TL;DR: Renting wins for almost everyone who fine-tunes open models. A QLoRA fine-tune of an 8B model costs $1–3 rented, so a $1,050 used RTX 3090 needs roughly 175 fine-tune runs per year — about 3.4 per week, every week, for three years — to pay for itself on fine-tuning alone. Buy the card only if it also runs your daily inference.

	Rent per job	Buy used 24GB card	Buy new 24GB card
Best for	Occasional or bursty fine-tuning	Fine-tuning + daily local inference	Same, plus warranty and resale value
Up-front cost	$0	~$1,250 (used RTX 3090, Sep 2026)	~$2,250 (used RTX 4090-class)
Cost per 8B QLoRA run	~$1–3 all-in	~$0.15 electricity	~$0.10 electricity
The catch	Your dataset leaves your machine	No warranty, mining-rig risk	Highest capital outlay

Honest take: rent on Vast.ai until the month your rental receipts pass ~$30, or until privacy of your training data becomes non-negotiable. Only then price out a used RTX 3090.

The question comes up every time someone finishes their first Unsloth fine-tune on a rented box: “should I just buy the GPU?” The answer is a number, not a feeling — it depends on how many fine-tune runs you do per year, and almost nobody does enough. All prices below verified September 2026.

How much does renting a GPU for one fine-tune cost in 2026?

Between $1 and $3 all-in for a typical 8B-model QLoRA job, and $5–20 for a 70B QLoRA job, at September 2026 market rates.

The two inputs are hourly rate and job length. On Vast.ai’s marketplace (September 2026), an RTX 3090 rents from $0.07/hr, an RTX 4090 from $0.14/hr, and an RTX 5090 from $0.25/hr — these are floating marketplace floors, so treat them as “from” prices, not guarantees. On RunPod, A100 80GB instances run roughly $1.20–2/hr depending on tier.

Job lengths, using Unsloth (tested numbers from our Llama 3.1 fine-tuning guide, 1,000 training examples, 3 epochs):

Job	Hardware	Compute time	Rental compute cost
Llama 3.1 8B QLoRA	RTX 4090 (from $0.14/hr)	~45–60 min	~$0.15–0.50
Llama 3.1 8B QLoRA	RTX 3090 (from $0.07/hr)	~1.5–2 hrs	~$0.15–0.40
Llama 3.1 70B QLoRA	RTX 4090 24GB	~8–12 hrs	~$1.50–5
Llama 3.1 70B QLoRA	A100 80GB (~$1.20–2/hr)	~4–6 hrs	~$5–12

Real bills land higher than the compute column because the meter runs during setup — more on that below. Budget double the compute cost and you will rarely be surprised. Larger datasets scale roughly linearly: 10,000 examples means roughly 10× the hours, still single-digit dollars for an 8B model on a 4090.

Boundary worth stating: this math is for LoRA/QLoRA adapter training. Full-parameter fine-tuning of even an 8B model wants 8×A100-class hardware and moves the discussion into hundreds of dollars per run — rent, always, unless you are a funded lab.

What does owning a fine-tuning GPU really cost per year?

About $350/year amortized, before electricity, for the standard pick — a used 24GB RTX 3090.

As of September 2026, used RTX 3090 cards typically sell near $1,050 on eBay (asking prices run $1,050–1,680; the used market is up ~11% in 90 days). Spread over a 3-year useful life, that is ~$350/year. A new RTX 4090 at ~$1,800+ roughly doubles that.

Electricity is a rounding error for fine-tuning workloads. A 3090 pulls ~350W under training load; a 2-hour 8B run is 0.7 kWh — about $0.13 at the September 2026 US average residential rate of ~18.6¢/kWh. Even 100 runs a year is ~$13. (Idle draw is a different story if the box runs 24/7 — see always-on server electricity cost by region.)

Resale softens ownership: a working 3090 will likely still fetch several hundred dollars in 2029, which effectively cuts the amortized cost — but you carry the risks a rental never has: no warranty on used cards, ex-mining stock, a PSU that needs 750W+ headroom, and the card being obsolete for whatever model sizes matter in three years.

Where is the rent-vs-buy break-even point?

Around 175 fine-tune runs per year — roughly 3.4 per week sustained for three years — if fine-tuning is the card’s only job.

Take $2 as a realistic all-in rental cost per 8B QLoRA run (Vast.ai 4090, including setup overhead) and $350/year as the amortized cost of a used 3090:

Fine-tune runs per year	Annual rental cost	Owned card (amortized)	Winner
12 (monthly)	~$24	~$350	Rent, by 14×
52 (weekly)	~$104	~$350	Rent, by 3×
~175 (3–4/week)	~$350	~$350	Break-even
250 (every workday)	~$500	~$350	Own

Run your own numbers — the break-even is just capital cost divided by per-job savings:

card = 1050          # used RTX 3090, Sep 2026
years = 3            # useful life
rent_per_job = 2.00  # all-in rental, 8B QLoRA
elec_per_job = 0.13  # 0.7 kWh at $0.186/kWh

breakeven = card / years / (rent_per_job - elec_per_job)
print(f"{breakeven:.0f} runs/year to break even")
# → 187 runs/year to break even

Two honest adjustments. If your jobs are 70B-class ($5–20 rented), break-even drops to roughly 20–70 runs/year — heavy 70B experimenters are the one group where ownership pays on fine-tuning alone, except a 70B QLoRA also takes 8–12 hours on a 24GB card, so weekly cadence is about the practical ceiling anyway. And if you credit resale value (~$500 back after 3 years), the 8B break-even drops to ~100 runs/year — still two per week, every week.

