How much does fine-tuning an LLM cost across platforms?

It varies by method and platform: full fine-tuning charges for all parameters across training tokens, while LoRA and QLoRA update far fewer parameters and cut GPU-hours dramatically. The calculator compares AWS Bedrock, Google Vertex AI, Azure OpenAI, and self-hosted GPU across training, inference, and hosting costs with 2026 pricing.

What is the difference between full fine-tuning, LoRA, and QLoRA?

Full fine-tuning updates every weight and is the most expensive; LoRA adds small trainable adapter matrices, cutting trainable parameters by orders of magnitude; QLoRA combines LoRA with 4-bit quantization of the base model for the lowest memory and cost. The calculator shows all three methods per platform, so the accuracy-vs-cost trade-off is explicit.

Which fine-tuning pricing traps should I watch for?

Per-token training prices that look small but compound over millions of samples, storage charges for the base model copies, and inference hosting that costs more than the training run. The calculator's platform traps section surfaces these line items so the annual TCO — not the training demo — is the decision number.

LLM Fine-tuning Cost — Bedrock vs Vertex vs Azure

📐 Fine-tuning Workload Profile

Configure model size, dataset, and method. The engine estimates training cost, GPU hours for self-hosted, and monthly inference TCO across all platforms.

Base Model

Dataset Size (tokens)

Epochs

Fine-tuning Method

Monthly Inference (M tokens)Millions of tokens served per month

Self-Hosted GPU Provider

🎯 Fine-Tuning Method Comparison

QLoRA fits 70B models on a single A100 80GB. Full fine-tuning needs 4-8 H100s. Click to compare costs across methods.

QLoRA (4-bit)1×A100 80GB or 1×RTX 4090 · 5–50¢/run · MB-sized adapters

LoRA (16-bit)1-2×A100 80GB · $3-50/run · Multi-tenant adapter serving

Full Fine-Tuning4-8×H100 · $50-500/run · Highest quality · GPU-obsolete risk

📊 Training Cost — Per Platform

Platform	Training Cost	Method / Notes	Model Storage

💰 Monthly Inference Cost + Annual TCO

Platform	Inference Rate	Monthly Inference	Monthly Hosting	Monthly Total	Annual TCO

Fine-Tuning Platform Pricing Traps

AWS Bedrock: Fine-tuned models require Provisioned Throughput (PTU) — a 24/7 hourly commitment with no scale-to-zero. At $7.10/hr for Titan Text Lite PTU (no commitment), the monthly floor is $5,183 before a single inference token. With 1-month commitment: $4,672/month. Even a 6-month commitment leaves you at $3,723/month. Bedrock fine-tuning is not for low-volume inference. It's for production workloads serving millions of tokens daily where the per-token PTU rate beats on-demand pricing.

Azure OpenAI: The training cost is modest ($5-25 for a typical fine-tuning run on GPT-4o-mini). The trap is hosting: $1.70/hour, 24/7, no idle discount. An idle deployment burns $1,224/month serving zero requests. Training cost is irrelevant — hosting is the entire cost decision.

Vertex AI: Gemini fine-tuning uses LoRA under the hood. Inference on tuned models runs at the base model's standard per-token rate with zero tuning surcharge and zero hosting fee. At $0.30/1M tokens input for Gemini Flash: 100M tokens/month inference = $30/month total. Vertex is the only platform where fine-tuning doesn't create a recurring cost obligation.

Self-hosted: QLoRA on an RTX 4090 via RunPod runs ~30 minutes for a 10K-sample 7B dataset at ~$0.50 total. Inference on the same GPU costs ~$0.39/hr. At 100M tokens/month with 8GB VRAM inference: $285/month GPU rental. No hosting fee, no minimum commitment, no surcharge. Self-hosted is cheaper than any managed platform at any inference volume — the trade is operational expertise.

Pricing Basis, Sources & Assumptions

Every rate on this page is a published vendor list price — no negotiated discounts, private pricing, or credit offsets. Totals are in USD and exclude tax. Rates were last checked against the sources below on 2026-09-23.

Pricing input	Basis used on this page
Region	US East (N. Virginia) for Bedrock, and each platform's published US pricing for Vertex AI and Azure OpenAI. Managed fine-tuning platforms price by region, so a deployment outside the US should re-check every line.
Currency	USD — on-demand list price, tax excluded
Last checked	2026-09-23 — The platform rates quoted here — Bedrock provisioned throughput at $7.10/hr for Titan Text Lite with the one- and six-month commitment tiers, Azure OpenAI hosting at $1.70/hr, and Gemini's per-token rate — were reviewed against the linked vendors' price pages on this date.

Pricing input

Basis used on this page

Region

US East (N. Virginia) for Bedrock, and each platform's published US pricing for Vertex AI and Azure OpenAI. Managed fine-tuning platforms price by region, so a deployment outside the US should re-check every line.

Currency

USD — on-demand list price, tax excluded

Last checked

2026-09-23 — The platform rates quoted here — Bedrock provisioned throughput at $7.10/hr for Titan Text Lite with the one- and six-month commitment tiers, Azure OpenAI hosting at $1.70/hr, and Gemini's per-token rate — were reviewed against the linked vendors' price pages on this date.

Modelling assumptions

Two costs, not one: a training run and a hosting charge. Training is billed per processed token or per run and is usually the smaller number; hosting is billed by the hour and runs continuously.

Managed platforms do not scale to zero on their own. A provisioned deployment continues to bill at its hourly rate while serving zero requests, which is what makes an idle model expensive.

Provisioned throughput is bought as a commitment, and the hourly rate falls as the commitment lengthens. The page shows the uncommitted rate alongside the discounted tiers so the floor is visible.

Self-hosted and LoRA paths are priced per GPU-hour at published rental rates, with the run length as the multiplier.

Adapter fine-tuning keeps the base model untouched, so only the adapter is stored and served; full fine-tuning produces a new set of weights to host.

What this model excludes

Dataset preparation, labelling, cleaning, and evaluation runs.

Storage for the training dataset, checkpoints, and the served adapter or weights.

Inference tokens consumed after deployment, which on managed platforms are billed on top of the hosting charge.

The engineering time to build the evaluation loop, and the cost of a bad fine-tune that regresses.

Sources

AWS (2026). "Amazon Bedrock Pricing." Custom-model training, storage, and provisioned-throughput rates including commitment tiers. aws.amazon.com

Microsoft Azure (2026). "Azure OpenAI Service Pricing." Fine-tuning training and hourly hosting rates for the Azure deployment model. azure.microsoft.com

Google Cloud (2026). "Generative AI on Vertex AI Pricing." Gemini supervised fine-tuning rates and per-token serving prices. cloud.google.com

AWS (2026). "AWS Lambda Pricing." Serverless compute baseline used for the non-GPU alternative in the comparison. aws.amazon.com

Vendor list prices change without notice — re-check the linked pages before committing spend. jslet takes no vendor sponsorship and carries no affiliate links; see about.