Magazine/Lower GPU cost/GUIDE · 5 STEPS

Lower your GPU bill: find the waste before changing the rate

Start with the cost of one completed job.

Remove avoidable runtime and idle time, then compare billing options and the full cost of moving.

Step 01 / 05 · Measure

Measure what one completed job costs

If three attempts were needed to produce one accepted result, all three belong in its cost.

  • Choose a useful unit: a training run, a batch or an inference task.
  • Record enough information to connect the result to the bill. Include failed attempts and repeated work.
Before moving on: be able to explain the bill for one completed job.
Exhibit 2Five records tie one completed job to its bill, beyond the advertised GPU rateone completed job
RecordWhy it matters
Code, data and configurationLets you repeat the same job.
GPU type and number of GPUsMakes hardware comparisons meaningful.
Start, finish and release timesSeparates useful runtime from time left allocated.
Result and acceptance checkShows whether the job delivered usable work.
Compute, storage and transfer chargesCaptures costs beyond the advertised GPU rate.
Source Lyceum
Step 02 / 05 · Remove waste

Remove time you do not need to pay for

Check what happens after the job finishes first: in the example, 40 of 60 allocated minutes come after the work.

  • On our VMs, termination wipes the machine’s disk; your account’s storage bucket persists. Save the output there, verify that it is readable, then terminate the VM. See our storage documentation.
  • If the job itself is slow, inspect its execution timeline. The GPU may be waiting for data, communication or another operation. A low utilisation reading does not identify the cause on its own.
  • Use a profiler to find where time goes. Change one suspected bottleneck, rerun the job and check both elapsed time and output quality.
Before moving on: have a tested shutdown path and evidence for any runtime change you plan to make.
Exhibit 3The longer the GPU stays allocated after the job, the larger the share you pay for idle timeallocated minutes for a 20-minute job on one GPU · % of allocated time left idle
jobleft allocated after the job
0 min left
20 min0%
10 min left
30 min33.3%
20 min left
40 min50%
40 min left
60 min66.7%
60 min left
80 min75%
Source Hypothetical timing at an unchanged rate; the page’s own example at each setting. Excludes startup, storage, transfers and minimum commitments. Save and verify output before terminating a machine.
Step 03 / 05 · Compare

Match the billing option to the workload

Per-second billing still charges for a machine left allocated after its work finishes.

  • Ask exactly what starts and stops the meter: it can run through setup, useful work and idle time.
  • Our compute options include GPU VMs and container-based training. Choose using your workload history, including quiet periods and failed runs.
  • A reserved machine may suit steady demand. An occasional experiment needs a different comparison. Measure the work you expect to complete across the whole billing period.
Before moving on: know the total commitment, interruption behaviour and expected cost per completed job for each option.
Exhibit 4Each billing option has its own check that decides the cost per completed job
OptionWhat to check
On-demand VMHow long the machine stays allocated, including setup and idle time.
Spot capacityWhether the job can resume after interruption, and how much work it repeats.
Reserved capacityThe full commitment and how much useful work you expect to complete.
Serverless trainingThe run’s resource requirements, billing terms and output storage.
Source Lyceum
Step 04 / 05 · Compare

Compare the full bill over the same period

Quotes compare only for the same workload, period and amount of completed work, priced without credits.

  • Check whether each quote covers one GPU or a whole machine. Compare memory, GPU count, storage, networking and billing terms alongside the price.
  • Forecast the first month after credits expire. Confirm which charges the credits cover and price the workload without that subsidy.
  • Keep data transfer separate from GPU runtime. Egress is data leaving a service or network. Record each route, its volume and frequency, then check the applicable charges.
  • Removing 40 idle minutes does not tell you what moving a large dataset will cost.
Before moving on: have comparable estimates for the same workload, period and amount of completed work.
Exhibit 5Owning and renting both carry costs beyond the quoted pricesame workload · same time period
ownership
  • Hardware
  • Power and cooling
  • Space
  • Maintenance
  • Staffing
renting
  • Capacity commitment
  • Supporting services
  • Work your team still operates
+ data transfer · + credits expire
Source Lyceum
Step 05 / 05 · Compare

Test one workload before moving the rest

A container packages the software environment, not the storage, permissions, logging and recovery around it.

  • Pick a small, representative job for a migration trial. Record its container version, launch command and configuration, data access, required secrets and output destination.
  • Run the same job in both environments and confirm the result passes the same acceptance check. Then compare elapsed and billed time; compute, storage and transfer charges; setup and ongoing team work; recovery from an interrupted run.
  • Our serverless training accepts a plain Docker container. Test that container and its surrounding workflow before moving a larger workload.
Before expanding the move: confirm the trial produces the required result and the full cost supports the decision.
Exhibit 6Both environments must pass the same acceptance check before their costs are compared
containerone representative jobpackages the software environment
environment A · Bstorage · permissions · logging · recovery
✓ checksame acceptance check
Source Lyceum

Your next workload starts here