英伟达 Vera CPU 深度解读 2026：88 核 Olympus 自研处理器挑战英特尔 AMD
Nvidia Vera CPU Explained 2026: 88-Core Olympus Chip Takes on Intel and AMD
> 📌 TL;DR
> 2026 年 5 月下旬，英伟达副总裁 Ian Buck 亲手把几颗芯片送进了 Anthropic、OpenAI、SpaceX 和甲骨文的大门——不是 GPU，而是 Vera，英伟达第一颗自己设计的数据中心 CPU。88 个自研 Olympus 核心、1.2 TB/s 内存带宽，号称比顶级 x86 快约 80%。这意味着英伟达不再只卖「算力」，而是要把 AI 数据中心的整块地皮——CPU、GPU、网络——全包了。英特尔和 AMD 最赚钱的数据中心 CPU 生意，第一次被一个 GPU 公司正面踩进门。
一颗被「亲手送货」的芯片
5 月下旬的一个周五，英伟达负责超大规模与高性能计算的副总裁 Ian Buck，亲自把第一批 Vera CPU 送到了三家公司：旧金山的 Anthropic、Mission Bay 的 OpenAI、帕罗奥图的 SpaceX；紧接着的周一，又送到了圣克拉拉的甲骨文云。英伟达在 6 月 1 日正式对外确认了这份「首发客户名单」（来源：Bloomberg，2026-06-01）。
CEO 团队级别的待客之道背后，是一件英伟达过去从没做过的事：卖一颗独立的 CPU。
过去几年，英伟达确实有一颗叫 Grace 的 CPU，但它从来不能单买——你只能买「Grace + Hopper」「Grace + Blackwell」这种 CPU+GPU 捆绑包，CPU 只是 GPU 的「保姆」。Vera 不一样：它是英伟达第一颗真正意义上的、可独立部署的自研数据中心 CPU，第一次和英特尔的 Xeon、AMD 的 EPYC 站在同一个货架上正面竞争。
Vera 到底是什么？
技术规格上，Vera 是一颗不折不扣的猛兽（来源：Tom's Hardware、NVIDIA 开发者博客，基于 GTC 2026 公布信息）：
| 维度 | Vera CPU | 上一代 Grace |
|------|----------|------------|
| 核心 | 88 个自研 Olympus 核 | 72 个 Arm Neoverse-V2 |
| 线程 | 176（空间多线程） | 72 |
| 指令集 | Armv9.2，支持 FP8 | Armv9 |
| L2 缓存 | 2 MB/核 | 1 MB/核 |
| L3 缓存 | 164 MB 统一 | 117 MB |
| 内存带宽 | 1.2 TB/s（LPDDR5X） | ~1 TB/s |
| 连接 | PCIe Gen6 / CXL 3.1 | PCIe Gen5 |
最关键的变化是「Olympus」——这是英伟达第一次彻底抛开 Arm 的公版核心、自己从头设计 CPU 核。Grace 用的是 Arm 现成的 Neoverse-V2，而 Olympus 是英伟达完全自研的架构：10 路取指译码前端、能在一个周期里预测两个跳转的神经分支预测器。在英伟达自己的基准测试里，Vera 比当下最强的 x86 CPU 快约 80%；在塞满 256 颗芯片的整机柜里，能给出最高 6 倍的 CPU 吞吐提升（来源：Tom's Hardware）。
> ⚠️ 别被「CPU」两个字骗了
> 这不是普通的服务器 CPU。Olympus 核心专门为「控制密集、延迟敏感」的工作负载设计——而这恰好就是 AI agent 在做的事：大量的分支判断、工具调用、任务编排。换句话说，Vera 是一颗为「智能体时代」量身定做的 CPU。
为什么是现在？因为 AI 变「agent」了
要理解英伟达为什么突然认真做 CPU，得先看清一件事：这一代 AI 的瓶颈，正在从 GPU 往 CPU 转移。
过去训练大模型，几乎所有重活都压在 GPU 上，CPU 只要把数据喂进去就行。但 2026 年的主角是 agentic AI——会自己做决策、调工具、跑多步任务、几乎不需要人盯着的智能体。这类工作有大量「控制流」：判断、循环、调度、读写状态、调用外部 API。这些活恰恰是 GPU 不擅长、CPU 才擅长的。
于是问题来了：当你的几十颗 Rubin GPU 在全力推理，旁边那颗 CPU 如果跟不上节奏、喂不动数据、调度不过来，整个机柜就会被这颗「保姆」拖死。英伟达的答案是：那我自己来做这颗保姆，而且让它强到能反客为主。
在英伟达的旗舰机柜 Vera Rubin NVL72 里，36 颗 Vera CPU 配 72 颗 Rubin GPU——每颗 Vera 通过第二代 NVLink-C2C 直连两颗 Rubin GPU，共享统一内存，CPU 和 GPU 之间不再有「翻译官」式的损耗。这是英伟达卖了十年 GPU 之后，第一次真正把「整台 AI 超算」攥在自己一家手里。
