Home / Buy vs rent
Buy or rent hardware for your AI workload?
Purchase prices come from the current validated build BOM or an exact workstation configuration. Rent comes from currently priced GPU server offers. Electricity uses active hours only. Token cost per 1M output tokens is shown only when Phase 11C serving evidence is comparable on both sides.
Advanced assumptions
Your workload
SmolLM2 1.7B Instruct · Q4 · 8K
- 200 active hours/month · RigForAI default assumption
- $0.15/kWh · RigForAI default assumption, not a national tariff
- 24-month comparison horizon · RigForAI default assumption
- 36-month amortization for the monthly equivalent view only · RigForAI default assumption
- Maintenance $0.00/month · RigForAI default assumption ($0)
- Idle wall power outside active hours is excluded (unknown idle draw).
- PSU rated wattage is not used as system power draw.
Buy
Cheapest validated custom build
Validated custom build (NVIDIA GTX 1660 SUPER 6 GB)
Purchase $725.32 · source Phase 9 current BOM
Electricity $19.63/month (estimated 654 W × 200 h)
Amortized hardware cost $20.15/month
Monthly equivalent under selected amortization assumptions $39.78
Cumulative at 24 months $1,196.52
Prebuilt workstation
Lenovo P5 · NVIDIA RTX 2000 Ada Generation · 64 GB RAM
Purchase $4175.00 · exact configuration, Phase 10 current price
Electricity —/month (unknown — CPU/GPU wall model incomplete; $0 unless you supply watts)
Monthly equivalent under selected amortization assumptions $115.97
Cumulative at 24 months $4,175.00
Rent
Memory-capable rental options with current Phase 11 prices. Hourly offers use your active hours. Monthly dedicated offers use the contract price regardless of hours. Setup fees are separate.
| Offer | Billing | Usage-adjusted / month | Setup | Cumulative 24 mo | Availability |
|---|---|---|---|---|---|
| Vast.ai · 1× NVIDIA RTX 5070 12 GB Memory-capable rental option · Vietnam, VN · VN | hour · $0.074/h | $14.81 | $0.00 | $355.56 | listed as available |
| Vast.ai · 1× NVIDIA RTX 5080 16 GB Memory-capable rental option · Ontario, CA · CA | hour · $0.147/h | $29.48 | $0.00 | $707.56 | listed as available |
| Vast.ai · 1× NVIDIA RTX 4090 24 GB Memory-capable rental option · Iowa, US · US | hour · $0.269/h | $53.78 | $0.00 | $1,290.67 | listed as available |
| Vast.ai · 2× NVIDIA RTX 5070 Ti 16 GB Memory-capable rental option · Poland, PL · PL | hour · $0.309/h | $61.78 | $0.00 | $1,482.67 | listed as available |
| Vast.ai · 4× NVIDIA RTX 3090 24 GB Memory-capable rental option · Japan, JP · JP | hour · $0.406/h | $81.16 | $0.00 | $1,947.73 | listed as available |
Lowest monthly rental cost under these usage assumptions: Vast.ai hour $14.81 · provider listing
Lowest cumulative cost after 24 months (rent side): Vast.ai hour $355.56
Break-even
At 200 active hours/month and $0.15/kWh, this Vast.ai hour offer remains cheaper through the cash-flow model — rental monthly cost is not higher than local electricity+maintenance, so buying does not catch up.
| BUY vs this RENT | State | Months |
|---|---|---|
| Validated custom build (NVIDIA GTX 1660 SUPER 6 GB) vs Vast.ai (hour) | RENT_ALWAYS_CHEAPER_WITHIN_MODEL | — |
| Validated custom build (NVIDIA GTX 1660 SUPER 6 GB) vs Vast.ai (hour) | BREAK_EVEN | ~73.7 |
| Validated custom build (NVIDIA GTX 1660 SUPER 6 GB) vs Vast.ai (hour) | BREAK_EVEN | ~21.2 |
| Validated custom build (NVIDIA GTX 1660 SUPER 6 GB) vs Vast.ai (hour) | BREAK_EVEN | ~17.2 |
| Validated custom build (NVIDIA GTX 1660 SUPER 6 GB) vs Vast.ai (hour) | BREAK_EVEN | ~11.8 |
Cumulative cost
Teal: custom build. Other lines: lowest-cost rentals under your usage. Not a ranking of “best”.
Serving economics
Cost per 1M output tokens: TOKEN_ECONOMICS_UNAVAILABLE
- Phase 5 llama.cpp tok/s is not used.
- Phase 11C vLLM serving evidence is required on both BUY and RENT under the same model, runtime, precision, workload shape, serving mode, concurrency, and EXACT_HARDWARE mapping.
- runtime_not_vllm_on_both_sides
- catalog_model_mismatch_or_missing
No EXACT_HARDWARE Phase 11C mapping on the ranked rental offers for this catalog model.
Buy vs rent by AI model
Model → GPU fit → custom build, workstation, or rental server → cash-flow comparison. Calculator query states are not indexed.
- Llama 3.1 8B Instruct · Cheapest GPU · Run comparison
- Llama 3.1 70B Instruct · Cheapest GPU · Run comparison
- Llama 3.3 70B Instruct · Cheapest GPU · Run comparison
- Qwen2.5 7B Instruct · Cheapest GPU · Run comparison
- Qwen2.5 14B Instruct · Cheapest GPU · Run comparison
- Qwen2.5 32B Instruct · Cheapest GPU · Run comparison
- Qwen2.5 72B Instruct · Cheapest GPU · Run comparison
- GPT-OSS 20B · Cheapest GPU · Run comparison
- GPT-OSS 120B · Cheapest GPU · Run comparison
- Mistral Small 24B Instruct · Cheapest GPU · Run comparison
- Gemma 2 9B Instruct · Cheapest GPU · Run comparison
- Gemma 2 27B Instruct · Cheapest GPU · Run comparison
- Mixtral 8x7B Instruct · Cheapest GPU · Run comparison
- Llama 3.2 3B Instruct · Cheapest GPU · Run comparison
Compare ways to run AI
How buy vs rent works
- Select an AI workload.
- Compare current compatible BUY and RENT options.
- See cumulative cost and break-even under your assumptions.