QLoRA on RTX 4090 in 2026: True Total Cost After 100 Training Runs vs RunPod

This article was originally published on runaihome.com

The standard pitch for buying an RTX 4090 for QLoRA fine-tuning goes like this: "Cloud rentals are a treadmill. A $1,500 card pays for itself in a few months." That math was honest in 2023. In May 2026 it is wrong, and the people repeating it haven't checked a used-card listing in a long time.

Used RTX 4090s on eBay now average $2,374.95, up from a $1,599 launch MSRP, and Amazon new units sit at $2,755. Meanwhile RunPod Community Cloud rents the same GPU for $0.34/hour. The breakeven point has moved hard against ownership for anyone whose only AI workload is fine-tuning.

This article does the math nobody else publishes—100 realistic QLoRA training runs, end to end, with verified prices and electricity rates—and names the cases where local still wins anyway.

What "100 Training Runs" Actually Means

Hand-wavy training-time numbers are why these comparisons usually fall apart. "QLoRA on RTX 4090 takes 3–4 hours" is true for a clean Llama-2 13B fine-tune on a small instruction dataset. It is not true for the iterative reality of an actual project, where you'll run dozens of short experiments before you have a single keeper.

Here is a realistic 100-run breakdown drawn from typical solo-dev fine-tuning projects:

Run type	Count	Hours each	Total hours
Quick experiments (small dataset, low rank, 1 epoch)	60	1.5–3	~135
Standard runs (full dataset, rank 16–64, 1–2 epochs)	30	3–6	~135
Long runs (longer context, full eval, 3+ epochs)	10	6–12	~90
Total	100		~360 hours

We'll round to 500 GPU-hours to account for the idle time you spend with the card spun up while you fix a dataset bug or rewrite a training script. That number—500 hours over a 3-year horizon—is the only honest unit for this comparison.

For reference, DigitalOcean's benchmark for a 13B Llama QLoRA fine-tune on a single RTX 4090 (24 GB) lands at the 3–4 hour mark for a clean run. The arXiv profiling paper Profiling LoRA/QLoRA Fine-Tuning Efficiency on Consumer GPUs confirms that QLoRA's memory savings come with a ~39% training-time penalty vs. plain LoRA, which means the "fast" QLoRA numbers you see quoted on Twitter usually need adjustment.

The Local Path: Buy a Used RTX 4090

Hardware acquisition

In May 2026 you have three real options:

1. Used RTX 4090 on eBay: average $2,374.95 (per the bestvaluegpu.com tracker, May 2026). Range $2,250–$2,470 across condition.
2. New RTX 4090 on Amazon: $2,755 if you can find stock—the card is officially end-of-life, so retail availability is patchy.
3. MicroCenter trade-in benchmark: $699 for a card in mint condition. Useful as a residual-value floor.

We'll model the used path at $2,374.

Power costs over 500 hours

The 4090 is a 450 W card per NVIDIA's official spec page. Under real fine-tuning load it pulls close to that ceiling, with brief drops between optimizer steps.

US average residential electricity in February 2026 was 17.65¢/kWh (EIA Electric Power Monthly, April 23, 2026 release). California and New England residents pay more; gulf-state residents pay less. We'll use the national average.

500 hours × 0.45 kW × $0.1765/kWh = $39.71

Add a generous 15% overhead for cooling air conditioner load (most US homes need extra AC when a 450W heater runs in the office) and PSU efficiency loss at sustained 80%+ load: ~$46 total electricity over the project.

Residual value at year 3

This is where the cloud-vs-local pitch usually cheats. People model the 4090 holding its value forever. It won't. The 4090 is already a 3.5-year-old discontinued architecture in May 2026, and an RTX 5090 with 32 GB outperforms it on both raw throughput and VRAM headroom.

A conservative 3-year residual estimate, assuming an RTX 60-series launches in late 2027 or 2028:

• Optimistic: $1,000 (held value because consumer GPU shortages persist)
• Conservative: $600 (used market collapses once 60-series hits volume)

We'll use the conservative $600 residual, mid-2029.

Local 3-year total

$2,374  acquisition (used)
+   $46  electricity, 500 hours
-  $600  residual at year 3
= $1,820  net 3-year cost for 500 fine-tuning hours
= $18.20  per fine-tuning hour

If your 500 hours land across 100 runs averaging 5 hours each, each fine-tuning run costs you about $91. Cooling, dust, and the inevitable PSU or fan replacement push this slightly higher.

The Cloud Path: RunPod RTX 4090

Community Cloud at $0.34/hour

RunPod's RTX 4090 page lists Community Cloud at $0.34/hr and Secure Cloud at $0.69/hr in May 2026. Community Cloud means individual host machines can drop off, so you need to checkpoint your training. For fine-tuning that's a non-issue—most QLoRA scripts checkpoint every N steps anyway.

500 hours at $0.34/hr: $170.

Add network-attached storage so your model checkpoints persist across pod instances. RunPod charges roughly $0.07/GB-month for network volumes. A 50 GB volume held for 3 months: $10.50.

Community Cloud total: ~$181 for 500 hours, or $1.81 per fine-tuning hour.

Secure Cloud at $0.69/hr

If your data has compliance requirements that rule out Community Cloud (HIPAA, certain corporate policies), Secure Cloud is the next tier. 500 hours × $0.69 = $345 + $10 storage = $356, or $3.56/hour.

The honest cloud caveat

Cloud-vendor pricing pages never mention this, but it matters: you pay for the wall-clock time the pod is running, not training compute time. If you leave the pod up overnight while you debug a dataset issue, that counts. If you fall asleep during a long run, that counts. Discipline matters—but a runpod-cli stop and pod start cycle takes 30 seconds, so the discipline is achievable.

The Real Comparison Table

                          Local 4090 (used)    RunPod Community    RunPod Secure
Hardware/setup            $2,374               $0                  $0
500 hr compute            $46 (electricity)    $170                $345
Storage                   $0                   $10                 $10
Year-3 residual           -$600                $0                  $0
─────────────────────────────────────────────────────────────────────────────
3-year total              $1,820               $181                $356
Per fine-tuning hour      $18.20               $0.38               $0.72
Per 5-hour run            $91                  $1.91               $3.58
Hours to breakeven        —                    4,789 hr            2,528 hr

Read that last row. To make the used-4090 path beat RunPod Community Cloud on pure fine-tuning, you need to put 4,789 hours of training on the card before year 3 ends. That is 26 hours per week of fine-tuning, every week, for three years. Even a research lab doesn't do that.

Against Secure Cloud the math is less brutal—2,528 hours, or 16 hours/week—but still nowhere near what 100 realistic runs deliver.

When Local Wins Anyway

The math above is the worst-case framing because it assumes the 4090 only does fine-tuning. Most people who buy one don't use it that way. Here is when the local path actually beats cloud, and the reasoning is honest:

1. The card is a shared workload device

If the same 4090 also runs:

• Daily inference for a self-hosted chatbot (Open WebUI, Ollama)
• ComfyUI for image generation
• Gaming on the off-hours
• Code completion via a local Llama 3.3 70B with offload

…then fine-tuning is the marginal use, not the primary one. The card was paid for by the other workloads. Ma