Why now?

Local LLM hardware buying is a live, confusing decision right now: the Show HN launch drew 323 points, which is a strong signal of genuine interest in the pre-purchase speed question. The tool itself is lightweight (just running models and timing tokens), so a solo dev can ship and iterate fast without heavy compute or frontier-model dependencies.

The gap

The pain point is clear and validated by the 323 HN points: people want to know how fast a local LLM will run before spending money on hardware. No incumbent is named in the signals, and the opportunity is a curated, trustworthy speed database rather than defensible tech. Demand for a paid product is not proven — the audience is niche and price-sensitive.

Main competitor

No specific competitor is identified in the provided signals. The closest substitute is scattered community benchmark posts and crowdsourced result tables, which are easy to replicate or scrape.

Execution plan

1. Ship a minimal benchmark runner that times tokens/sec for popular open models across common quantization levels, and publish results as a public, filterable table.
2. Add a hardware picker: user selects GPU/CPU/RAM budget, tool returns expected tokens/sec from the benchmark data.
3. Seed the database with your own reproducible runs on a handful of common consumer GPUs, documenting methodology and model/quantization versions so results stay credible.
4. Open submissions with version tagging so results can be validated and stale entries flagged; keep a changelog as models and drivers update.
5. Add hardware affiliate links next to each result and a vendor sponsorship slot to test monetization before building any paid tier.

Monetization

Affiliate links to hardware vendors alongside benchmark results, plus vendor sponsorship of the speed database. A paid tier is unproven — the signals show interest in the free decision tool, not willingness to pay, so treat subscriptions as a later experiment.