NVIDIA Vera CPU (November 2026): The Processor Built for AI Agents — 88 Cores, 2.2x Faster Orchestration, and What It Means for AMD and Intel

AI-generated illustration — a concept image, not a photograph of a real product.

NVIDIA Vera CPU is the most interesting processor NVIDIA has ever built, not because it chases the fastest single-core score or the highest clock speed, but because it was designed for a job that barely existed a few years ago: running AI agents. On 5 November 2026, NVIDIA published new results for Vera, and an independent test from DeepInfra backed up the central claim with real production traffic. This guide explains what Vera is, how it works, what the independent numbers actually show, where the company claims go further than the evidence, and what the chip means for AMD, Intel, data-center buyers, and regular PC users. Everything here comes from the sources listed at the end. Vendor claims are labelled as claims, independent results are labelled as independent, and anything that has not been announced is stated plainly as unknown.

AI agents are changing what a CPU has to do. A chatbot answers one question and stops. An agent plans a task, calls a model, reads the result, runs code, checks the output, calls the model again, and keeps going until the job is done. Each step is small, but there are many of them, and the CPU does the work between every model call. That in-between work is where Vera is aimed. If you buy laptops, follow processors, run servers, or simply want to understand where computing is heading, Vera matters because it shows NVIDIA moving beyond the graphics processor and into the heart of the machine that serves AI.

1. What Is NVIDIA Vera? The CPU Built for AI Agents

Vera is NVIDIA's custom CPU designed for agentic AI, built to do the work that happens around each model call rather than the model call itself. That single sentence explains almost everything about the design. When an AI agent works, the graphics processor handles the large model computation, while the CPU handles orchestration logic, parsing the model's output, spawning sandboxed code execution, and moving data between the parts of the system. Vera is the processor NVIDIA built specifically for that surrounding work, and it can operate as a host CPU in Vera Rubin systems or as a standalone processor for agentic AI, reinforcement learning, and related workloads.

The name matters because it marks a shift in how NVIDIA talks about its own products. For years, the company was known first for graphics processors, and its earlier Grace CPU was presented mainly as a companion that fed data to those graphics processors. Vera is presented differently. It is a processor with its own job, its own architecture, and its own benchmarks, measured on agent work rather than on traditional server chores. NVIDIA calls its cores Olympus, a custom design rather than a licensed core used unchanged, and the whole chip is tuned for software that branches constantly, jumps between tasks, and manages many small jobs at once.

2. Why AI Agents Needed a New Kind of CPU

AI agents stress a CPU in a different way from ordinary software, because one request becomes a long chain of model calls, tool uses, and growing context. NVIDIA's developer blog, published in August 2026 and updated on 29 October 2026, explains the scale of the change with figures from OpenRouter's State of AI report, drawn from 100 trillion tokens of real-world usage. Average prompt tokens per request grew roughly fourfold, and single agentic requests consume 15 times the tokens of ordinary chat. Those two numbers describe a workload that keeps expanding while it runs, rather than a workload that arrives, gets processed, and ends.

Traditional server CPUs were designed for a broad mix of work: databases, web serving, virtual machines, and general computing. They are good at many things, which is why they remain the default choice for most servers. The question Vera raises is whether a CPU tuned for one specific pattern, the branch-heavy, tool-calling, sandbox-spawning pattern of agents, can beat a general-purpose CPU at that pattern by enough to matter. NVIDIA says yes, and the independent tests in this guide provide the first serious evidence either way.

3. Vera's Architecture: 88 Olympus Cores and Spatial Multithreading

Vera's headline design is 88 custom NVIDIA-designed Olympus cores, with Spatial Multithreading supplying 176 threads from those cores. The core count is the first thing to understand, because it sets the scale of the chip. Eighty-eight physical cores is a large CPU by any standard, and the threading design doubles the number of hardware threads the operating system can schedule work onto. A dual-socket Vera machine, the configuration used in the ClickHouse test described later in this guide, has 176 physical cores and 352 hardware threads, which is a very large amount of parallel capacity in a single system.

Olympus is the name NVIDIA gives to its custom core design, and the independent evidence suggests the design choice that matters most is how well the core handles instructions per clock on branch-heavy code. DeepInfra measured Vera sustaining 3.62 instructions per clock on its agent workload, compared with 2.44 for the best measured x86 chip in the same test, with sub-1 percent cache misses. Instructions per clock is a measure of how much useful work a core completes in each tick, and a higher figure on this kind of software means the core wastes less time waiting, guessing wrong about branches, or fetching data it should already have nearby.

Spatial Multithreading is the second design idea to understand. In simple terms, it lets each physical core handle two threads, so the chip presents 176 threads to software. That alone is not unusual, because many server CPUs offer two threads per core. What matters is how well the second thread performs on real agent work, and DeepInfra found that Spatial Multithreading delivered 20.5 percent more deployed throughput at high load, because the second thread absorbs virtualization housekeeping while the first thread does agent work. That is a practical benefit, not just a specification on a slide, because agent hosting relies heavily on isolated sandboxes that create extra housekeeping work.