对英特尔和 AMD 意味着什么
数据中心 CPU 是英特尔和 AMD 最后的、也是最赚钱的堡垒之一。Xeon 和 EPYC 长期瓜分这块市场，利润率极高。Vera 的出现，等于英伟达带着自己庞大的 AI 客户群，直接踩进了这扇门。
更要命的是捆绑效应：当一家公司已经买英伟达的 GPU 买到停不下来，英伟达再顺手说一句「CPU 也用我的吧，和 GPU 配合最好、延迟最低」，客户几乎没有理由拒绝。甲骨文已经表态，要从 2026 年起部署数十万颗 Vera。这对靠卖 CPU 给云厂商吃饭的英特尔、AMD 来说，是结构性的威胁。
当然也别急着唱衰 x86。Vera 是 Arm 架构，迁移有成本；英特尔、AMD 的软件生态、x86 兼容性、价格弹性依然是护城河。Vera 全面量产要到 2026 年第三季度，真正的市场效果还得再等几个季度才看得清。
> ✨ 一句话看懂
> 英伟达卖 GPU 的时候，英特尔和 AMD 还能说「CPU 是我们的地盘」。Vera 落地的那一刻，这句话失效了——AI 数据中心的最后一块拼图，英伟达决定自己拿。
---
本文关键数据来自 Bloomberg、Tom's Hardware、NVIDIA 官方博客与 Phoronix 的公开报道（2026 年 5–6 月），技术规格以英伟达 GTC 2026 及官方开发者博客公布的信息为准。
> 📌 TL;DR
> In late May 2026, Nvidia VP Ian Buck personally hand-delivered a handful of chips to Anthropic, OpenAI, SpaceX and Oracle — and they weren't GPUs. They were Vera, Nvidia's first in-house data-center CPU. With 88 custom Olympus cores, 1.2 TB/s of memory bandwidth and a claimed ~80% edge over top x86 chips, Vera signals that Nvidia no longer just sells "compute" — it wants the entire AI data center: CPU, GPU and networking. For the first time, the most profitable corner of Intel's and AMD's business is being invaded head-on by a GPU company.
A chip delivered by hand
On a Friday in late May 2026, Ian Buck — Nvidia's VP of Hyperscale and High-Performance Computing — personally carried the first Vera CPUs into three companies: Anthropic in San Francisco, OpenAI in Mission Bay, and SpaceX in Palo Alto. The following Monday, a batch went to Oracle Cloud in Santa Clara. Nvidia formally confirmed this "launch customer list" on June 1 (source: Bloomberg, 2026-06-01).
Behind the white-glove treatment is something Nvidia has never done before: sell a standalone CPU.
For the past few years, Nvidia did have a CPU called Grace — but you could never buy it on its own. You bought "Grace + Hopper" or "Grace + Blackwell" CPU+GPU bundles, where the CPU was just a nanny for the GPU. Vera is different. It is Nvidia's first genuinely standalone, in-house data-center CPU, and the first to sit on the same shelf as Intel's Xeon and AMD's EPYC, competing head-to-head.
What exactly is Vera?
