The Economics of Model Training
Lesson 4 of 4Intermediate 9 min

Rent, buy or don't train at all

Most teams should rent, and many shouldn't train at all. Here's how to work out which applies to you.

In this lesson you'll

  • Calculate a break-even point for buying a GPU
  • List the costs a rental or purchase price leaves out
  • Choose the cheapest approach that meets your goal

The first question isn't where to run training. It's whether you need to. Each step below costs far more than the one before, so move on only when evaluation shows that the current step has hit its limit.

ApproachTypical costUse it when
Prompt tuningAPI callsThe model can do the task but needs better instructions
Model searchAPI callsYou haven't tested which model suits the task
LoRA or QLoRATens to hundreds of dollars per runYou need a consistent format, style or lower cost at scale
Full fine-tuningThousands of dollars and upLoRA has plateaued and you have a lot of good data
PretrainingHundreds of dollars for a small research model, millions for a competitive oneYou're doing research or working in a domain open models don't cover

Those costs are our rough guide. As one reference point, QLoRA's best models took about 24 hours of fine-tuning on a single GPU [1], which is around $100 at H100 rates. Budget for several runs, because the first rarely wins.

The break-even calculation

Buying only pays off if the machine stays busy. GPU prices also moved sharply in 2026. Tom's Hardware's tracker showed a best US price of $6,919 for an RTX 5090 on 3 November, against a $1,999 MSRP [2]. So a complete workstation lands near $10,000 by our estimate.

Example: a $10,000 RTX 5090 workstation vs renting the same GPU

rental    = $1.19 per hour
power     = 0.8 kW × $0.182 per kWh = $0.15 per hour
saving    = $1.19 − $0.15           = $1.04 per hour
breakeven = $10,000 / $1.04         = about 9,600 hours

That's about 13 months running nonstop,
or about 2.2 years at 50% utilisation.

This example uses Runpod's Secure Cloud rate [3], an assumed 800 W system draw and the August 2026 US residential average of about 18.2 cents per kWh [4]. Your rate may differ, and resale value could shorten the payback. Cheaper Community Cloud pricing would stretch it.

What the sticker price leaves out

  • Failed and repeated runs: hyperparameter sweeps and bugs can cost more than the final run.
  • Idle time: rented GPUs bill while you debug, and owned GPUs lose value while they sit unused.
  • Data: collecting, cleaning and licensing it takes people and time.
  • Staff: R&D staff made up 29–49% of frontier development costs in Epoch AI's estimate [5].
  • Serving: a trained model still needs hosting, and heavy traffic can outweigh training cost over the model's life [6].

When buying makes sense

  • Your GPUs will stay busy for a year or more.
  • Your data can't leave your own machines.
  • You run lots of small experiments, and fast local iteration matters.

Rule of thumb

Rent until you have months of measured, steady usage. Rented capacity lets you try newer GPUs and scale up for a single run, and you can stop paying the day a project ends.

Key takeaways

  • Start with the cheapest approach, and let evaluations justify each step up.
  • A $10,000 workstation takes about 13 months of nonstop use to beat renting the same GPU.
  • People, data, failed runs and serving often cost more than the final run.

Practise in the dojo

Compare your project's model versions side by side. If a prompt-tuned version already meets your target, you've just avoided a training bill.

Open Compare

Check your understanding

Score 70 or more to complete the lesson. Progress is saved in this browser.

1. What decides whether buying a GPU beats renting one?
2. A team hasn't tested other models and wants better accuracy. What should it try first?

Further reading

Sources

Checked against the original sources in November 2026.

  1. [1]Dettmers, T. et al. (2023). QLoRA: Efficient Finetuning of Quantized LLMs. NeurIPS 2023 · arXiv:2305.14314.
  2. [2]Tom's Hardware (2026). GPU price tracking 2026 (updated 3 November 2026). tomshardware.com.
  3. [3]Runpod (2026). GPU pricing (updated 4 November 2026). runpod.io.
  4. [4]U.S. Energy Information Administration (2026). Electricity Monthly Update: end-use prices (August 2026 data). eia.gov.
  5. [5]Cottier, B. et al. (2024). The rising costs of training frontier AI models. arXiv:2405.21015.
  6. [6]Sardana, N. et al. (2023). Beyond Chinchilla-Optimal: Accounting for Inference in Language Model Scaling Laws. ICML 2024 · arXiv:2401.00448.