On May 19, 2026, NVIDIA executive Ian Buck carried a server chassis into Anthropic’s San Francisco office and plugged it in himself. The next day, he did the same thing at OpenAI’s Mission Bay headquarters, then at SpaceXAI in Palo Alto, then at Oracle’s Santa Clara campus. The chip inside wasn’t a GPU. It was something NVIDIA had never shipped before: a CPU called Vera, purpose-built for a category of computing that didn’t have its own hardware until now.
The NVIDIA Vera CPU is the world’s first processor designed specifically for agentic AI, the type of computing where AI models don’t just generate answers but execute tasks, call tools, run code, and manage complex workflows autonomously. NVIDIA reported $81.6 billion in Q1 revenue the same week, up 85% year over year, and CEO Jensen Huang told investors that Vera opens a “brand new $200 billion market” the company has never addressed.
That sentence should matter to every founder building with AI right now. When the most valuable semiconductor company on Earth builds an entirely new chip category and books $20 billion in sales before the product even ships broadly, it’s confirming something the startup world has been debating for two years: AI agents aren’t a feature. They’re an infrastructure layer. And that layer just got its own silicon.
Last updated: May 2026
Quick answers
What is the NVIDIA Vera CPU?
The NVIDIA Vera CPU is a data center processor featuring 88 custom Olympus cores, built specifically for agentic AI workloads. Unlike GPUs that handle model training and inference, Vera handles the execution layer: tool calls, code sandboxes, task orchestration, and data retrieval that AI agents perform after generating a response. It delivers 1.2 TB/s memory bandwidth and 50% faster per-core performance than standard server CPUs.
Why did NVIDIA build a CPU?
NVIDIA built Vera because AI agents require a different type of computing than model training. GPUs power the “thinking” phase, but agents spend most of their time executing tasks on CPUs: running code, querying databases, calling APIs, and managing workflows. Jensen Huang said this represents a $200 billion market NVIDIA has never addressed, separate from its existing GPU business.
What is the NVIDIA Vera CPU?
Vera is NVIDIA’s first custom-designed CPU, built from scratch on the Arm v9.2-A architecture. It’s not an upgrade to NVIDIA’s previous Grace CPU, which used licensed Arm Neoverse V2 cores. For Vera, NVIDIA designed its own core called Olympus. That distinction matters.
The spec sheet tells part of the story: 88 Olympus cores, 1.2 TB/s of LPDDR5X memory bandwidth (double what Grace delivered), up to 1.5 TB of memory capacity (triple Grace), and NVLink-C2C connectivity running at 1.8 TB/s for coherent CPU-GPU communication. Vera also introduces Spatial Multithreading, which physically partitions each core’s resources to run 176 total threads without the performance penalties of traditional time-slicing.
But the architecture isn’t the point for most founders. The point is what Vera runs. When an AI agent at Anthropic generates a plan, breaks it into steps, executes Python code in a sandbox, retrieves documents from a vector database, calls three external APIs, and synthesizes the results, almost none of that runs on GPUs. It runs on CPUs. And until Vera, those CPUs were general-purpose chips from Intel and AMD that weren’t designed for this workload pattern.
Vera also introduces full Confidential Computing support, a notable advance over Grace. This means CPU and GPU workloads run in a fully protected domain, which matters for enterprise customers running sensitive data through AI agents. It’s the kind of feature that turns a “we’re testing agents” conversation into a “we’re deploying agents in production” conversation.
NVIDIA claims agent sandboxes run 50% faster on Vera than on traditional server CPUs and that enterprise data queries execute up to 3x faster. Oracle Cloud Infrastructure, the first hyperscaler to deploy Vera at scale, plans to roll out hundreds of thousands of units beginning in 2026.
Why does agentic AI need its own hardware?
Think about what happens when you ask Claude or ChatGPT to do something complex: research a topic, write code, test it, fix the errors, then write a summary. The model generates each step using GPU inference. But between those generation steps, the system is executing code, reading files, parsing API responses, managing state across tool calls, and orchestrating the whole pipeline. That’s CPU work, and there’s a lot of it.
Deloitte reported that 93% of IT leaders plan to deploy autonomous agents within two years. Gartner predicts 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% a year earlier. IDC estimates agentic AI already represents 10-15% of enterprise IT spending this year. The AI side hustle economy is growing alongside the enterprise market.
