
发布于 11 月 6 日 • 编辑于 11 月 6 日 • 最初刊载于 toolsdig.com
NVIDIA 多年来一直在向训练模型的人销售 GPU。2026 年 11 月 5 日,它公布了一个不同方向的结果:一款面向“训练之后”那部分工作的 CPU——运行智能体(agents)。这款芯片名为 Vera,NVIDIA 的核心宣称是:在一项 100 MW 数据中心对比中,它的每瓦可运行智能体数量比最快的 x86 CPU 高 1.6 倍,并且智能体运行速度比最节能的 x86 CPU 快 2 倍。
这些数字来自 NVIDIA,而且是 NVIDIA 自己的对比。我没有亲自跑过这些测试,目前也没有独立方复现结果;本文会一直强调这一点——因为即使带着这种保留,这些数字仍然值得你关注。智能体并不像多数人以为的那样在“模型调用”上遇到瓶颈。它们的瓶颈在于模型调用周围的一切:启动沙箱、编排工具调用、运行数据库、在不同步骤之间搬运上下文。这些都属于 CPU 工作,也正是 Vera 瞄准的方向。
Vera 是 NVIDIA 首款专为 AI 智能体工作负载设计的数据中心 CPU。按其公布的规格:
88 个定制 Olympus 核心,通过 NVIDIA 称为 Spatial Multithreading 的技术提供 176 线程
最高 1.2 TB/s LPDDR5X 内存带宽,最高 1.5 TB 内存
250–450 W TDP——范围很宽,也说明这是一族配置,而不是单一固定芯片
内存带宽这个数字最值得你盯住。智能体对内存的饥渴方式与 Web 服务器完全不同:超长上下文、持续的工具输出搬移、在推理过程中还要查询数据库。1.2 TB/s 是让上述智能体性能宣称“看起来成立”的关键规格,也是等硬件送到评测者手上后最容易被独立验证的规格。
Vera 位于 Vera Rubin 平台之中——也就是 NVIDIA 正在与 Oracle 一起出货的 NVL72 机柜系统。这也引出了同一天(11 月 5 日)的第二条消息。
同一天,NVIDIA 与 Oracle 宣布 Vera Rubin NVL72 系统以及 Vera CPU 在 Oracle Cloud Infrastructure(OCI)上全面可用,并且 Superclusters 正朝着多吉瓦(multi-gigawatt)规模扩展。
这条信息比芯片规格更重要,原因很简单:你几乎肯定不会直接购买 Vera CPU。它是面向云服务商与超大规模客户的数据中心部件,不是放工作站里的东西。你真正会购买的,是它所支撑的云端算力——智能体沙箱、托管式编程智能体、运行在 OCI 上的编排服务,以及预计后续也会出现在其他云上。对你“账单”产生影响的,是这条可用性公告。
NVIDIA 不只是发布了自家对比结果,它还发布了由合作伙伴报告的结果。下面给出完整表格,并标注每个数字的来源——因为来源本身就是重点:
用怀疑的态度读这张表——至少我是这样。Daytona 的数字最具体:10.8 秒启动 1000 个沙箱(对比为 32.3 秒)。沙箱启动是每个编程智能体在开始“有用工作”之前都要交的税,而把这里降低 3 倍,是你在真实产品中能感知到的:等待更短、每个智能体运行更便宜、同样的钱能跑更多并行智能体。DeepInfra 的数字最精确:编排步骤 29 ms,并且是在裸金属上对着具名芯片测出来的。Perplexity 的数字最含糊——“约快 1.5 倍”,但没有给出基准名称——在 Perplexity 提供更多信息之前,我会对这一条折扣最大。
也给出理解这一切的正确框架:这些都是合作伙伴报告的数字,由 NVIDIA 进行挑选,并出现在发布公告中。没有独立实验室运行过这些测试。ClickHouse 与 DuckDB 的结果最具可独立核验性——ClickBench 是公开基准,DuckDB 的数字也可复现——所以等确认时,值得重点看这两项。
真正关键的地方在这里,而且它比标题更低调,但确实存在。
如果你在云端运行智能体,你花钱买的瓶颈越来越集中在编排层,而不是模型。编程智能体的账单=模型 token + 沙箱时间 + 工具调用延迟。Vera 针对的是第二项和第三项。如果云服务商把这些效率红利传导下去——而且这是个真正的“如果”,定价是他们的决定,不是 NVIDIA 的决定——那么即使你一行代码都不改,智能体运行也会在后台变得更便宜、更快。
如果你构建智能体基础设施,需要关注的是沙箱启动时间以及每一步的编排延迟。这是 NVIDIA 表中与用户体验直接对应的两个指标:智能体多久才开始干活,以及它的步骤之间有多少无意义的“空转时间”。去问你的服务商:在你们当前硬件上,这些数字分别是多少。如果他们答不上,那是一个值得先补上的缺口——在你去关心底层用哪块 CPU 之前。
如果你只是使用编程智能体,明天什么都不会变。你的智能体不会因为一块芯片发布了就自动更快。它要更快,需要等你的服务商部署基于 Vera 的容量,并且你的账单能体现出来——这是一条漫长链路,而不是发布当天就会发生的事件。
本周就可以检查的:你的智能体账单里该看什么
你不必等 Vera 落地,也能从这次公告里获取价值。对你已经在付费使用的任意智能体环境,做一个 5 分钟检查:
沙箱启动时间。从“开始任务”到智能体真正开始工作,需要多久?如果你的服务商公布了这一数据,就和 Daytona 在 Vera 上的 10.8 秒对比。如果他们没公布,就去问。