The table below puts Vera's published specifications side by side with the comparison chips named in the independent tests, so you can see the scale of the design at a glance. The comparison chips are included because DeepInfra tested them directly, not because they are identical products, and the differences in core design and memory help explain why the benchmark results in later sections look the way they do.

Feature	NVIDIA Vera	AMD Zen 5 Turin	Intel Granite Rapids	Intel Sapphire Rapids
Role in test	New agent-focused CPU	Best x86 overall in DeepInfra test	Best Intel chip in DeepInfra test	Older Intel reference chip
Custom cores	88 Olympus cores	Zen 5 cores	Granite Rapids cores	Sapphire Rapids cores
Hardware threads	176 via Spatial Multithreading	Two threads per core design	Two threads per core design	Two threads per core design
Memory bandwidth	Up to 1.2 TB/s LPDDR5X	Server DDR5 platform	Server DDR5 platform	Server DDR5 platform
Memory capacity	Up to 1.5 TB	Platform dependent	Platform dependent	Platform dependent
Measured IPC on agent work	3.62	Not captured	2.44	2.44

4. Memory: 1.2 TB/s LPDDR5X and Up to 1.5 TB

Vera offers up to 1.2 TB/s of LPDDR5X memory bandwidth and up to 1.5 TB of memory, and that memory system is central to why the chip performs well on agents. Bandwidth is the rate at which the CPU can read and write data, and capacity is how much data it can hold at once. Agents need both, because they juggle growing context, tool results, code, and session state, and they do it for many agents at the same time. A fast CPU with slow memory would spend its time waiting, so NVIDIA paired the Olympus cores with a memory system designed to keep them fed.

The capacity figure also matters for a specific job NVIDIA describes: KV-cache offload. When a model serves a long conversation, it stores intermediate results, known as the KV cache, so it does not have to recompute everything from scratch on each turn. That cache grows with context, and offloading part of it to CPU memory is one way to serve longer sessions efficiently. A CPU with up to 1.5 TB of memory can hold a large amount of that cached data, which supports the longer, context-heavy sessions that agents create. The RAM market context is useful here too, and our guide to RAM prices in November 2026 explains why memory pricing and availability matter to anyone planning a system build right now.

5. NVIDIA's November 5 Claims — What the Company Says

NVIDIA's own claims, published on 5 November 2026, say Vera is dramatically faster than the best x86 CPUs at agent work, and those claims must be read as the company's own marketing figures until independent tests confirm them. The headline claim NVIDIA puts forward is 80 percent faster agentic CPU performance, which the company presents as a 1.8x bar. In a 100-megawatt data-center comparison, NVIDIA claims 2x faster agent speed than the most energy-efficient x86 CPU and 1.6x more agents per watt than the fastest x86 CPU. Those are large numbers, and they are the reason Vera received so much attention on the day the results were published.

The strongest reason to take the claims seriously is that NVIDIA published them alongside partner results and independent testing, rather than as a standalone slide. The weakest reason to accept them at face value is that every chip maker presents its best results in its best light, and the history of processor launches is full of headline numbers that looked smaller in independent hands. The honest position is in the middle: Vera is clearly a serious agent CPU with real design work behind it, and the size of its advantage depends on the workload, as the next sections show in detail.

6. The Independent Test: How DeepInfra Measured Vera

DeepInfra ran the most important independent Vera test so far, using a real agentic workload from its production environment, with the methodology locked before the hardware arrived. That last detail matters more than any single result, because it removes the suspicion that the test was tuned after seeing the numbers. DeepInfra replayed captured traffic, with no live model tokens, so differences in the results come from the CPU, not from randomness in model output. The workload is the kind of agent orchestration work Vera is designed for, run the way DeepInfra actually runs it for customers.

The fairness controls in the test deserve a close look, because they are what make the results worth trusting. Each chip ran in an identical fair configuration: 20 physical cores, one thread per core, a single NUMA node, with confinement verified and any run that escaped the core budget discarded. That means no chip received more cores, more memory locality, or a looser setup than the others. The field was NVIDIA Vera, AMD Zen 5 Turin, Intel Granite Rapids, and Intel Sapphire Rapids. A fifth chip, a DeepInfra production Xeon built on Granite Rapids silicon and tuned for inference, ran as an unranked reference and is discussed separately later in this guide.

DeepInfra tested two runtimes, and the difference between them is central to understanding agent hosting. The first runtime, runc, is close to bare metal, with minimal isolation overhead. The second, Kata microVMs, uses hardware-isolated sandboxes, which is how DeepInfra deploys untrusted multi-tenant agents in production. Agents from different customers cannot be allowed to interfere with each other, so real deployments pay for isolation. Testing both runtimes shows how Vera performs in a clean lab setup and in the messier, safer setup that real services actually use.

