Quick Answer
Nvidia’s Vera CPU is an 88-core, in-house Arm-compatible processor built to improve the single-threaded work that AI agents use for orchestration and tool calling. At Hot Chips on August 24, 2026, Nvidia cited roughly 1.8x agentic-workload speedup over Grace Blackwell, while AMD answered with a 3.3x rack-level claim for Venice. Treat both performance figures as vendor claims and wait for comparable independent tests.
Key Takeaways
- Nvidia’s Vera CPU uses 88 custom Olympus cores with spatial multithreading.
- Vera uses LPDDR5X memory and a monolithic compute die rather than a chiplet design.
- Nvidia says Vera improves agentic workloads by roughly 1.8x versus Grace Blackwell.
- AMD says its upcoming 256-core Zen 6 Venice CPU delivers 3.3x higher rack-level performance than Vera.
- Nvidia says Dell, HPE, Lenovo, and Supermicro are building Vera systems at scale.
What did Nvidia reveal about the Vera CPU at Hot Chips 2026?
Nvidia revealed that the Vera CPU uses 88 custom Olympus cores, LPDDR5X memory, spatial multithreading, and a monolithic compute die. The company presented the architecture at Hot Chips on August 24, 2026, providing a more detailed view of the CPU component in its Vera Rubin platform than its earlier product announcement supplied.
Vera matters because Nvidia is positioning the CPU as part of an AI system rather than as a standalone general-purpose server processor. Nvidia’s design emphasis is on the host-side work around AI models, including orchestration and tool calling, where response speed from an individual CPU thread can affect the broader system’s behavior.
The Vera CPU also marks a shift from Nvidia’s prior Grace approach. According to ServeTheHome’s Hot Chips coverage, Vera uses Nvidia’s Olympus architecture instead of the licensed Arm Neoverse cores used in earlier Grace chips. The practical result is that Nvidia has more control over how the CPU is tuned for its own AI platforms, although independent testing will determine whether that control translates into a meaningful advantage for customers.
Why is Nvidia designing Vera around AI agents?
Nvidia is designing Vera around AI agents because agentic workloads require CPU coordination work alongside GPU model processing. Nvidia described Vera as a processor for the orchestration and tool-calling tasks that accompany AI agents, rather than presenting raw core count as the primary measure of success.
Agentic systems can require repeated interactions among a model, external tools, storage, and other services. Nvidia’s emphasis on single-thread performance suggests that the company sees delays in this coordination layer as an important constraint, particularly when an AI system must make sequential decisions rather than complete one isolated model request.
Nvidia’s framing also reflects the company’s broader effort to own more of the AI infrastructure stack. That strategy is visible in large data center commitments, where compute hardware, facilities, and AI deployment plans increasingly overlap. Vera does not replace GPUs in Nvidia’s platform, but it gives Nvidia a purpose-built CPU component for the workloads that surround GPU inference and training.
How does the Olympus architecture differ from Nvidia’s prior Grace CPUs?
Nvidia’s Olympus architecture differs from prior Grace CPU designs because Olympus is an Nvidia in-house core architecture that is Arm v9.2-compatible. Grace chips used licensed Arm Neoverse cores, while Vera gives Nvidia a custom CPU design that can be developed around the company’s AI platform requirements.
The Hot Chips presentation also identified spatial multithreading as part of Vera’s architecture. Nvidia paired that core design with LPDDR5X memory and a monolithic compute die, meaning the principal compute die is not divided into separate chiplets. The design choice is important because CPU architecture, memory configuration, and physical package design all influence how efficiently a processor handles data movement and coordination work.
The available details support a clearer distinction between Vera and Grace, but they do not establish a universal performance winner against every competing CPU. Workload mix, software configuration, system design, and rack-level deployment all affect real-world results. The most sensible approach is to treat the architectural disclosure as evidence of Nvidia’s direction, not as a complete replacement for independent server testing.
| Processor platform | Known architecture detail | Performance claim | Important limitation |
|---|---|---|---|
| Nvidia Vera | 88 custom Olympus cores, spatial multithreading, LPDDR5X memory, monolithic compute die | Roughly 1.8x speedup in agentic workloads versus Grace Blackwell, according to Nvidia | The figure is vendor-reported and focuses on Nvidia’s stated agentic-workload framing |
| AMD Zen 6 Venice | Upcoming 256-core CPU, according to AMD’s public claim | 3.3x higher rack-level performance than Vera, according to AMD | The claim uses a rack-level comparison and requires independent, like-for-like testing |
What does Nvidia’s 1.8x agentic-workload claim mean?
Nvidia’s roughly 1.8x agentic-workload claim means the company reports a substantial improvement over its prior Grace Blackwell CPU in the specific AI coordination tasks it highlighted. The figure came from Nvidia’s Hot Chips presentation and is a self-reported performance result, not an independent benchmark conclusion.
