NVIDIA’s push beyond graphics processors has reached an important milestone: the NVIDIA Vera CPU, its first custom processor designed around agentic artificial intelligence, is now shipping to cloud providers and major AI laboratories. The company said on 27 August 2026 that AWS had received its first Vera CPU server and Vera Rubin GPU, following earlier deliveries to Oracle Cloud Infrastructure, Anthropic, OpenAI and SpaceXAI.
This matters because modern AI agents do much more than generate a reply. They call tools, execute code, search files, manage context and coordinate multi-step workflows. Those supporting tasks can become a CPU bottleneck even when powerful GPUs are available. Vera is NVIDIA’s attempt to redesign that side of the system for the age of AI agents.
Background: why AI agents need more than GPUs
GPUs remain central to model training and inference, but an agentic system is a larger software pipeline. A model may decide to run Python, query a database, launch an isolated software environment, retrieve documents and evaluate the result before producing an answer. CPUs handle much of this orchestration, data preparation and general-purpose execution.
Traditional server processors were not created specifically for thousands of concurrent agent sandboxes or repeated tool calls. If the CPU cannot prepare work and move data quickly enough, expensive accelerators may sit underused. That can increase response times and raise the cost of every useful AI task.
What NVIDIA announced
NVIDIA says Vera systems are shipping at scale, moving the processor from product announcement towards production deployment. AWS is the latest named recipient and plans to work with NVIDIA on Vera-based infrastructure. Oracle Cloud Infrastructure has also said it plans to deploy hundreds of thousands of Vera CPUs beginning in 2026.
The rollout is not limited to cloud operators. Anthropic, OpenAI and SpaceXAI have received systems for evaluation or development, giving some of the largest AI builders early access to the platform. That does not mean every cloud customer can order a Vera instance immediately; broad service availability, pricing and regional access will depend on each provider’s rollout.
NVIDIA Vera CPU specifications
According to NVIDIA’s product documentation, Vera includes:
- 88 custom Olympus CPU cores and 176 threads through NVIDIA Spatial Multithreading.
- Up to 1.2TB/s of LPDDR5X memory bandwidth and up to 1.5TB of memory.
- Up to 1.8TB/s of coherent NVLink-C2C bandwidth between Vera CPUs and compatible NVIDIA GPUs.
- A single-die design with NVIDIA’s Scalable Coherency Fabric to reduce cross-chiplet communication delays.
- Arm compatibility, confidential-computing features and support for both standalone CPU systems and accelerated platforms.
Vera also acts as the host processor in the Vera Rubin NVL72 rack-scale system, which combines 36 Vera CPUs with 72 Rubin GPUs, networking and data-processing hardware.
Why the NVIDIA Vera CPU matters
Vera reflects a broader change in AI infrastructure design. Performance is increasingly determined by the complete system rather than the GPU alone. CPU execution, memory bandwidth, networking, storage and software isolation all affect how quickly an agent completes a real task.
NVIDIA claims Vera delivers up to 1.8 times faster performance than leading x86 CPUs across selected agentic sandbox workloads. Its product page also claims twice the memory bandwidth and three times the bandwidth per core compared with leading x86 systems using DDR5. These are vendor-provided comparisons, and NVIDIA notes that the measured results are subject to change. Independent benchmarks will be needed to show how the processor performs across a wider range of production software.
Practical impact for businesses and developers
Faster tool-using AI services
Developers building coding assistants, research agents or automated operations tools could see lower delays when agents execute code and manage many isolated environments. Better CPU throughput may also allow more simultaneous users on the same infrastructure.
Potentially better infrastructure economics
Keeping GPUs supplied with data and instructions can improve overall utilisation. For cloud providers and large enterprises, that may translate into more completed tasks per rack and better energy efficiency. The ultimate customer benefit will depend on cloud pricing, software optimisation and real workload behaviour.
A stronger full-stack NVIDIA platform
Vera deepens NVIDIA’s control over the AI data-centre stack. The company can now combine its own CPU, GPU, interconnect, networking, DPU and rack architecture. That tight integration may simplify optimisation for customers already committed to NVIDIA systems, while increasing the importance of portability planning.
Risks, limitations and concerns
The biggest unanswered questions are price, availability and independent performance. Initial deliveries to hyperscalers and frontier AI laboratories do not guarantee rapid access for smaller organisations. Arm compatibility also does not eliminate migration work: enterprises must validate operating systems, libraries, monitoring tools and proprietary software on the new platform.
There is also a concentration risk. A tightly integrated system can offer strong performance, but customers may become more dependent on one vendor’s hardware, networking and software roadmap. Procurement teams should compare the operational gains with lock-in, supply-chain and long-term cost considerations.
Finally, more efficient agent infrastructure does not resolve AI governance problems. Organisations still need controls for data access, tool permissions, audit logs, cybersecurity and human oversight. Faster autonomous execution can magnify mistakes if those safeguards are weak.
What to watch next
The next useful signals will be general cloud availability, instance pricing and benchmark results from independent users. Watch for details from AWS and OCI about which regions and services will expose Vera, as well as software support from Linux distributors, observability vendors and enterprise application providers.
It will also be important to compare standalone Vera CPU systems with the integrated Vera Rubin platform. Different workloads may benefit from different balances of CPU cores, memory, GPU capacity and networking. Real-world results from coding agents, reinforcement-learning environments and data pipelines will be more informative than a single headline benchmark.
Conclusion
The NVIDIA Vera CPU is a notable step in the evolution of AI infrastructure. Its arrival shows that the agentic AI race is shifting from model capability alone to the efficiency of the entire computing system. Vera’s high-bandwidth memory, custom cores and close GPU connection are designed to accelerate the less visible work behind every tool-using agent.
For most businesses, the immediate action is to monitor rather than migrate. Once pricing, cloud access and independent tests become available, organisations can judge whether Vera delivers meaningful gains for their own AI workloads. If NVIDIA’s claims hold up outside controlled comparisons, purpose-built agent CPUs could become an important new layer of the AI stack.