7. Orchestration Speed: 29 Milliseconds vs 64 — the 2.2x Result

On orchestration work-step latency per core, Vera measured 29 milliseconds in the runc runtime and 58 milliseconds in Kata microVMs, against 64 and 122 milliseconds for Sapphire Rapids, giving Vera a 2.2x advantage over the oldest chip in the field. The full table below shows every chip DeepInfra measured, and it is worth reading slowly, because the pattern is consistent across both runtimes. Vera is fastest in each row, AMD Zen 5 is the best x86 chip overall, and Intel Granite Rapids sits between AMD and the older Sapphire Rapids design.

Chip	Kata microVM latency per core	runc latency per core	Instructions per clock	Vera advantage in runc
NVIDIA Vera	58 ms	29 ms	3.62	Baseline, fastest
AMD Zen 5 Turin	87 ms	35 ms	Not captured	Vera 1.2x faster
Intel Granite Rapids	103 ms	51 ms	2.44	Vera 1.76x faster
Intel Sapphire Rapids	122 ms	64 ms	2.44	Vera 2.2x faster

Three comparisons in that table matter most. First, Vera at 29 milliseconds against Sapphire Rapids at 64 milliseconds is the 2.2x figure in the title of this article, and it is the largest gap in the test. Second, Vera is 1.76x faster than the best Intel chip, Granite Rapids, which is the more relevant comparison for anyone buying a new Intel server today. Third, Vera is 1.2x faster than the best x86 chip overall, AMD Zen 5 Turin, which is the closest contest and the one that shows Vera winning without running away from the field. DeepInfra's honest framing for the whole result was simple: Vera was the fastest of all four chips tested.

The mechanism behind the result is as interesting as the result itself. Vera sustained 3.62 instructions per clock on this branch-heavy workload, compared with 2.44 for the best measured x86 chip, with sub-1 percent cache misses. Instructions per clock measures how efficiently a core turns each tick into useful work, and the gap suggests Olympus was designed for exactly this kind of software: code that branches constantly, jumps between small tasks, and makes decisions at every step. The low cache-miss rate means the core usually found the data it needed close by, instead of waiting for slower memory. Together, those two facts explain why Vera completes each orchestration step faster.

8. More Agents per Core: 256 vs 160 at the Same Quality Bar

At the strict quality bar of a 99th-percentile response at or below 6 seconds with zero errors, Vera sustained 256 concurrent agents while the incumbents managed 160, giving Vera 1.33x to 1.6x more agents per core at identical quality of service. The range matters, because Granite Rapids reached 192 agents on its best run, so the advantage over the strongest incumbent on its best day was 1.33x, while the advantage over the 160-agent result was 1.6x. Either way, the direction is clear: Vera serves more agents on the same number of cores without letting the slowest responses slip past the limit or allowing errors.

The latency tail tells the same story in a different way. At 192 agents, Vera's latency tail matched what the incumbents showed at around 128 agents, which means Vera was serving roughly a third more agents while delivering the same worst-case responsiveness. Tail latency is the experience of the unluckiest user, the one whose request arrives at the busiest moment, and holding the tail down while adding load is harder than improving the average. A service that keeps its tail under control feels fast for everyone, not just for users who arrive at quiet moments.

DeepInfra also tested Vera at full socket scale, and the result is striking. On a full socket, Vera held more than 574 concurrent agents within the service objective and never broke it, with the ceiling set by the microVM boot path rather than by the CPU itself. That means the processor still had room to go, but the system that starts new sandboxes could not start them fast enough to find the limit. It is a good problem for a new platform to have, because it suggests the CPU is not the bottleneck, and it points to the software around the chip as the next place to improve.

9. Daytona's Sandbox Test: 1,000 Sandboxes in 10.8 Seconds

Daytona reported on 7 October 2026 that Vera prepared 1,000 sandboxes in 10.8 seconds, against 32.3 seconds for AMD EPYC 9575F and 88.4 seconds for EPYC 9755. Sandbox preparation is the work of starting an isolated environment where an agent can safely run code, and it matters because agents start sandboxes constantly. A service that prepares sandboxes slowly makes users wait before their agent can even begin, while a service that prepares them quickly feels responsive from the first moment. The gap here is large: Vera was roughly three times faster than the faster EPYC chip on this task and roughly eight times faster than the slower one.

The result holds at larger scale as well. Vera needed 27.4 seconds for 2,000 sandboxes, showing that the advantage did not collapse as the number of sandboxes doubled. With 2,000 concurrent sandboxes executing the same coding-agent task, Vera delivered 12.7 agent jobs per second, against 10.9 on EPYC 9755 and 8.1 on EPYC 9575F. Jobs per second is a throughput measure, and it shows Vera not only starting sandboxes faster but also completing agent work faster once those sandboxes are running. The two results together cover the full cycle: start fast, work fast, finish fast.

It is worth noting that the two EPYC chips trade places between the two measures. The 9575F prepares sandboxes faster than the 9755, but the 9755 completes more agent jobs per second once the sandboxes are running. That is a useful reminder that no single number captures a whole platform, and that buyers should match the benchmark to the job they actually run. For a service that starts many short-lived sandboxes, preparation speed matters most. For a service that runs long sessions inside a smaller number of sandboxes, sustained job throughput matters more. Vera leads on both measures in this test, which is why Daytona's result attracted attention.