The hidden rental cost: the meter runs during setup

The first time I ran a fine-tune on a rented marketplace instance, the $0.60 job cost $2.10 — because the instance spent 40 minutes downloading the 16GB base model and installing dependencies at full hourly rate before training started.

Three fixes that keep rental bills close to the compute cost:

1. Filter for fast networking. On Vast.ai, sort by bandwidth, not just price — a $0.14/hr instance with 100 Mbps download costs more in practice than a $0.20/hr one with 1 Gbps:

vastai search offers 'gpu_name=RTX_4090 num_gpus=1 inet_down>500' -o 'dph'

2. Enable fast HuggingFace transfers. pip install hf_transfer and export HF_HUB_ENABLE_HF_TRANSFER=1 before downloading weights — this alone routinely triples download throughput on well-connected instances.

3. Do a 60-step smoke test first. Set max_steps=60 in your training config to confirm loss is moving before committing to the full run. A crashed 10-hour job you notice the next morning bills all 10 hours.

When does buying the card still win?

When the GPU has a second job. The break-even above assumes the card sits idle between fine-tunes, and that assumption is what kills the ownership case — so change the assumption.

A 24GB card that runs your daily coding model, a local RAG stack, or image generation is being paid for by those workloads; fine-tuning rides along at electricity cost (~$0.15/run). That is a different calculation — the full year-one cost math is here — and it breaks even against heavy API usage in ways pure fine-tuning never does. Data privacy is the other hard trigger: if your training set is customer data or anything under NDA, a marketplace rental (community hosts on Vast.ai are literally other people’s machines) may be disqualified regardless of price. Buy the card, or at minimum pay the premium for datacenter-tier instances like RunPod Secure Cloud.

For what hardware to put around a 24GB card — PSU, case airflow, or whether a unified-memory mini-PC covers you instead — runaihome.com’s Ryzen AI Max 395 guide covers the current alternatives to a discrete-GPU build.

When NOT to buy a GPU for fine-tuning

• You have run fewer than 20 fine-tunes ever. You do not yet know your real cadence. $40 of rentals teaches you; $1,050 of GPU assumes.
• Your models are 70B+ and you want full fine-tunes, not LoRA. No consumer card does this. Rent A100/H100 clusters per job.
• You pay 30¢+/kWh (California, much of the EU). Always-on ownership economics get materially worse — the electricity line stops being a rounding error once the box does inference duty too.
• You expect to want a different card within 18 months. Model VRAM appetites keep shifting; renting transfers obsolescence risk to the host.
• You were only buying it to “learn fine-tuning.” Learn on a $2 rental first. If self-hosting in general is the question, read when NOT to self-host AI before spending anything.

What to actually buy

Prices as of September 2026, all taken from the comparison above:

Your situation	The pick	Price	Where
Fine-tune 8B models, weekly or less	Rented RTX 4090	from $0.14/hr (~$1–3/run)	Vast.ai
70B QLoRA jobs, or private data on datacenter hardware	Rented A100 80GB	~$1.20–2/hr	RunPod
3+ runs/week AND daily local inference	Used RTX 3090 24GB	~$1,050	Check price
Same, but you want warranty + 2× training speed	RTX 4090 24GB	~$2,150–$2,350	Check price

FAQ

How much does it cost to fine-tune an 8B model on a rented GPU? Roughly $1–3 all-in at September 2026 rates: a Vast.ai RTX 4090 from $0.14/hr finishes a 1,000-example QLoRA run in under an hour, and setup/download overhead typically doubles the raw compute cost. RunPod runs the same job for under $3.

How many fine-tunes per year justify buying a GPU? About 175 per year for 8B-class QLoRA jobs against a $1,050 used RTX 3090 amortized over three years — roughly 3.4 runs per week. Credit ~$500 of resale value and it is still ~100/year. If the card also handles daily inference, the math changes entirely and ownership can win at much lower fine-tuning volume.

Is renting a GPU safe for confidential training data? Not by default on marketplace platforms — Vast.ai community hosts are individuals’ machines, and your dataset and adapter weights sit on their disk during the job. For NDA or customer data, use datacenter-tier offerings (RunPod Secure Cloud) or own the hardware.

Sources

• Vast.ai marketplace pricing (cloud.vast.ai, verified September 2026): RTX 3090 from $0.07/hr, RTX 4090 from $0.14/hr, RTX 5090 from $0.25/hr
• RunPod pricing page (runpod.io/pricing, September 2026): A100 80GB ~$1.20–2/hr by tier
• eBay sold listings for RTX 3090 24GB, September 2026: typical sold prices near $1,050
• Unsloth documentation (docs.unsloth.ai): QLoRA VRAM minimums and training throughput, v2026.5.8
• U.S. EIA average residential electricity rate, ~18.6¢/kWh, September 2026

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe