Local inference / mining / token economics
Where does your GPU pay back first?
Compare current crypto mining profit with the value of running Qwen3.8 27B locally. Change the assumptions. The table recalculates in your browser.
RTX 3090 through frontier API savings.
Change the assumptions
41 cards shown. Click a GPU name for the calculation detail.
Selected card
RTX 3090
The selected row shows the inputs used for each payback path. Mining values use the current Hashrate.no best row. API values use the local generation estimate and the selected token prices.
Calculation detail
What the numbers mean
Mining net subtracts GPU power from the current Hashrate.no gross daily estimate. Local API savings multiply Qwen generation speed by 86,400 seconds, then subtract AI power. Missing mining rows show n/a instead of inventing a crypto return.
Benchmark boundary
The RTX 3090 Qwen result is measured. Other generation rates use reported results, memory bandwidth, architecture estimates, or a named range. Runtime, quantization, context length, and drivers change the result. Legacy warnings cover cards that need pinned drivers, older ROCm, or custom builds.
Source links
- Hashrate.no GPU data
- Hashrate.no hosting and card index
- Qwen3.8 27B benchmark table
- Market API pricing
- Anthropic official pricing
- llama.cpp CUDA architecture support
- Ollama GPU support
- ROCm compatibility matrix
- ROCm Ryzen and Strix Halo support
- AMD ROCm llama.cpp guide
- AMD Ryzen AI Max local LLM result
- NVIDIA DGX Spark
- AMD Strix Halo street price
- Used GPU price reference