On paper, Vera is a beast (source: Tom's Hardware, NVIDIA Developer Blog, based on GTC 2026 disclosures):
| Dimension | Vera CPU | Previous-gen Grace |
|-----------|----------|--------------------|
| Cores | 88 custom Olympus cores | 72 Arm Neoverse-V2 |
| Threads | 176 (spatial multithreading) | 72 |
| ISA | Armv9.2, FP8 support | Armv9 |
| L2 cache | 2 MB/core | 1 MB/core |
| L3 cache | 164 MB unified | 117 MB |
| Memory bandwidth | 1.2 TB/s (LPDDR5X) | ~1 TB/s |
| Connectivity | PCIe Gen6 / CXL 3.1 | PCIe Gen5 |
The pivotal change is "Olympus" — the first time Nvidia has fully abandoned Arm's off-the-shelf cores and designed a CPU core from scratch. Grace used Arm's stock Neoverse-V2; Olympus is Nvidia's own architecture, with a 10-wide fetch-and-decode frontend and a neural branch predictor capable of evaluating two taken branches per cycle. In Nvidia's own benchmarks, Vera runs roughly 80% faster than the strongest current x86 CPUs, and a full rack packed with 256 chips delivers up to a 6x gain in CPU throughput (source: Tom's Hardware).
> ⚠️ Don't be fooled by the word "CPU"
> This is not an ordinary server CPU. The Olympus cores are purpose-built for "control-heavy, latency-sensitive" workloads — which is exactly what AI agents do: endless branching, tool calls, task orchestration. In other words, Vera is a CPU tailored for the age of agents.
Why now? Because AI became "agentic"
To understand why Nvidia suddenly got serious about CPUs, you have to see one thing clearly: the bottleneck of this AI generation is shifting from the GPU to the CPU.
In the old world of model training, almost all the heavy lifting fell on the GPU, and the CPU just had to feed it data. But the protagonist of 2026 is agentic AI — agents that make their own decisions, call tools, run multi-step tasks, and need almost no human babysitting. This kind of work is full of "control flow": conditionals, loops, scheduling, reading and writing state, calling external APIs. These are precisely the tasks GPUs are bad at and CPUs excel at.
So here's the problem: when dozens of Rubin GPUs are running inference at full tilt, if the CPU next to them can't keep pace — can't feed the data, can't schedule fast enough — that "nanny" chip drags the whole rack down. Nvidia's answer: I'll build the nanny myself, and make it strong enough to take over the house.
In Nvidia's flagship rack, the Vera Rubin NVL72, 36 Vera CPUs pair with 72 Rubin GPUs — each Vera connects directly to two Rubin GPUs over second-generation NVLink-C2C, sharing a unified memory pool with no "translator" overhead between CPU and GPU. After a decade of selling GPUs, this is the first time Nvidia truly holds an entire AI supercomputer in one company's hands.
What it means for Intel and AMD
Data-center CPUs are one of the last — and most profitable — fortresses Intel and AMD have left. Xeon and EPYC have long split that high-margin market between them. Vera's arrival means Nvidia, dragging its enormous AI customer base along, just kicked open that door.
The deadlier part is the bundling effect. When a company is already buying Nvidia GPUs and can't stop, and Nvidia casually adds, "use my CPU too — it pairs best with the GPU, lowest latency," the customer has almost no reason to say no. Oracle has already signaled plans to deploy hundreds of thousands of Vera CPUs starting in 2026. For Intel and AMD, who live on selling CPUs to cloud providers, that is a structural threat.
That said, don't write off x86 just yet. Vera is Arm-based, and migration carries real costs; Intel and AMD's software ecosystems, x86 compatibility, and pricing flexibility remain genuine moats. Vera doesn't reach full production until Q3 2026, and the real market impact will take several more quarters to come into focus.
> ✨ The one-line takeaway
> While Nvidia only sold GPUs, Intel and AMD could still say "CPUs are our turf." The moment Vera shipped, that sentence stopped being true — the last piece of the AI data center puzzle, Nvidia has decided to take for itself.
---
Key figures in this article are drawn from public reporting by Bloomberg, Tom's Hardware, the official NVIDIA blog, and Phoronix (May–June 2026). Technical specifications follow the information disclosed at Nvidia GTC 2026 and on the official developer blog.