10. Perplexity, DuckDB and ClickHouse: What the Other Partners Found

Three more partners reported independent Vera results in October 2026, and all three point in the same direction, even though they tested very different kinds of work. Perplexity, DuckDB, and ClickHouse each used Vera for a different job, and each found a clear advantage over the baseline they compared against. No single partner result proves the whole case, but three results in different domains make the pattern harder to dismiss as a quirk of one test.

Perplexity ran early Vera tests on agentic coding workflows and found them roughly 1.5x faster than its current production reference, with concurrent sandboxes starting up to 1.9x faster. The company plans to deploy Vera in an upcoming production system, which is a meaningful signal, because a production deployment is a commitment that goes beyond a press quote. Perplexity's result also matches the Daytona finding on sandbox starts, which suggests the sandbox advantage is real across different companies and different setups, not an artifact of one test harness.

DuckDB published its result on 16 October 2026, testing a database workload rather than an agent workload. Vera ran 1.53x to 1.61x faster in its power test and 1.37x to 1.41x faster in its throughput test than an Intel Xeon 6 baseline. A database test stresses different parts of a CPU from an agent test, because it involves scanning data, filtering rows, and aggregating results, so a win here suggests Vera's strengths extend beyond the narrow agent pattern. The range in each figure reflects different queries and configurations inside the test, and the honest way to read it is that Vera was roughly one and a half times faster on power and roughly one and four tenths times faster on throughput.

ClickHouse published its result on 21 October 2026, and it is the most striking of the three for a different reason. Its standard binary, with no Vera-specific tuning, reached the top of the ClickBench hardware leaderboard on a dual-socket Vera machine with 176 physical cores and 352 hardware threads. Reaching the top with an untuned binary matters, because it means customers do not need to wait for special software work to see the benefit. The table below collects the three partner results so you can compare them at a glance.

Partner	Workload tested	Result reported	Baseline compared
Perplexity	Agentic coding workflows	Roughly 1.5x faster, sandboxes up to 1.9x faster to start	Current production reference
Daytona	Sandbox preparation and agent jobs	1,000 sandboxes in 10.8 seconds, 12.7 jobs per second	AMD EPYC 9575F and 9755
DuckDB	Database power and throughput	1.53 to 1.61x on power, 1.37 to 1.41x on throughput	Intel Xeon 6
ClickHouse	ClickBench analytical queries	Top of leaderboard, standard binary, no tuning	All leaderboard entries

11. MLPerf Inference v6.1: Vera Rubin's First Peer-Reviewed Numbers

In MLPerf Inference v6.1, run by MLCommons, the Vera Rubin NVL72 system's DeepSeek-R1 and Qwen3-VL entries both landed in Preview, which means the system behind the headline multipliers is not a shipping, orderable product yet. Preview status is an important qualification, and it is confirmed by an independent MLPerf v6.1 writeup. It means the submission was made and the numbers were recorded, but the system is not yet in the fully validated, shipping category that buyers usually rely on for purchasing decisions. Readers should treat the MLPerf figures in this section as early, promising evidence rather than as final product benchmarks.

The numbers themselves are large. On Qwen3-VL, NVIDIA's own MLPerf recap puts Vera Rubin NVL72 at up to 3.7x the throughput of GB300 NVL72 across offline, server, and interactive scenarios. On DeepSeek-R1, using TensorRT-LLM, the figure is up to 2.5x. On the SemiAnalysis AgentX agentic benchmark, Vera Rubin NVL72 delivered 30x the performance of GB300 NVL72, with 2x at 110 tokens per second per user, 10x at 130, and 30x at 160, all three figures from the same preview submission. NVIDIA's own developer blog says the AgentX results are pending SemiAnalysis review, which is a further reason to read the 30x figure with care.

The AgentX pattern deserves explanation, because it shows why agent benchmarks behave differently from ordinary model benchmarks. At low load, the two systems are close. As the load per user rises, the gap widens sharply, which is the signature of a system that holds its performance under pressure while the comparison system falls behind. That is consistent with the DeepInfra finding that Vera holds its latency tail while serving more agents, and it is the reason agent benchmarks are becoming a separate category. MLPerf v6.1 added a new agentic-inference workload category, which shows the industry itself now treats agent serving as a distinct job with its own measures.

12. Vera vs EPYC and Xeon: The Full Comparison

Across every head-to-head test in this guide, Vera leads the best current EPYC and Xeon chips on agent work, with the gap largest against older designs and smallest against the newest AMD silicon. That sentence is the fair summary of the evidence so far. It avoids the trap of quoting only the biggest number, and it avoids the opposite trap of pretending the contest is close everywhere. Vera wins clearly, but the size of the win depends on which chip it faces and which job it runs.

On orchestration latency, the ranking is Vera first, AMD Zen 5 second, Intel Granite Rapids third, and Sapphire Rapids fourth, with Vera 1.2x faster than AMD, 1.76x faster than the best Intel, and 2.2x faster than the oldest Intel in the field. On capacity at a strict quality bar, Vera serves 256 agents against 160 for the incumbents, with Granite Rapids reaching 192 on its best run. On sandbox preparation, Vera is roughly three times faster than EPYC 9575F and roughly eight times faster than EPYC 9755 at 1,000 sandboxes. On database work, Vera is roughly one and a half times faster than Xeon 6 on power and roughly one and four tenths times faster on throughput. No test in this guide shows an incumbent beating Vera on agent work.

