How Much Does AI Fine-Tuning Cost?

Self-Hosted GPU Costs

Self-hosted fine-tuning means running the training on hardware you own or rent by the hour. With QLoRA, the GPU requirements are surprisingly modest. An 8B model fine-tunes on any NVIDIA GPU with 24GB or more VRAM, which includes the consumer RTX 3090, RTX 4090, and the professional A5000.

Consumer hardware (RTX 4090, $1,600-2,000): A QLoRA fine-tune of Llama 3.1 8B on 5,000 examples takes 4-8 hours. Electricity cost at $0.15/kWh with the card drawing 450W is roughly $0.30-0.55 per training run. If you already own the card for gaming or other GPU work, fine-tuning is essentially free. The card pays for itself after a few dozen training runs compared to cloud GPU pricing.

Cloud GPU (A100 80GB, $1.50-3.00/hour): The same training run on a cloud A100 takes 2-4 hours because of faster memory bandwidth and higher compute throughput. Total cost: $3 to $12 per run. Spot instances on Lambda Labs, RunPod, or Vast.ai cut this by 50-70%, putting most runs under $5. The A100 also handles 70B models with QLoRA, which consumer cards cannot.

Cloud GPU (H100, $2.50-4.00/hour): H100s are the fastest option, completing 8B fine-tunes in 1-2 hours. Total cost: $2.50 to $8 per run. The H100's advantage grows with larger models and datasets, where its superior memory bandwidth reduces training time significantly. For one-off fine-tunes, an A100 is more cost-effective. For iterative experimentation where you are running 10+ training runs, the H100's speed saves hours of wall-clock time.

Larger models (70B+): QLoRA fine-tuning of a 70B model requires at least one A100 80GB or H100. Training takes 8-24 hours depending on dataset size, costing $12 to $96 in cloud compute. Full fine-tuning (not QLoRA) of a 70B model requires 4-8 A100s and costs $500 to $2,000+ per run. This is why most teams use QLoRA unless they have specific requirements that demand full fine-tuning.

API-Based Fine-Tuning Costs

API fine-tuning services handle all the infrastructure, so you upload your data and get back a fine-tuned model endpoint. Pricing is per training token.

Together AI offers the most competitive pricing for open-source model fine-tuning. LoRA fine-tuning on models up to 16B parameters costs $0.48 per million training tokens. For a 5,000-example dataset averaging 500 tokens per example over 3 epochs, the total is 7.5 million training tokens, which costs $3.60. Even a 10,000-example dataset at 1,000 tokens per example (30 million tokens over 3 epochs) costs only $14.40. Together supports Llama, Mistral, Qwen, and other popular open-source models.

OpenAI pricing varies by model. GPT-4.1 Nano (the smallest and cheapest) trains at $0.20 per million tokens. GPT-4.1 Mini is approximately $0.80 per million tokens. GPT-4o runs $25 per million training tokens, which is significantly more expensive. Note that OpenAI wound down fine-tuning for newer models (GPT-5 series) in May 2026, so fine-tuning is only available on the GPT-4.1 family and o4-mini for existing customers. A 5,000-example dataset on GPT-4.1 Nano costs under $1. The same dataset on GPT-4o costs approximately $62.

Google Vertex AI offers fine-tuning for Gemini models, with pricing based on node-hours rather than tokens. Tuning a Gemini model starts at approximately $3-5 per node-hour, with most jobs completing in 2-8 hours. Total cost for a standard fine-tune is $10 to $40, competitive with cloud GPU self-hosting but with less control over hyperparameters.

The API approach makes sense when you want to avoid GPU infrastructure entirely, when you need to fine-tune a closed-source model, or when your dataset is small enough that the per-token pricing is cheaper than renting a GPU for even one hour.

Enterprise Fine-Tuning Services

Enterprise fine-tuning is a managed service where a vendor handles everything from data curation through deployment and monitoring. Pricing typically ranges from $5,000 to $50,000 per engagement, depending on scope.

What you get for the premium includes: expert data scientists who review and optimize your training data, multiple training runs with systematic hyperparameter sweeps, comprehensive evaluation against domain-specific benchmarks, deployment support including model hosting and monitoring, and ongoing maintenance with scheduled retraining. Providers in this space include Scale AI, Anyscale, Lamini, and the professional services teams at Anthropic, Google, and Microsoft.

Enterprise services make economic sense when the cost of getting it wrong is high (medical, legal, financial applications), when your team lacks ML expertise, or when the time saved by having experts handle the project is worth more than the fee. For a company that bills $200/hour for engineering time and would spend 100 hours learning fine-tuning from scratch, a $10,000 enterprise engagement is breakeven on labor alone, before accounting for the likely better results from experienced practitioners.

The Hidden Cost: Data Preparation

Compute is the cheapest part of fine-tuning. Data preparation is the expensive part, and it is where most budgets underestimate.

Collecting examples: If you have production logs with user feedback, extracting training examples is a scripting task that takes a few hours. If you need to create examples from scratch, each high-quality example takes 5-15 minutes for a domain expert to write and validate. At 1,000 examples, that is 80-250 hours of expert time. At $50-100/hour for domain expertise, the data preparation cost is $4,000-25,000, dwarfing the $10-30 compute cost.

Cleaning and validation: Even with production-sourced data, cleaning takes significant effort. Removing PII, fixing errors, standardizing formatting, and validating quality across a 5,000-example dataset takes 20-40 hours. This is tedious work that cannot be fully automated because it requires domain judgment to determine whether an example is correct.

Iteration: First fine-tunes rarely produce perfect results. Plan for 3-5 training cycles where you identify weaknesses in the model's behavior, add targeted training examples to address those weaknesses, retrain, and evaluate. Each cycle adds a few hours of data work and a few dollars of compute. The total data preparation cost typically lands at 2-5x the initial estimate.

Cost Comparison Table

Self-hosted RTX 4090: $0.30-0.55 per run (electricity only), 4-8 hours training time, 8B models only with QLoRA.

Cloud A100 (spot): $3-6 per run, 2-4 hours, handles up to 70B with QLoRA.

Cloud H100 (spot): $2.50-8 per run, 1-2 hours, fastest option.

Together AI API: $3-15 per run, no GPU management, Llama/Mistral/Qwen.

OpenAI API (4.1 Nano): $0.50-2 per run, no GPU management, OpenAI models only.

Google Vertex AI: $10-40 per run, no GPU management, Gemini models.

Enterprise managed: $5,000-50,000 per engagement, full service including data curation.