Skip to content
ANSWER · COST OF CAPACITY

What does an 8-GPU H100 server cost for a month?

Across 54 real H100 listings as of 2026-11-05, the median is $2.388/GPU/hour. Eight GPUs running flat out for a month (730 hours) costs about $13,948, or $167,374 a year. The same eight cards range from $12,463 to $19,887 a month between the 10th and 90th percentile — choosing the platform saves more than negotiating.

Quote this: An 8-GPU H100 server costs about $13,948 a month at asking prices (median $2.388/GPU/h x 8 x 730 hours). — CVIN, 2026-11-05, https://cvin.org/en/answers/h100-monthly-cost

Prices: device-level public listings on vast.ai / Clore.ai, latest observation per machine · as of 2026-11-05 · same caliber as /v1/market/models

01

Monthly bill by GPU count

配置整机时价月(730h)年(8760h)P10–P90 月区间
1× H100 SXM$2.39 $1,743$20,922 $1,558 – $2,486
2× H100 SXM$4.78 $3,487$41,844 $3,116 – $4,972
4× H100 SXM$9.55 $6,974$83,687 $6,231 – $9,944
8× H100 SXM$19.11 $13,948$167,374 $12,463 – $19,887

730 = 365 × 24,按连续占用算,不扣停机;8760 = 365 × 24。真实账单通常低于此数,因为多数团队不会跑满。

02

Platform spread

平台在架机器单卡中位8 卡月账单主要地区
vast.ai54 $2.389$13,951 Japan, JP · Taiwan, TW · Thailand, TH · California, US

同一型号跨平台的价格差,本质是「谁在卖自己的机器、谁在转租别人的机器」——转租要加价,所以同一台机器在两个平台看到两个价是正常的,本文取的都是各自平台自己的挂牌。

03

How the number is computed

步骤做法
取数拉取 H100 SXM 的全部设备级挂牌(当前 54 台机器 / 54 条有价观测)
归一整机价 ÷ 卡数,得到「单卡/小时」——不归一就会把 8 卡整机当成一张卡的 8 倍价
去重每台机器只保留最新一次观测;同一台机器反复调价不重复计入
统计取中位数(不用均值:少数 $40/h 的离群挂牌会把均值整体拉高)
外推月 = 单卡价 × 卡数 × 730;年 = × 8760

代入自己的数:月成本 = 单卡时价 × 卡数 × 730。单卡时价在 $2.134–$3.405 之间都算正常,超出这个区间要么是特殊配置,要么是被加价转租。

04

Where this flips: three caveats

① 利用率不等于 100%。上面算的是「连续跑满」。若一个月实际只用 300 小时,按需支出只有 $5,732——先量自己的利用率,再谈包月。
② 挂牌价不是成交价。大客户长期包机会有折扣,本站看不到也不猜折扣;看到的是公开挂牌这一层。
③ 不含存储与出网。模型权重、数据集、结果回传的存储与流量在多数平台单独计费,价格不含这部分。

At a similar decision point, compare H200: H200 medians $3.685/GPU/h across 63 machines, about $21,519 a month for eight.

05

Related

H100 SXM model file H200 model file A100 vs H100 Cost to train a 70B Payback calculator All model prices