liveBook · Manning

Return on investment

As explained in LLM Customization and Fine-Tuning

Return on investment: the economic benefit of an investment relative to its cost.

Return on investment (ROI) compares an investment’s economic benefit with its cost. For model adaptation, the decision is not just whether training is affordable: it is whether the resulting savings or business gains justify development, deployment, and ongoing maintenance against the best alternative. section 1.3 section 6.13

A business case can combine cost, privacy and data sovereignty, latency, and competitive advantage, weighed against human effort. Calculate the economic case against a baseline, and consider whether prompting, retrieval-augmented generation (RAG), or an existing model already meets the need. section 1.3 section 1.11

Compare serving costs before training

At one million requests a month, late-2026 estimates put fast-tier API calls at $1,000–$3,000, mainstream calls at $7,000–$10,000, and frontier calls at $15,000–$30,000. A continuously running enterprise GPU costs roughly $1,500–$4,400 per month. On those assumptions, self-hosting loses against the fast tier, while one GPU wins against the mainstream tier and clearly wins against the frontier tier. Check capacity and redundancy as well: a redundant GPU pair roughly breaks even against mainstream pricing. Prompt caching and batch discounts can also reduce API costs, while an idle GPU still incurs charges. section 1.3

Include the full investment and its expected shelf life:

  • Count data collection and annotation, engineering time, compute, evaluation, and serving—not just the training GPU bill. section 5.7 section 6.13
  • Budget recurring retraining and evaluation. Adapters are tied to their base weights, so a base-model refresh requires retraining and another evaluation gate. section 1.3
  • Account for adoption and opportunity cost: projected benefits require actual use, and engineering time spent on fine-tuning cannot also improve retrieval, prompting, or data pipelines. section 6.12.5 section 6.13

Value time saved in clinical lookups

A fine-tuned model used as a first-pass tool reduced average lookup time from 12 to 5 minutes. At approximately 800 queries per day, the estimate was 33,600 clinician hours saved annually, valued at $1.7 million using loaded hourly cost, with payback in under a week. Clinicians still verified and refined answers. The estimate depended on adoption, and the model needed monthly retraining and continuous safety monitoring. section 6.12.5

For supervised fine-tuning (SFT), estimate net value as quality improvement multiplied by business value per unit of quality, minus training, maintenance, and opportunity costs. Measure the improvement against the best non-SFT alternative; a gap below 5% rarely justifies SFT. A small quality gain can nevertheless have very different value across domains. section 6.13

Table 6.15 When to use SFT vs. when to skip it
Use SFT if…
Skip SFT if…
LoRA has been tried and falls measurably short of quality targets
Prompt engineering or RAG achieves acceptable quality
The task requires new reasoning patterns not present in the base model
The task primarily requires factual knowledge (use RAG instead)
You have 1,000+ high-quality, curated training examples
You have fewer than 500 examples (risk of overfitting; try LoRA or few-shot first)
Consistent output format and behavior are critical (e.g., regulatory compliance)
You can meet output format requirements with structured prompting.
Inference cost matters: a smaller fine-tuned model can replace a larger general model
You need the model to handle a wide variety of unpredictable tasks.
You have a budget and team capacity for ongoing model maintenance
The task requirements change frequently (the retraining cycle becomes burdensome)
Latency requirements demand a smaller, specialized model
A larger general-purpose model with prompting meets latency requirements
Domain-specific terminology and reasoning are essential.
The domain vocabulary is already well-represented in the base model.

Compare conditions favoring SFT with conditions favoring alternatives, including data availability, quality targets, and maintenance capacity. section 6.13

How LLM Customization and Fine-Tuning covers it

How deep it goes
Works through serving-cost and payback comparisons, then provides a net-value formula and criteria for SFT decisions.
Examples
Uses a million-request serving comparison and a clinical lookup time-savings estimate with adoption and maintenance caveats.
What you can do after
Evaluate model adaptation investments against alternatives, including recurring maintenance and opportunity costs.
How it connects
Connects ROI to supervised fine-tuning, LoRA, retrieval-augmented generation, and model evaluation.

Also explained in

Compare how 14 books explain return on investment →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Customization and Fine-Tuning. The section links point to the parts of the book it draws on.