每步编排延迟。模型的工具调用之间有多少时间差?DeepInfra 的 29 ms 是要超越的目标——但前提是你的服务商确实在测量它。
成本应看“每个 agent 小时”,而不是每 token。模型 token 是可见成本;沙箱时间与空闲编排是隐藏成本,也正是 Vera 要解决的问题。如果你只盯 token,你最多只算了这项工作的一半。
你的服务商算力到底跑在哪里。Vera 现在已经在 OCI 上全面可用。如果你的智能体平台运行在 OCI 上,问他们 Vera 容量何时进入你所在区域。如果运行在其他地方,问题就是何时会来,以及是否真的会来。
你的数据库层是不是真正的瓶颈。ClickHouse 和 DuckDB 都在 Vera 上放出了很强的数字,这说明一件事:智能体负载会非常猛地压在分析型数据库上。如果你的智能体很慢,先怀疑下面的数据库,再怀疑 CPU。
所有标题里的大数字,都来自 NVIDIA 自行挑选的对比以及合作伙伴报告的结果。目前不存在独立基准。请把每个数字都当作待验证的宣称,而不是既定事实。
Vera 不是消费级产品。你买不到它、也没法把它拿来做基准测试,或塞进你自己的机器里。它触达你的唯一方式,是通过云端算力,而且时间表与价格都由服务商决定。
“每瓦可运行智能体多 1.6 倍”这个标题是数据中心指标,并且是在 100 MW 对比中测出的。它没有说明在你笔记本上跑单个智能体的任何情况,也没有说明价格——服务商机柜里更高效的芯片,并不会自动意味着你的账单更便宜。
250–450 W 的 TDP 范围很宽。这表明存在多种配置,而我们不知道哪个配置对应哪个数字。别拿 250 W 的 Vera 去对比 450 W 的 x86 芯片,然后说这是公平比较。
本文没有任何上手实测。撰写本文没有测试过 Vera 硬件,除了 OCI 的公告之外,也没有确认任何价格、云实例类型,或区域可用性。
结论是审慎的。NVIDIA 在下注:智能体时代不只存在 GPU 问题,也存在 CPU 问题——而合作伙伴给出的数字,尤其是 Daytona 的沙箱启动时间与 DeepInfra 的编排延迟,是一类具体且可核验的宣称,足以让这个赌注看起来有意思。如果这些数字能在独立测试中站得住,你已经付费使用的智能体就会在你毫无操作的情况下更快、更便宜。如果站不住,它就只是又一张发布当天的图表。真正值得盯的,是公开可核验的指标——ClickBench、DuckDB、沙箱启动——因为这些数字没人能躲在新闻稿后面隐藏。
你的智能体运行慢,是因为模型,还是因为模型周围的一切?这其实就是 Vera 在问的问题。我很好奇你自己测出来的数字是什么样子。
本文来源:https://dev.to/paullovelace/nvidia-built-a-cpu-for-ai-agents-heres-what-changes-for-your-stack-1la9
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Posted on Nov 6 • Edited on Nov 6 • Originally published at toolsdig.com
NVIDIA has spent years selling GPUs to people who train models. On 5 November 2026 it published results for something different: a CPU built for the part of the job that happens after training — running agents. The chip is called Vera, and NVIDIA's headline claim is that in a 100 MW data-center comparison it delivers 1.6x more agents per watt than the fastest x86 CPU, and 2x faster agent speed than the most energy-efficient x86 CPU.
Those are NVIDIA's numbers, from NVIDIA's own comparisons. I haven't run them, nobody independent has reproduced them yet, and this article will keep saying so — because the numbers are still worth your attention even with the caveat attached. Agents don't bottleneck on the model call the way most people assume. They bottleneck on everything around it: spinning up sandboxes, orchestrating tool calls, running databases, shuttling context between steps. That's CPU work. And that's exactly what Vera is aimed at.