The math problem is straightforward. If a single AI agent call involves one GPU inference step and ten CPU execution steps, then scaling to millions of concurrent agents means you need an enormous amount of CPU capacity tuned for this specific pattern: rapid context switching, high memory bandwidth for retrieval, and low-latency tool execution. General-purpose server CPUs handle this, but not efficiently. That’s the gap Vera fills.
Jensen Huang framed it on NVIDIA’s Q1 earnings call: the buildout of AI factories is “the largest infrastructure expansion in human history,” and it’s accelerating. NVIDIA reported record Data Center revenue of $75.2 billion for the quarter, up 92% year over year, per the company’s SEC filing.
How does the Vera CPU compare to NVIDIA’s GPU?
They’re not competing products. They’re complementary layers of the same AI stack, and understanding the distinction is useful for any founder making infrastructure decisions.
GPUs handle parallel processing at massive scale. They’re what powers model training (teaching an AI to understand language) and inference (generating responses). NVIDIA’s Blackwell GPUs dominate this market and will continue to.
Vera handles what comes after inference: the serial, task-oriented work that agents perform. Code execution, database queries, API calls, file operations, workflow orchestration. This work is inherently sequential and benefits from fast single-threaded performance and high memory bandwidth, not the thousands of parallel cores that make GPUs powerful for matrix math.
| Feature | NVIDIA Vera CPU | NVIDIA Blackwell GPU | Traditional server CPU (Intel/AMD) |
|---|---|---|---|
| Primary workload | Agent execution, tool calls, orchestration | Model training, inference | General-purpose computing |
| Core design | 88 custom Olympus cores (Arm v9.2-A) | Thousands of CUDA/Tensor cores | 64-128 x86 cores |
| Memory bandwidth | 1.2 TB/s (LPDDR5X) | 8+ TB/s (HBM3e) | ~500-600 GB/s (DDR5) |
| Built for | Agent sandboxes, retrieval, code execution | Matrix math, parallel processing | Broad compatibility |
| Best for founders building | Agent-heavy products at scale | Custom models, fine-tuning | Traditional web apps |
The Vera Rubin NVL72, NVIDIA’s full-rack system that pairs Vera CPUs with Rubin GPUs, costs an estimated $7.8 million per unit according to Morgan Stanley Research, with individual Vera CPUs priced around $5,000 each in volume. That’s data center pricing, not startup pricing. But the relevant signal for founders isn’t the hardware cost. It’s what the hardware tells you about where the industry is heading.
What does the Vera CPU mean for founders building with AI?
The most important thing Vera confirms is that the AI stack is splitting into two distinct layers, and founders should plan for both.
Layer 1: Intelligence (GPUs). This is model training and inference. If you’re fine-tuning a model, running custom inference, or building products that need real-time AI generation, you’re paying for GPU compute. Most founders building on top of OpenAI, Anthropic, or Google APIs are already paying for this indirectly through API pricing.
Layer 2: Execution (CPUs). This is where agents act. Every time your AI product runs code, queries a database, calls an API, processes a file, or manages a multi-step workflow, that’s CPU work. As products move from simple chatbots to autonomous agents that complete tasks end-to-end, this layer grows faster than the inference layer.
For founders using cloud APIs, the split is already happening in pricing. Starting an AI business in 2026 means understanding that your costs will increasingly come from execution compute, not just model calls. When Anthropic’s Claude executes a 20-step agentic workflow, the Claude API charges for the inference tokens, but the compute cost of running those 20 tool calls, sandboxes, and retrievals falls on the infrastructure provider’s CPU fleet. Vera is NVIDIA’s bet that this CPU fleet will be enormous.
The practical implications for founders watching Jensen Huang’s moves:
If you’re building agent-heavy products, your architecture should separate inference calls from execution compute. The two will scale differently and cost differently. Products that treat every agent step as an API call to a model are overspending. The execution steps (code running, data retrieval, orchestration) should run on purpose-built CPU infrastructure, which is exactly what Vera provides at scale.
If you’re choosing between cloud providers, Oracle’s early Vera deployment gives it a specific advantage for agentic workloads. AWS, Azure, and GCP will follow, but Oracle’s hundreds-of-thousands-unit commitment to Vera puts it ahead on this specific infrastructure layer for now.
If you’re evaluating the “build vs. buy” decision for AI agents, Vera’s existence confirms that the major cloud providers expect agent execution to be a first-class compute category. You won’t need to provision and tune your own CPU clusters for agent workloads. Purpose-built infrastructure is coming to the cloud platforms you already use.
Who is already using the NVIDIA Vera CPU?
The first four recipients tell you who NVIDIA thinks the Vera market belongs to.