The table below brings the comparison together in one place, with a practical note on what each result means for a buyer choosing between platforms. The table is a summary, not a substitute for the detailed sections above, and readers who need the full method should return to the DeepInfra and Daytona sections before making a decision.

Comparison	Vera result	Best rival result	What it means
Orchestration latency, runc	29 ms per step	35 ms, AMD Zen 5	Agents respond faster at every step
Orchestration latency, Kata	58 ms per step	87 ms, AMD Zen 5	Secure sandboxes cost Vera less speed
Concurrent agents at strict quality bar	256 agents	160 agents, up to 192 best run	More users per machine, same quality
Sandbox preparation, 1,000 sandboxes	10.8 seconds	32.3 seconds, EPYC 9575F	Agents start work much sooner
Agent jobs, 2,000 sandboxes	12.7 jobs per second	10.9 jobs per second, EPYC 9755	Higher throughput at full load
Database power test	1.53 to 1.61x faster	Intel Xeon 6 baseline	Strength extends beyond agents

13. Vera vs Grace: NVIDIA's Own CPU, Beaten

Vera beats NVIDIA's own earlier Grace CPU on agent work by a wide margin, and the comparison is useful because it shows how far NVIDIA's CPU design has moved in one generation. The clearest evidence comes from a bonus finding in the DeepInfra test. With the 256-agent fleet running at full tilt, the remaining 56 cores on the Vera machine served a real language model, gpt-oss-20b, on the CPU at 118.5 tokens per second decode under full agent contention. That figure is more than an entire 72-core NVIDIA Grace socket, the GH200, doing nothing else, which measured 84.5 tokens per second. The agent fleet stayed within its service objective while this was happening, losing under 3 percent of its performance.

Read that result slowly, because it contains two wins at once. First, the leftover cores on a busy Vera machine outperformed a whole Grace socket that had nothing else to do. Second, the main agent workload barely noticed the extra job, losing under 3 percent while sharing the machine. That is the kind of result that changes how a data-center operator thinks about capacity, because it means a Vera machine can do agent serving and CPU inference at the same time without either job collapsing. The capstone demonstration made the point even clearer: 256 secure agent microVMs, plus CPU inference on leftover cores, plus graphics-processor model serving on an installed RTX PRO 6000, with every layer within roughly 2 to 4 percent of its performance running alone.

Grace was a good CPU for its time, designed mainly to feed data to graphics processors and to serve as a capable host in NVIDIA systems. Vera is a different kind of product, designed to do agent work itself, and the gap between them shows how quickly the job has changed. A buyer who owns Grace systems should not read this as a reason to discard working hardware, because Grace remains a capable platform for the jobs it was bought for. The lesson is narrower and more useful: for new agent deployments, Vera is a large step forward from NVIDIA's own previous CPU, not just from rival chips.

14. The Sandbox Tax: Why Isolation Costs More Than Silicon

DeepInfra found that across every chip including Vera, the biggest tax on agent hosting is the secure-isolation layer, not the CPU, and that finding cuts both ways for Vera. It is good news for NVIDIA, because Vera is the best-behaved chip under that stack, and it is a warning for the whole industry, because the cost of isolation is large on every platform. MicroVM sandboxing costs 1.5x to 2x on typical work and roughly 5x to 13.5x on process spawning across architectures, with Arm paying about half the x86 spawn tax. Those are large penalties, and they apply before the CPU even gets to do the useful work.

To understand why, think about what a sandbox must do. An agent runs code it did not write, often code produced by a model, on behalf of a customer who must not be able to reach another customer's data. The system must start an isolated environment, set its limits, load the code, run it, collect the result, and tear the environment down, all without letting anything leak across the boundary. Each of those steps costs time, and process spawning, the act of starting a new process inside the sandbox, is the most expensive step of all. The Kata microVM numbers in the DeepInfra test show that cost directly: Vera rises from 29 milliseconds in runc to 58 milliseconds in Kata, AMD rises from 35 to 87, and Intel rises from 51 to 103 and from 64 to 122.

Vera's advantage under isolation is that it loses less speed than its rivals when the sandbox layer is added, and that is a design win, not a lucky result. The Olympus cores handle the branching, housekeeping, and context-switching work of isolation more efficiently, and Spatial Multithreading helps by letting the second thread absorb virtualization housekeeping. The result is that Vera in Kata microVMs, at 58 milliseconds, is still faster than the best x86 chip in the near-bare-metal runc runtime, at 35 milliseconds for AMD only in the sense that the gap narrows, and Vera remains clearly ahead of every rival in both runtimes. Readers should compare the rows carefully rather than taking a single number away from this section.