Vera is NVIDIA's first data-center CPU designed specifically for AI agent workloads. The specs, as published:
88 custom Olympus cores, 176 threads via what NVIDIA calls Spatial Multithreading
Up to 1.2 TB/s LPDDR5X memory bandwidth, up to 1.5 TB of memory
250–450 W TDP — a wide range, which tells you this is a family of configurations, not one fixed chip
The memory bandwidth number is the one to stare at. Agents are memory-hungry in a way that looks nothing like a web server: long contexts, constant tool-output shuffling, databases being queried mid-reasoning. 1.2 TB/s is the spec that makes the agent claims plausible, and it's also the spec that would be easiest to verify independently once hardware reaches reviewers.
Vera sits in the Vera Rubin platform — the NVL72 rack system NVIDIA is now shipping with Oracle. Which brings us to the second piece of 5 November news.
On the same day, NVIDIA and Oracle announced that Vera Rubin NVL72 systems and Vera CPUs are generally available on Oracle Cloud Infrastructure (OCI), with Superclusters expanding toward multi-gigawatt scale.
That matters more than the chip specs, for a simple reason: you will almost certainly never buy a Vera CPU. It is a data-center part, sold to cloud providers and hyperscalers, not something that goes in a workstation. What you will buy is the cloud capacity it powers — agent sandboxes, hosted coding agents, orchestration services running on OCI and, presumably, other clouds later. The availability announcement is the part that reaches your bill.
NVIDIA didn't just publish its own comparisons. It published partner-reported results. Here's the full set, with the source of each number, because the source is the whole story:
Read that table like a sceptic, because I am one. The Daytona number is the most concrete: 1,000 sandboxes in 10.8 seconds instead of 32.3. Sandbox spin-up is the tax every coding agent pays before it does anything useful, and a 3x cut there is the kind of thing you'd feel in a real product — shorter waits, cheaper per-agent runs, more parallel agents for the same money. The DeepInfra number is the most specific: 29 ms orchestration steps, measured on bare metal against named chips. The Perplexity number is the vaguest — "~1.5x faster" with no benchmark named — and it's the one I'd discount hardest until Perplexity says more.
