--- title: "Modal将3亿美元AI智能体业务迁至英伟达Vera CPU:自测吞吐量最高翻倍" description: "Modal将把年入3亿美元的产品放到英伟达基于Arm架构的Vera CPU上运行。此前公司自测显示,在装满容器的服务器上,Vera的任务吞吐量最高达到英特尔机器的两倍。Modal给出的理由是:CPU一侧的任务没做完,昂贵的GPU只能闲置等待——这让选一颗CPU,变成了“昂贵芯片能赚多少钱”的决策。" publisher: "The Inference" section: "Data Center" published: 2026-10-21T18:07:11.682Z modified: 2026-10-21T18:07:11.682Z canonical: https://theinference.org/article/modal-moves-300-million-of-ai-agent-work-to-nvidia-s-vera-cpu-after-its-own-tests-showed-up-to-2x-throughput/zh-CN language: zh-CN keywords: "Modal, Nvidia, AI Agents, Cloud Computing, Semiconductors" --- # Modal将3亿美元AI智能体业务迁至英伟达Vera CPU:自测吞吐量最高翻倍 > Modal将把年入3亿美元的产品放到英伟达基于Arm架构的Vera CPU上运行。此前公司自测显示,在装满容器的服务器上,Vera的任务吞吐量最高达到英特尔机器的两倍。Modal给出的理由是:CPU一侧的任务没做完,昂贵的GPU只能闲置等待——这让选一颗CPU,变成了“昂贵芯片能赚多少钱”的决策。 [Modal](https://theinference.org/markets/companies/modal)正把智能体沙箱迁往[英伟达](https://theinference.org/markets/companies/nvidia)基于[Arm](https://theinference.org/markets/companies/arm)架构的Vera CPU,这是公司首次离开一直使用的[英特尔](https://theinference.org/markets/companies/intel)和[AMD](https://theinference.org/markets/companies/amd)芯片。此前,[自测基准显示,在装满容器的服务器上,任务吞吐量最高翻倍](https://thenextweb.com/news/modal-nvidia-vera-cpu-sandboxes)[1]。 Modal约三分之一的收入系于这一层:沙箱是供AI智能体编写、运行代码的隔离计算环境,已带来3亿美元年化收入——Modal自称[年化收入已达10亿美元](https://cloudnews.tech/modal-tops-1-billion-in-annualized-revenue-raises-750m/)——累计创建的沙箱接近30亿个[1][2]。Modal给出的理由关乎利用率:在强化学习中,若CPU一侧的任务尚未完成,昂贵的GPU往往只能闲置等待[1]。 ## CPU为何决定GPU能赚多少 Modal表示,多数训练流程仍是同步的:模型每一轮rollout(即对一项任务的完整尝试,其结果将转化为训练信号)都有一部分时间花在CPU上;整批rollout必须全部跑完,闲置下来的GPU才能进入下一步,所以在CPU上省下的时间,就是在GPU上省下的时间[3]。就这样,便宜的处理器反而决定了昂贵处理器的利用率。更快的CPU,实际上等于买到了GPU时长。 ## 智能体负载波动剧烈,Vera的带宽足以消化 沙箱租户并不安生:智能体先是闲着等待语言模型,转头又在安装软件时调用10倍于平时的算力;数百个沙箱共用一台宿主机,一家客户的用量尖峰就可能拖慢另一家的任务[1]。 在Modal覆盖50项编程任务的基准测试中,逐项单测时Vera都胜过英特尔、AMD和[亚马逊](https://theinference.org/markets/companies/amazon)的芯片,但只是小胜;差距要等负载上来才拉开——容器越堆越多,Vera保持节奏,其他芯片先后撞上天花板,Modal把这归功于Vera的内存带宽[1][3]。结果出自Modal自己之手,尚未经独立验证[1]。 ## 溢价正从隔离转向训练循环 10月14日,Modal按157.5亿美元估值完成7.5亿美元融资,称这笔钱买的是更高的层级,强化学习和后训练(post-training)工具都在其中[2]。竞争对手、云厂商[Together AI](https://theinference.org/markets/companies/together-ai)[在10月19日开放的正是这一层](https://www.together.ai/blog/together-post-training):预留B200 GPU,训练循环由客户定义,入门价每GPU小时9美元[4][5]。 再往下一层正在商品化:[Perplexity已将生产环境沙箱SPACE开源](https://forkast.news/the-sandbox-layer-is-commoditizing-modal-just-raised-at-15-75x-revenue-anyway/),[CoreWeave](https://theinference.org/markets/companies/coreweave)、[Akamai](https://theinference.org/markets/companies/akamai)和[Crusoe](https://theinference.org/markets/companies/crusoe)也已提供同类服务[6][8]。各家智能体公司为这份隔离买单,付给Modal的每虚拟CPU小时费用仍比E2B或Daytona高出约41%[6]。 这大约是Modal自家普通算力费率的三倍[6]。Modal的容量租自二十多家云厂商,名下没有一座数据中心,因此一颗让部署密度翻倍的芯片,就能抬高每台租用服务器的收入[2]。 ## 英特尔与AMD承担插槽风险 英伟达3月推出Vera,定位为面向智能体的CPU,内置88核,[Anthropic](https://theinference.org/markets/companies/anthropic)和[OpenAI](https://theinference.org/markets/companies/openai)位列首批用户[1]。计划部署的云厂商中,就有阿里云、[字节跳动](https://theinference.org/markets/companies/bytedance)、CoreWeave、[Nebius](https://theinference.org/markets/companies/nebius)和[甲骨文](https://theinference.org/markets/companies/oracle)[3]。这份好处,英伟达赚两次:一次在CPU上,一次在因它而保持忙碌的GPU上;而在智能体运行代码的这唯一一层,[英特尔和AMD正失去默认占据的插槽](https://theinference.org/article/amd-raises-its-cpu-forecast-a-third-time-to-200-billion-nvidia-already-ships-the-rival-rack)。 ## 投产之前,“翻倍”只是一种说法 客户何时迁上Vera,Modal并未给出时间点;英伟达则表示,[合作伙伴渠道的供应要等到2026年下半年](https://www.unite.ai/modal-plans-nvidia-vera-rollout-for-sandboxes-in-first-arm-move/)[1][3]。Modal自己也留了后手:英特尔和AMD的芯片继续留在轮换序列里,每个任务都会调度到跑得最好的那款芯片上[7]。最终检验这一判断的,一是[装满容器的服务器带来的性能提升,能否在Modal自家基准测试之外复现](https://theinference.org/article/cognition-s-4-8x-benchmark-proves-throughput-not-cheaper-agent-work);二是41%的溢价,在已然免费的隔离层面前还能否保住[6]。 ## Takeaway CPU选型如今是一道GPU经济学题:如果Vera真能让满载服务器的吞吐量翻倍,Modal租用的服务器收入大致也随之翻倍;但这一成绩出自Modal自己的基准测试,而下方的沙箱层正在走向商品化。 ## Sources 1. [The Next Web, "Modal says Nvidia's Vera CPU doubles throughput for its AI sandboxes", 19 October 2026](https://thenextweb.com/news/modal-nvidia-vera-cpu-sandboxes) 2. [Cloudnews, "Modal Tops $1 Billion in Annualized Revenue, Raises $750M", 21 October 2026](https://cloudnews.tech/modal-tops-1-billion-in-annualized-revenue-raises-750m/) 3. [Unite.AI, "Modal Plans NVIDIA Vera Rollout for Sandboxes in First Arm Move", 19 October 2026](https://www.unite.ai/modal-plans-nvidia-vera-rollout-for-sandboxes-in-first-arm-move/) 4. [Together AI, "Together Post-Training: frontier-grade model shaping, the way that fits you", 19 October 2026](https://www.together.ai/blog/together-post-training) 5. [TokenPost, "Together AI Launches Post-Training Service With Reserved GPU Capacity", 21 October 2026](https://www.tokenpost.com/news/technology/35231) 6. [AInvest, "Modal Raised at 15.75x Revenue. Its Compute Is Rented, and Its Sandbox Is a Commodity", 19 October 2026](https://www.ainvest.com/news/modal-raised-15-75x-revenue-compute-rented-sandbox-commodity-2610/) 7. [AI Weekly, "Modal Adds Nvidia Vera CPU to Host AI Agents, Claims 2x Throughput", 20 October 2026](https://aiweekly.co/alerts/modal-adds-nvidia-vera-cpu-to-host-ai-agents-claims-2x-throughput) 8. [Forkast News, "The Sandbox Layer Is Commoditizing. Modal Just Raised at 15.75x Revenue Anyway.", 19 October 2026](https://forkast.news/the-sandbox-layer-is-commoditizing-modal-just-raised-at-15-75x-revenue-anyway/) --- Modal将3亿美元AI智能体业务迁至英伟达Vera CPU:自测吞吐量最高翻倍 — The Inference. Canonical: https://theinference.org/article/modal-moves-300-million-of-ai-agent-work-to-nvidia-s-vera-cpu-after-its-own-tests-showed-up-to-2x-throughput/zh-CN