Anthropic received the first delivery in San Francisco. The company’s Claude models power some of the most complex agentic workflows in production, including Claude Code (an autonomous coding agent) and the Computer Use feature that lets Claude operate software directly. These agent products generate huge volumes of CPU-side compute: code execution, file operations, browser automation, and multi-step task management. Anthropic is a natural early adopter because its products are already CPU-constrained at scale.
OpenAI took delivery at its Mission Bay office. The company’s agent ambitions are growing, and its recent deals with YC startups signal a push to become infrastructure for the next generation of agent-powered companies.
SpaceXAI received Vera at its Palo Alto campus, connecting the chip to Elon Musk’s broader AI ambitions. After the SpaceX-Cursor deal earlier this year, the company has been building out AI coding infrastructure at a pace that requires dedicated agent compute. SpaceXAI’s use case is telling: autonomous coding agents that can write, test, debug, and deploy software generate massive CPU workloads with every cycle. A single coding session can involve hundreds of sandboxed code executions.
Oracle Cloud Infrastructure is deploying at a different scale entirely. OCI’s commitment to hundreds of thousands of Vera CPUs makes it the first hyperscaler to offer purpose-built agentic infrastructure as a cloud service. For founders, this is the most relevant signal: Vera won’t stay in the lab. It’s coming to the cloud platforms startups use.
Is the agentic AI market real or overhyped?
The hype-vs-reality debate around AI agents has been running since late 2024. Vera is NVIDIA’s $20 billion answer.
Here’s where the numbers stand. The agentic AI market is projected at $40 billion in 2026, according to Information Matters, with estimates ranging up to $140 billion by 2030. IDC projects that 40% of roles in Global 2000 companies will involve direct engagement with AI agents by the end of this year. But there’s a credibility gap: a 2026 survey found that 79% of enterprises say they’ve adopted AI agents, while only 11% run them in production.
NVIDIA’s Vera investment cuts through the survey noise. Companies don’t book $20 billion in CPU orders for products that aren’t shipping real workloads. Oracle doesn’t commit to hundreds of thousands of specialized chips for pilot projects. The hardware orders are a harder signal than the adoption surveys.
The trajectory NVIDIA’s CEO has been describing is playing out: AI isn’t just about bigger models. It’s about models that do things. And doing things requires a different kind of computing than thinking about things. Vera is the hardware acknowledgment of that distinction.
For founders still deciding whether to build agent-powered products, the question has shifted. It’s no longer “will agents work?” The infrastructure investment has answered that. The question now is “how many agents will my product run, and what infrastructure do they need?”
The bigger lesson behind the Vera CPU
Hardware follows software waves. This pattern repeats throughout tech history, and recognizing it early separates founders who ride the wave from those who get dragged by it. GPUs existed for gaming before deep learning needed them. NVIDIA’s market cap was $15 billion in 2016; ten years later it passed $3 trillion because Jensen Huang spotted the deep learning wave before anyone else built hardware for it. Smartphones existed as communication devices before the app economy needed them. Google’s TPUs didn’t exist until the company needed custom chips for its AI search workloads.
Vera follows the same pattern. AI agents existed as software products before anyone built hardware specifically for running them. Now the hardware exists, which means the infrastructure providers, cloud platforms, and the entire compute supply chain will gear toward agent workloads over the next 2-3 years. That’s the kind of commitment signal that’s hard to walk back.
Analysts at Tom’s Hardware estimate NVIDIA will ship 4 million Vera CPUs in FY2027, potentially capturing two-thirds of the data center CPU market from Intel and AMD. Even if those projections prove optimistic, the direction is unmistakable. NVIDIA’s pivot from graphics to AI created a $3 trillion company. Its pivot from GPUs-only to GPUs-plus-CPUs is the next chapter.
Amazon is already making parallel moves. Meta recently tapped AWS Graviton processors to run AI agents, signaling that even companies not using NVIDIA GPUs are investing in CPU-side agent infrastructure. The agentic compute market isn’t a single-vendor story. But NVIDIA’s $20 billion in booked Vera sales and 4-million-unit delivery target make it the clearest signal of where the industry is heading.
For solopreneurs building micro-SaaS products and enterprise teams alike, the takeaway is the same. The companies providing your AI infrastructure are betting billions that the next phase of AI is about execution, not just generation. Your product roadmap should reflect that bet. The founders who thrive in this next phase won’t be the ones with the biggest models. They’ll be the ones who figured out what to do with agents once the thinking is done.
