groq Cloud
- Custom silicon for orbital or radiation-tolerant deployment → SpaceX Terafab products
Getting Started
- Vera is an enterprise data-center SKU — access requires an NVIDIA enterprise sales relationship
- Existing NVIDIA enterprise customers should ask their account team about Vera availability and the Vera Rubin rack reference design
- Evaluate agentic workloads first — the agent orchestration loop has to be where your latency bottleneck lives for Vera to deliver its strongest benefit
- Plan around software stack compatibility — Vera's strongest fit is when CUDA and NVIDIA AI Enterprise are already in production use
- Track NVIDIA's earnings disclosures for capacity availability — Vera sales have already absorbed $20 billion year-to-date, and supply may be constrained for new customers in 2026
Key Takeaways
- Vera is NVIDIA's first CPU purpose-built for agentic AI — the first CPU SKU in the company's data-center product line, designed to host the orchestrating loop for autonomous agents next to GPU inference
- The framing is a new $200 billion total addressable market — CPUs for agents that behave like users, distinct from the existing GPU training and inference markets
- Vera Rubin is the rack-scale option — pairing Vera with NVIDIA's next-generation Rubin GPUs for end-to-end agentic workloads
- Early traction is real — $20 billion in standalone year-to-date sales as of NVIDIA's Q1 fiscal 2027 earnings, alongside $81.6 billion in total revenue and $75.2 billion in data-center revenue
- The structural effect on the AI stack — Vera deepens NVIDIA's role across CPU, GPU, networking, AND software, sharpening the concentration question rather than resolving it
