Costs, Budgets & Images

GPU catalog

Every compute picker shows the estimated price per hour next to each option.

GPU	VRAM	Max per container	Notes
T4	16 GB	8	Light inference / exploration
L4	24 GB	8	Best price/performance for inference
A10	24 GB	4	Formerly A10G
L40S	48 GB	8	Strong inference
A100 / A100-40GB / A100-80GB	40 / 80 GB	8	A100 may be upgraded to 80GB at no extra cost
RTX-PRO-6000	96 GB	8	Blackwell workstation GPU
H100	80 GB	8	May be upgraded to H200 at no extra cost
H100!	80 GB	8	Pinned H100, never upgraded
H200	141 GB	8	HBM3e
B200	180 GB	8	Blackwell
B200+	180 GB	8	B200 or B300, whichever is free first, billed as B200
B300	288 GB	8	Blackwell Ultra

hub, dev and preset create ask how many GPUs to attach (H100:4).

Prices

m-gpux billing rates             # prices used for estimates
m-gpux billing rates --refresh   # fetch live prices from Modal (modal>=1.5.4)

Live prices are cached in ~/.m-gpux/rates.json. Without a cache, a built-in snapshot of Modal's list prices is used.

Budgets

A budget is a monthly spending limit per account. It can be below the free $30 credit (to keep a reserve) or above it (if you pay for extra usage).

m-gpux budget set 20                 # default for every account
m-gpux budget set 50 -a team-a100    # one account
m-gpux budget show                   # used / limit / left for every account
m-gpux budget check --stop           # stop m-gpux apps on accounts over budget; exit code 2 if any
m-gpux budget watch --interval 600   # keep enforcing until Ctrl+C
m-gpux budget clear -a team-a100

Budgets also change how accounts are picked:

• AUTO chooses the account with the most money left under its budget (or free credit).
• Picking an account by hand that has less than $1 left offers to switch to the best account instead (MGPUX_SKIP_CREDIT_CHECK=1 turns this off).

budget watch only enforces limits while it runs. For unattended enforcement, run m-gpux budget check --stop from cron or Task Scheduler.

Billing reports

m-gpux billing usage --all --resources   # per-app cost + breakdown by CPU / memory / GPU type
m-gpux billing summary --all             # metered vs. credits vs. billed, this month

Published images

Installing the same packages on every cold start is slow. Build once, publish, and reuse:

m-gpux image build torch --pip torch,torchvision,transformers
m-gpux image build myproj -r requirements.txt --all   # publish to every account
m-gpux image list

When the selected account has published images, m-gpux hub and m-gpux dev up offer them as the base image. Image names are per workspace, which is why --all builds on each account; m-gpux remembers where each image exists in ~/.m-gpux/images.json.