15. What Vera Means for AMD and Intel

For AMD and Intel, Vera is a serious competitive challenge on agent work, but it is not the end of either company, because both still lead in areas Vera has not yet proven itself. The honest assessment starts with what Vera does well and what it does not yet claim. Vera leads on orchestration latency, on agents per core at a strict quality bar, on sandbox preparation, and on the database tests reported so far. Those are real wins, and they are wins on the workload that is growing fastest in data centers. AMD and Intel must respond, and they have the engineering depth to do so.

AMD's position is the stronger of the two in the current evidence. Zen 5 Turin was the best x86 chip overall in the DeepInfra test, at 35 milliseconds in runc and 87 milliseconds in Kata, and AMD's EPYC line remains a proven, widely deployed platform with broad software support and a large customer base. The gap between Vera and Zen 5 is real, at 1.2x on orchestration latency, but it is not the kind of gap that forces an immediate switch for customers who already run AMD infrastructure. AMD can respond with tuning, with new silicon, and with pricing, and the company has done so before when a rival took a lead in one workload.

The CPU market has already seen a shift in recent weeks, and our article on how the CPU market just flipped gives useful background on the competitive picture. For readers who want the Intel and AMD comparison in depth, our Nova Lake versus Zen 6 guide covers the next generation of desktop and server silicon from both companies. The short version for Vera is this: NVIDIA has taken a real lead in agent CPUs, AMD is close enough to fight back, Intel has more ground to recover, and the next year of benchmarks will decide whether the lead grows or narrows.

16. What Vera Means for Data-Center Buyers

For data-center buyers, Vera changes the capacity calculation for agent services, because serving more agents per machine at the same quality reduces the number of machines, the power, and the floor space a service needs. The DeepInfra capacity result is the number to start with: 256 agents on Vera against 160 on the incumbents, at a 99th-percentile response of 6 seconds or less with zero errors. If a service needs to serve 2,560 agents, the difference between 256 and 160 per machine is the difference between 10 machines and 16 machines, which is a large saving in hardware, power, cooling, and management. The exact saving depends on the service, the workload, and the price of the hardware, which has not been announced, but the direction is clear.

Buyers should also weigh the full-socket result, because it shows Vera has headroom beyond the tested configuration. Holding more than 574 concurrent agents within the service objective, with the ceiling set by the microVM boot path rather than the CPU, means the processor is not the limiting factor at that scale. A buyer planning a large deployment should ask NVIDIA and its partners about the boot path, the orchestration software, and the tooling that starts sandboxes, because those layers will decide how much of Vera's capacity a real service can actually use. The bonus inference finding also matters for capacity planning: leftover cores can serve a real model on the CPU while the agent fleet runs, which means one machine can do two jobs at once without either job losing much performance.

There are three practical steps a buyer should take before committing. First, test Vera on the buyer's own agent workload, not just on a published benchmark, because the gap between Vera and the incumbents varies by workload and the buyer's workload is the only one that matters. Second, check software support, including the orchestration framework, the sandbox runtime, and the model-serving stack, because a fast chip with weak software will not deliver its benchmark numbers in production. Third, get a firm price and a firm delivery date, because neither has been announced for a standalone Vera product, and a capacity saving on paper means little without a price to compare against the incumbents.

17. What Vera Means for Regular PC Buyers

For regular PC buyers, Vera means nothing to buy right now, and that is the most important fact in this section. There is no standalone price for Vera, no retail or consumer channel exists, and Vera Rubin systems are reservation-only. There is nothing for a regular PC buyer to purchase, and no Pakistan price exists, so this guide will not invent one. If you are building or buying a PC today, Vera should not change your decision, because you cannot buy it, and the chips you can buy remain the right choice for the jobs a home PC actually does.

That does not mean Vera is irrelevant to PC buyers. The ideas behind Vera, fast orchestration, efficient sandboxing, and high memory bandwidth for agent work, will shape the software and services PC users rely on, even if the chip itself stays in data centers. When you use an AI agent in a browser, in a coding tool, or in a productivity app, that agent runs on a server, and a faster server CPU means a faster response for you. The benefit arrives indirectly, through the service, rather than directly, through a chip in your own machine. That is how most server advances reach home users, and Vera is no exception.

PC buyers should also watch for the longer-term effect on laptop and desktop processors. If agent work becomes a standard part of everyday computing, future consumer chips may borrow ideas from Vera, such as cores tuned for branch-heavy code or memory systems built for large context. That is speculation, not a product announcement, and no such consumer chip has been announced. For now, the practical advice is simple: choose your next laptop or desktop on the processors you can actually buy, using verified reviews and buying guides, and treat Vera as a sign of where the industry is heading rather than as a shopping option.

18. Availability and Price: What Is Known and What Is Not

No standalone price for Vera has been announced, no retail or consumer channel exists, and Vera Rubin systems are reservation-only, so there is nothing for an ordinary buyer to order today. Vera Rubin NVL72 went generally available on OCI on 27 October 2026, according to prior blog coverage, which means cloud customers can access the system through that platform, but that is not the same as buying a Vera chip on its own. For price context only, Spheron rents B200 at 10.25 dollars per hour and B300 at 11.52 dollars per hour on demand as of 6 November 2026, while GB300 and Vera Rubin stay reservation-only. Those rental figures describe different graphics processors, not Vera, and they are included only to show how cloud access is priced. The MLPerf entries for Vera Rubin NVL72 are in Preview, which confirms the system is not yet a shipping, orderable product in the usual sense. Anyone who tells you a Vera price, a retail launch date, or a Pakistan price is inventing it, because none of those facts has been announced.

