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。

阅读原文

Richard TangRichard Tang · 3 min read
Share
Modal首席执行官埃里克·伯恩哈德松在公司Runtime大会上登台
1 / 6Slide 1 of 6
Modal首席执行官埃里克·伯恩哈德松(Erik Bernhardsson)在公司Runtime大会上登台演讲。该公司将把年收入3亿美元的智能体沙箱业务放到英伟达Vera CPU上运行。

Modal$87M — Modal, private, latest valuation $87M正把智能体沙箱迁往英伟达+1.92% — 英伟达, up 1.92 percent today基于Arm+0.88% — Arm, up 0.88 percent today架构的Vera CPU,这是公司首次离开一直使用的英特尔−0.21% — 英特尔, down 0.21 percent today和AMD+2.23% — AMD, up 2.23 percent today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——累计创建的沙箱接近30亿个12。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−0.32% — 亚马逊, down 0.32 percent today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13。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~$3.3B — Together AI, private, latest valuation ~$3.3B在10月19日开放的正是这一层:预留B200 GPU,训练循环由客户定义,入门价每GPU小时9美元45。

再往下一层正在商品化:Perplexity已将生产环境沙箱SPACE开源,CoreWeave−2.35% — CoreWeave, down 2.35 percent today、Akamai+0.69% — Akamai, up 0.69 percent today和Crusoe$26.61B — Crusoe, private, latest valuation $26.61B也已提供同类服务68。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$350B — Anthropic, private, latest valuation $350B和OpenAI$1.18T — OpenAI, private, latest valuation $1.18T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、CoreWeave、Nebius−2.06% — Nebius, down 2.06 percent today和甲骨文−2.15% — 甲骨文, down 2.15 percent today3。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。

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13。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;二是41%的溢价,在已然免费的隔离层面前还能否保住6。

Get Data Center by email, free

Who is building AI's data centers, and who is paying for them. In our daily email, every weekday.

Loading ARM…

Reader comments

0 comments

    More in Data Center

    All in this topic →

    Data Center

    Goldman Sachs and JPMorgan Chase had to lift Volta's $5 billion AI loan yield past 11% to sell it

    The banks had to lift an unrated $5 billion AI loan's yield to roughly 11.2 to 11.75 percent and add protections before buyers would take it, while Goldman's own credit provision tripled the same quarter.

    Data Center

    AI fiber prices may run 5 times 2025 levels, but Corning's CEO says price is not the profit story

    TrendForce says AI fiber prices will average more than five times 2025 levels on a shortage stretching to 2028, and Corning reports in one week. But CEO Wendell Weeks says price is not how the company drives profit, so the 61.7 P/E is paying for a story the seller himself discounts.

    Data Center

    Iron Mountain's survey says 86% of AI projects stalled; it signed 110 megawatts anyway

    The company's own survey says 86% of AI projects stall on data debt, while its data-center arm books 110 megawatts of leases on borrowed money. On November 5 the market finds out which half of Iron Mountain the numbers pay.

    Data Center

    Oracle's filing shows $288 billion off-balance-sheet leases as S&P sees no disconnect and CDS hits 295

    The rating agency says it cannot see all the AI debt and sees no need to act. Oracle's quarterly filing puts a number on the unseen part, and the market's price for Oracle default keeps setting records.

    Data Center

    SpaceX posts its first profit, $4.4 billion, as AI compute revenue passes Starlink and launch combined

    The rocket company's first profit since its June listing came from the AI segment, now bigger than Starlink and launch combined. Whether Musk's $500 billion to $1 trillion forecast for 2027 arrives turns on how fast SpaceX can build, and on lease terms that let tenants walk.

    Data Center

    Nvidia picked Schneider Electric to build 14-day AI data centers as Barclays warns peak pricing may end

    Nvidia named Schneider a factory builder of its 14-day DSX POD data centers on October 21, the same day Schneider launched its RapidScale design for Vera Rubin, while Barclays warned the day before that peak pricing for power-gear makers is ending. Volume and price now point opposite ways on the same company.

    Get the newsletter

    The day's AI and market news, in your inbox every weekday.