And the honest frame for all of it: these are partner-reported numbers, selected by NVIDIA, in a launch announcement. No independent lab has run these tests. The ClickHouse and DuckDB results are the most independently checkable — ClickBench is a public benchmark, and DuckDB numbers are reproducible — so those are the ones to watch for confirmation.
Here's the part that matters, and it's smaller than the headlines but real.
If you run agents in the cloud , the bottleneck you pay for is increasingly the orchestration layer, not the model. A coding agent's bill is model tokens plus sandbox time plus tool-call latency. Vera attacks the second and third items. If cloud providers pass the efficiency through — and that's a real "if," pricing is their decision, not NVIDIA's — agent runs get cheaper and faster in the background, without you changing a line of code.
If you build agent infrastructure , the numbers to watch are sandbox spin-up time and orchestration latency per step. Those are the two metrics in NVIDIA's table that map directly to user experience: how long before the agent starts working, and how much dead time sits between its steps. Ask your provider what those numbers are today, on your current hardware. If they can't answer, that's a gap worth closing before you care about which CPU is underneath.

If you just use coding agents , nothing changes tomorrow. Your agent doesn't get faster because a chip launched. It gets faster when your provider deploys Vera-based capacity and your bill reflects it — which is a months-long chain, not a launch-day event.
What to check in your own agent bill — this week
You don't need to wait for Vera to get value from this announcement. Run this five-minute check on whatever agent setup you already pay for:
Sandbox spin-up time. How long from "start task" to the agent actually doing work? If your provider publishes it, compare it against Daytona's 10.8 s on Vera. If they don't publish it, ask.
Orchestration latency per step. How much time sits between the model's tool calls? DeepInfra's 29 ms is the number to beat — but only if your provider measures it at all.
Cost per agent-hour, not per token. Model tokens are the visible cost. Sandbox time and idle orchestration are the hidden ones, and they're the ones Vera attacks. If you only track tokens, you're pricing half the job.
Where your provider's capacity actually runs. Vera is generally available on OCI now. If your agent platform runs on OCI, ask when Vera capacity reaches your region. If it runs elsewhere, the question is when — and whether — it arrives there.
Whether your database layer is the real bottleneck. ClickHouse and DuckDB both posted strong Vera numbers, which tells you something: agent workloads hammer analytical databases hard. If your agent is slow, the database underneath it is a suspect before the CPU is.
All headline numbers are NVIDIA's own selected comparisons and partner-reported results. No independent benchmarks exist yet. Treat every figure as a claim to be verified, not a fact.
Vera is not a consumer product. You can't buy one, benchmark one, or put one in your machine. The only way it reaches you is through cloud capacity, on the provider's timeline and pricing.
The "1.6x more agents per watt" headline is a data-center metric measured in a 100 MW comparison. It says nothing about a single agent run on your laptop, and it says nothing about price — a more efficient chip in a provider's rack doesn't automatically mean a cheaper bill for you.
The 250–450 W TDP range is wide. That suggests multiple configurations, and we don't know which configuration produced which number. Don't compare a 250 W Vera against a 450 W x86 chip and call it fair.
Nothing here is hands-on. No Vera hardware was tested for this article, and no prices, cloud instance types, or regional availability beyond the OCI announcement have been confirmed.
The bottom line is a measured one. NVIDIA is betting that the agent era has a CPU problem, not just a GPU problem — and the partner numbers, especially Daytona's sandbox times and DeepInfra's orchestration latency, are the kind of concrete, checkable claims that make the bet interesting. If they hold up under independent testing, the agents you already pay for get faster and cheaper without you lifting a finger. If they don't, it's another launch-day chart. The numbers to watch are the public ones — ClickBench, DuckDB, sandbox spin-up — because those are the ones nobody can hide behind a press release.
Are your agent runs slow because of the model, or because of everything around it? That's the question Vera is really asking. I'd be curious what your own numbers look like.