19. The Bigger Picture: The CPU Is Becoming an AI Component

Vera shows the CPU becoming an AI component in its own right, rather than a general-purpose helper that merely supports the graphics processor. For years, the story of AI hardware was simple: the graphics processor did the model work and the CPU did everything else. Vera breaks that story, because the CPU now has a named job inside the AI system, orchestration, tool execution, sandbox management, and KV-cache offload, and it is measured on that job with its own benchmarks. The supporting technologies in the Vera Rubin platform, including NVLink 6, ConnectX-9, BlueField-4, Spectrum-X, Dynamo disaggregated serving, NVFP4, TensorRT-LLM WideEP, and speculative decoding, show a system designed as a whole, with Rubin graphics processors handling large context and efficient decoding while Vera handles the work around each model call. AgentX, the new agentic-inference workload category in MLPerf v6.1, confirms the industry now treats agent serving as a distinct job. The CPU market has already shifted in response, and the question for the next year is how quickly AMD and Intel answer with agent-tuned designs of their own.

20. Verdict: Should You Care About Vera?

Yes, Vera is worth caring about if you run agents, buy servers, or follow processors, and it is worth ignoring as a purchase until a price and a channel exist. The evidence in this guide is consistent: DeepInfra measured Vera at 29 milliseconds against 64 for Sapphire Rapids, 1.2x faster than the best x86 chip, with 256 agents against 160 at the same strict quality bar, while Daytona, Perplexity, DuckDB, and ClickHouse each found a clear advantage on different work. NVIDIA's headline claims go further than the independent evidence, and the MLPerf figures are Preview entries, so the full picture is still forming, but the direction is clear and the independent tests agree. For regular PC buyers, the next step is nothing, because there is nothing to buy and no Pakistan price exists. Vera matters most as a sign of where computing is heading: the CPU is no longer just the host for the AI accelerator, it is part of the AI system itself, and the companies that build the best agent CPU will shape the next generation of AI services.

Frequently Asked Questions

What is the NVIDIA Vera CPU?

Vera is NVIDIA's custom CPU designed for agentic AI, built to do the work that happens around each model call rather than the model call itself. It has 88 custom Olympus cores, with Spatial Multithreading supplying 176 threads, up to 1.2 TB/s of LPDDR5X memory bandwidth, and up to 1.5 TB of memory. When an AI agent works, the graphics processor handles the large model computation, while Vera handles orchestration logic, parsing the model's output, spawning sandboxed code execution, and moving data between the parts of the system. Vera can operate as a host CPU in Vera Rubin systems or as a standalone processor for agentic AI, reinforcement learning, and related workloads. It is not a consumer desktop chip, and no standalone price or retail channel has been announced, so ordinary buyers cannot purchase it today.

How fast is Vera compared with x86 CPUs?

On DeepInfra's independent test, Vera measured 29 milliseconds per orchestration work step in the runc runtime, against 35 milliseconds for AMD Zen 5 Turin, 51 milliseconds for Intel Granite Rapids, and 64 milliseconds for Intel Sapphire Rapids. That gives Vera a 1.2x advantage over the best x86 chip overall, a 1.76x advantage over the best Intel chip, and a 2.2x advantage over Sapphire Rapids. In Kata microVMs, the secure sandbox runtime used in production, Vera measured 58 milliseconds against 87 for AMD, 103 for Granite Rapids, and 122 for Sapphire Rapids. NVIDIA's own claims go further, with a headline of 80 percent faster agentic CPU performance, but those are the company's figures, while the DeepInfra numbers come from an independent test with a locked methodology and identical configurations for every chip.

How many AI agents can Vera serve at once?

At DeepInfra's strict quality bar, a 99th-percentile response at or below 6 seconds with zero errors, Vera sustained 256 concurrent agents on identical 20-core partitions inside secure microVMs, while the incumbents managed 160, with Granite Rapids reaching 192 on its best run. That is 1.33x to 1.6x more agents per core at identical quality of service. On a full socket, Vera held more than 574 concurrent agents within the service objective and never broke it, with the ceiling set by the microVM boot path rather than the CPU itself. Vera's latency tail at 192 agents matched what the incumbents showed at around 128 agents, which means it served roughly a third more agents while delivering the same worst-case responsiveness for the unluckiest user.

What is Spatial Multithreading?