Nvidia’s claim matters because the company is arguing that AI infrastructure needs stronger CPU performance for more than traditional server management. If AI agents repeatedly call tools, move between services, and coordinate model responses, the CPU can influence overall system responsiveness even when GPUs perform the main model computation.
At the same time, the 1.8x figure does not show how Vera will perform across every database, cloud application, enterprise workload, or model deployment. Tom’s Hardware’s Vera analysis reported both Nvidia’s agentic-workload figure and AMD’s competing response. Buyers should wait for benchmarks that use the same software, system configuration, power constraints, and measurement method before treating either number as a procurement verdict.
Why did AMD challenge Nvidia’s Vera CPU so quickly?
AMD challenged Nvidia’s Vera CPU quickly because CPUs are becoming a more visible part of the AI server competition. AMD publicly said its upcoming 256-core Zen 6 Venice CPU would beat Vera by 3.3x in rack-level performance, creating an immediate counterpoint to Nvidia’s agent-focused architecture presentation.
AMD’s claim and Nvidia’s claim do not measure the same thing in the information currently available. Nvidia emphasized agentic workloads compared with Grace Blackwell, while AMD emphasized rack-level performance compared with Vera. Those differences matter because a rack-level result can reflect a broader system configuration than a CPU-focused workload result.
The competitive pressure is also visible in the market context. CNBC reported that AMD and Intel shares had risen 128% and 149%, respectively, during 2026, compared with Nvidia’s 8% gain. Share performance does not prove one company’s hardware is better, but it shows investors are treating the AI server CPU market as a meaningful contest rather than an uncontested Nvidia category.
Where is Nvidia deploying the Vera CPU?
Nvidia says Vera is already moving toward large-scale system deployment, with Dell, HPE, Lenovo, and Supermicro building Vera systems. Nvidia also says Anthropic, OpenAI, and SpaceXAI are among the customers exploring Vera, while Tom’s Hardware reported that Nvidia announced a new SpaceXAI deployment the day before the Hot Chips presentation.
The deployment list matters because a CPU architecture only becomes commercially important when system manufacturers can package it into complete servers and customers can test it in operational environments. Dell, HPE, Lenovo, and Supermicro are major server suppliers, so their involvement gives Vera a route into enterprise and data center purchasing channels.
Nvidia’s official announcement describes the manufacturers and customer interest, but exploration is not the same as a disclosed volume purchase or a completed production rollout. Nvidia’s Vera announcement identifies the companies involved, while buyers still need confirmed availability, system specifications, pricing, and independent performance data before making a deployment decision.
What should AI infrastructure buyers watch next?
AI infrastructure buyers should watch for independent Vera and Venice testing that compares equivalent racks, software stacks, power limits, and AI workloads. Nvidia’s 1.8x agentic-workload result and AMD’s 3.3x rack-level result both offer useful signals, but the claims cannot yet be treated as directly comparable measurements.
Buyers should also ask whether a planned workload is genuinely limited by CPU orchestration. Organizations running AI agents that repeatedly call tools may find Nvidia’s Vera design relevant, while organizations with different bottlenecks may prioritize GPU capacity, memory, networking, software support, or total rack efficiency instead.
The most important practical question is whether a server platform improves the complete workload that an organization intends to run. AI infrastructure costs remain sensitive to component availability and system configuration, especially as memory pricing pressure affects new PC and server purchasing decisions. Procurement teams should require workload-specific evidence before selecting a platform based on a single vendor benchmark.
FAQ
What is Nvidia Vera?
Nvidia Vera is an 88-core CPU in Nvidia’s Vera Rubin AI platform that uses custom Olympus cores, spatial multithreading, LPDDR5X memory, and a monolithic compute die. Nvidia is positioning Vera for the CPU orchestration work associated with AI agents and tool calling.
How fast is Nvidia’s Vera CPU?
Nvidia says Vera delivers roughly 1.8x faster agentic-workload performance than its prior Grace Blackwell CPU. The performance result is self-reported by Nvidia, so independent testing is needed to determine how it translates across different server workloads.
What is AMD’s response to Nvidia Vera?
AMD says its upcoming 256-core Zen 6 Venice CPU delivers 3.3x higher rack-level performance than Vera. AMD’s result is also a vendor claim, and the available information does not establish that AMD and Nvidia measured identical workloads or system configurations.
Does Vera use Arm CPU cores?
Nvidia Vera uses Nvidia’s custom Olympus cores, which are Arm v9.2-compatible. The Olympus design differs from prior Grace chips because Grace used licensed Arm Neoverse cores rather than Nvidia’s in-house CPU architecture.
Which companies are building Nvidia Vera systems?
Nvidia says Dell, HPE, Lenovo, and Supermicro are building Vera systems at scale. Nvidia also identifies Anthropic, OpenAI, and SpaceXAI as customers exploring Vera, although exploration does not confirm a completed production deployment for every company.