GPU Hashrate to Inference Capacity Converter
Turn your crypto-mining hashrate into an estimated local LLM tok/s number. Compare electricity cost with cloud-equivalent inference value.
Scenario preset
Pivoting from mining to AI inference? Check the rest of the stack.
Compare local ROI, hosting revenue, and hardware resale value before you switch rigs.
Local LLM ROI → Miner Pivot Calculator →⛏️ Mining input
⚡ Power & economics
Estimated tok/s
—
—
Tokens / day
—
Tokens / month
—
Electricity cost / mo
—
Cloud-equivalent value / mo
—
Net value vs power
—
GPU comparison for this algorithm & model
| GPU | VRAM | Hashrate | Tok/s | Tokens/day |
|---|
Sensitivity to hashrate
| % of reference | Hashrate | Tok/s | Tokens / mo |
|---|
Verdict
—
How the math works
- The converter looks up the selected GPU's reference hashrate for the chosen algorithm and its reference tok/s for the chosen model.
- It computes a ratio = your measured hashrate ÷ reference hashrate, then multiplies reference tok/s by that ratio.
- For custom GPUs, the same ratio is applied using the reference GPU you select.
- Monthly tokens assume the selected inference utilization % applied to the hours-per-day the rig is online.
- Cloud-equivalent value = monthly tokens ÷ 1,000,000 × your chosen API price per 1M tokens. It shows what you would pay for the same output at that API rate.
Last updated: 2026-07-20. See notes.
Frequently asked questions
How does mining hashrate relate to LLM inference speed?▼
Both depend on GPU memory bandwidth, compute, and power. A card that hashes faster on memory-bound algorithms usually also runs LLMs faster, but the correlation is rough because different algorithms stress different parts of the chip. This tool scales measured hashrate against known benchmarks to estimate tok/s.
Which algorithm gives the best proxy for LLM inference?▼
Memory-bound algorithms like Etchash, ZelHash, and Octopus are closer to LLM inference than compute-bound algorithms like KawPow. Use the algorithm you are actually mining for the most relevant pivot estimate.
Can I use this for multi-GPU rigs?▼
Yes. Enter the total measured hashrate of the whole rig and set the power draw and hours to match the full system. The converter scales linearly, so a rig with 2x the hashrate of a single card is estimated to deliver roughly 2x the tok/s.
Why does the 70B model show zero tok/s for some cards?▼
A 70B parameter model at 4-bit quantization needs roughly 40 GB of VRAM. Cards with 16 GB or less cannot run it without aggressive offloading, which is usually too slow to be useful, so the calculator marks them as unsupported.
Is this a guarantee of actual inference revenue?▼
No. The cloud-equivalent value is a benchmark, not a price you will automatically earn. Rental pricing, demand, downtime, and model quality all affect real hosting revenue. Use the Local LLM Hosting Revenue Estimator for a full business model.
This converter estimates local LLM inference capacity by scaling known GPU benchmarks with your measured crypto-mining hashrate. Hashrate and inference throughput are not perfectly correlated because different algorithms stress memory bandwidth, compute, and cache differently, but the ratio is a useful rule of thumb for pivot planning. Actual tok/s depends on quantization, context length, batching, and backend.