Spatial Multithreading is the threading design that lets Vera's 88 physical Olympus cores present 176 hardware threads to software, so the operating system can schedule two threads on each core. Many server CPUs offer two threads per core, so the idea itself is not new, but the measured benefit on agent work is what matters. DeepInfra found that Spatial Multithreading delivered 20.5 percent more deployed throughput at high load, because the second thread absorbs virtualization housekeeping while the first thread does agent work. Agent hosting relies heavily on isolated sandboxes that create extra housekeeping work, so a second thread that handles that housekeeping frees the first thread to serve agents. A dual-socket Vera machine has 176 physical cores and 352 hardware threads in total.

What did Daytona find about Vera and sandboxes?

Daytona reported on 7 October 2026 that Vera prepared 1,000 sandboxes in 10.8 seconds, against 32.3 seconds for AMD EPYC 9575F and 88.4 seconds for EPYC 9755, and Vera needed 27.4 seconds for 2,000 sandboxes. With 2,000 concurrent sandboxes executing the same coding-agent task, Vera delivered 12.7 agent jobs per second, against 10.9 on EPYC 9755 and 8.1 on EPYC 9575F. Sandbox preparation matters because agents start isolated environments constantly, and a service that prepares them slowly makes users wait before work can begin. Perplexity found a matching result, with concurrent sandboxes starting up to 1.9x faster on Vera, which suggests the sandbox advantage is real across different companies and setups.

What did Perplexity, DuckDB and ClickHouse find?

Perplexity ran early Vera tests on agentic coding workflows and found them roughly 1.5x faster than its current production reference, with concurrent sandboxes starting up to 1.9x faster, and the company plans to deploy Vera in an upcoming production system. DuckDB reported on 16 October 2026 that Vera ran 1.53x to 1.61x faster in its power test and 1.37x to 1.41x faster in its throughput test than an Intel Xeon 6 baseline, on a database workload rather than an agent workload. ClickHouse reported on 21 October 2026 that its standard binary, with no Vera-specific tuning, reached the top of the ClickBench hardware leaderboard on a dual-socket Vera machine with 176 physical cores and 352 hardware threads. Three different domains, one consistent picture of a clear Vera advantage.

What do the MLPerf Inference v6.1 results show?

In MLPerf Inference v6.1, the Vera Rubin NVL72 system's DeepSeek-R1 and Qwen3-VL entries both landed in Preview, which means the system behind the headline multipliers is not a shipping, orderable product yet, and readers should treat the figures as early evidence. On Qwen3-VL, NVIDIA's MLPerf recap puts Vera Rubin NVL72 at up to 3.7x the throughput of GB300 NVL72 across offline, server, and interactive scenarios, and on DeepSeek-R1 the figure is up to 2.5x. On the SemiAnalysis AgentX agentic benchmark, Vera Rubin NVL72 delivered 30x the performance of GB300 NVL72, with 2x at 110 tokens per second per user, 10x at 130, and 30x at 160, all from the same preview submission, and NVIDIA's own developer blog says the AgentX results are pending SemiAnalysis review. MLPerf v6.1 added a new agentic-inference workload category.

Can I buy a Vera CPU for my PC?

No, there is nothing for a regular PC buyer to buy. No standalone price for Vera has been announced, no retail or consumer channel exists, and Vera Rubin systems are reservation-only. Vera Rubin NVL72 went generally available on OCI on 27 October 2026, which gives cloud customers access through that platform, but that is not the same as buying a chip for a home PC. No Pakistan price exists, and this guide will not invent one. If you need a PC today, choose from the processors you can actually buy, using verified reviews and buying guides. Vera's benefit reaches home users indirectly, through faster agent services running on servers, rather than directly through a chip in your own machine, and any future consumer product will be covered here when it is announced.

Is Vera better than AMD EPYC and Intel Xeon?

On the agent work tested so far, yes, Vera leads the best current EPYC and Xeon chips, with the gap largest against older designs and smallest against the newest AMD silicon. Vera is 1.2x faster than AMD Zen 5 Turin on orchestration latency, 1.76x faster than Intel Granite Rapids, and 2.2x faster than Sapphire Rapids, and it serves 256 agents against 160 at the same strict quality bar. On sandbox preparation it is roughly three times faster than EPYC 9575F, and on database power it is roughly one and a half times faster than Xeon 6. That does not prove Vera wins on every server workload, because the tests focus on agents, sandboxes, and databases, and neither price nor long-term reliability has been established yet. AMD remains close enough to fight back, and Intel retains enormous scale.

What is the sandbox tax, and why does it matter?

The sandbox tax is the performance cost of running agents inside secure, isolated environments, and DeepInfra found it is the biggest tax on agent hosting, larger than the differences between CPUs. MicroVM sandboxing costs 1.5x to 2x on typical work and roughly 5x to 13.5x on process spawning across architectures, with Arm paying about half the x86 spawn tax. A sandbox must start an isolated environment, set its limits, load code, run it, collect the result, and tear the environment down without letting anything leak between customers, and each step costs time. Vera is the best-behaved chip under that stack, losing less speed than its rivals when isolation is added, but the industry still needs faster, lighter isolation before agent hosting reaches its full potential.

Sources

Key facts in this article come from the sources below. NVIDIA claims are labelled as claims in the text, and independent results are labelled as independent. Readers who want the full method, the complete tables, or the original wording should read the original reports.