Skip to content
Bitpute
ToolsModelsGPUsCloudLearn

Training ROI

When does the
card pay for itself?

If you fine-tune on your own hardware, every run is “free” once the GPU is bought. We estimate the hours a run takes, price the electricity and the cloud alternative, and tell you how many runs it takes to break even.

Real run hours

6·N·D compute, ~40% MFU — editable if you know yours.

Per-run cost

Electricity owned vs the same run rented on cloud.

Break-even

The run count where owning overtakes renting.

Training ROI modellive

Hours / run
estimated
Electricity / run
owned, at the wall
Cloud cost / run
same run, rented
Break-even
runs to recoup card
Saving per run vs cloud
cloud cost − electricity, once the card is paid off

Own vs rent, run by run

The recoup curve.

Owning starts at the sticker price and inches up by the electricity per run. Renting starts at zero and climbs by the cloud cost each run. Where they cross, the card has paid for itself.

Cumulative cost
Own (price + power/run) Rent (cloud/run) Break-even

How we estimate a run

Assumptions on the table.

Compute

6 × params × tokens FLOPs — the standard forward + backward cost for a dense transformer step.

Time

FLOPs ÷ (peak BF16 × MFU), MFU ≈ 40%. Peak TFLOPS are indicative — real throughput shifts with batch, sequence length and framework, so the field is editable.

Money

Electricity TDP × 1.3 × hrs × rate; cloud $/hr × hrs. Break-even = price ÷ (cloud − power) per run.

Size the training VRAM →

Before you commit the capital

Buy the card that earns it back.

ROI only matters if the card fits the job. The Workspace ranks GPUs for your model, then lands on buy-vs-rent — this page proves the payback.

Evidence & method

How this calculation works

Estimates a fine-tuning run's hours, electricity and cloud cost, then shows how many runs it takes for owning the GPU to beat renting it.

Data sources

  • NVIDIA, AMD & Intel GPU documentation (TDP, price class)
  • Cloud provider pricing
  • Bitpute Methodology

Assumptions

  • Energy = TDP × system-overhead factor × hours/day × 30.4 × months × rate
  • Default utilisation 8 hours/day; electricity rate and hours are adjustable
  • Purchase price amortised over the period; cloud-rent equivalent at the provider hourly rate
  • Excludes cooling, networking, maintenance and datacentre cost beyond the TDP overhead factor

Limitations

  • Electricity rates, utilisation and cloud prices vary widely by region, provider and contract
  • Real power draw depends on workload, not just rated TDP
  • Prices are indicative and change frequently
  • Estimates are for comparison, not financial advice

Data status: Hardware specs — vendor documentation. Prices & rates — indicative and user-adjustable.

Why does this estimate differ from other calculators?
  • Different electricity rates and utilisation assumptions
  • Cloud pricing variance by provider and region
  • Power measured at the wall vs at the card
  • Amortisation period and residual value
  • Rounding