Independent, reproducible cost research. Every tool below is free, zero-signup and runs arithmetic in your browser or your terminal — no account, no email wall, no "request a demo".
The premise is simple: vendor price lists are quoted in units your business does not actually buy (GPU-hours, gigabytes, requests, MoR percentage). These calculators and CLIs convert them into the unit you do buy — cost per usable GPU-hour, per 1,000 clean pages, per completed flow, per retained dollar.
Every formula is visible on the page. Nothing is uploaded; all math runs client-side.
| Calculator | What it prices | Hosted | Dataset |
|---|---|---|---|
| 🖥️ LLM VRAM & GPU Cloud True TCO Calculator | Weights + KV cache (GQA/MLA), NVLink vs PCIe topology, card-count rounding, RunPod/Vast.ai bundle billing → cost per useful GPU-hour | gpu.mybskm.top/calculator · HF Space | HF Dataset |
| 🕷️ LLM/RAG Web Scraping & Anti-Bot True TCO Calculator | 1/(1-block_rate) retry amplification, challenge-page bandwidth bloat, Chromium render cost, managed-credit tiers → $ per 1,000 clean Markdown pages |
scrape.mybskm.top/calculator · HF Space | HF Dataset |
🌐 Residential Proxy Sticky Session & cf_clearance Loss Calculator |
Premature sticky-IP drops invalidating cf_clearance, geometric re-run tax, real GB burned → sticky sessions vs per-request rotation, monthly |
proxy.mybskm.top/calculator · HF Space | HF Dataset |
| 💳 Non-US Indie SaaS: MoR (Paddle/LS) vs US LLC + Stripe Breakeven | 5% Merchant-of-Record drag vs Wyoming/Delaware franchise tax, registered agent, Form 5472 exposure → at which MRR the LLC wins | pay.mybskm.top · HF Space | HF Dataset |
Every model above also ships as a machine-readable dataset with a documented schema, so you can audit or re-run the arithmetic instead of trusting a screenshot. CSV + JSONL, 10–252 rows each, rendered by the HuggingFace Dataset Viewer.
| Dataset | What is inside | Rows |
|---|---|---|
gpu-cloud-h100-a100-pricing-benchmark |
RunPod / Vast.ai / Lambda / CoreWeave / AWS / GCP / Azure list rates for H100 SXM & PCIe, A100 80GB, L40S, RTX 4090 + cold-start, NVMe and 85%-utilisation effective cost per useful GPU-hour | 28 |
web-scraping-anti-bot-api-benchmark |
ScraperAPI, ZenRows, ScrapingBee, Crawlbase, Bright Data, Smartproxy residential and self-hosted Playwright: JS/premium credit multipliers, block rates, retry amplification, $/1k clean pages (measured + hard-target stress) | 14 |
residential-proxy-commercial-policy-matrix |
9-dimension commercial matrix for Smartproxy (Decodo), Bright Data, Oxylabs, IPRoyal, SOAX, Webshare and the free curl_cffi + WARP baseline, plus sticky-session re-run tax across 3 flow shapes |
21 |
global-saas-mor-vs-us-llc-fee-matrix |
Paddle / Lemon Squeezy 5% + $0.50 vs US LLC + Stripe 2.9% + intl/FX surcharges across $1k–$50k MRR, $15–$99 AOV and 3 card mixes, with 3-year net retained and Firstbase / Atlas / Doola entity costs | 252 |
Usage:
pip install datasets
python3 -c "from datasets import load_dataset; print(load_dataset('600me/residential-proxy-commercial-policy-matrix', split='train')[:3])"| Repository | What it does |
|---|---|
600me/check-proxy-tax |
Zero-dependency Python CLI (stdlib only). Fires a concurrent burst at one URL, classifies every response as usable or blocked, and reports usable rate, p50/p95 latency, bytes transferred and effective cost per 1,000 usable requests — the retry tax plus bandwidth amplification your invoice never shows. |
600me/awesome-web-scraping-apis-benchmark |
Cloudflare/Turnstile bypass and cost benchmark across managed scraping APIs: credit multipliers, render surcharges, and the only metric that survives production. |
600me/awesome-residential-proxies-benchmark |
9-dimension commercial policy and sticky-session telemetry matrix — session TTL, pool transparency, pricing granularity, concurrency limits, refund terms. |
600me/awesome-gpu-cloud-pricing-benchmark |
Daily H100/A100 SXM vs PCIe pricing and hidden-cost tracker: egress, NVMe, cold-start pre-pull, bundle minimums. |
600me/awesome-indie-global-payments-matrix |
Cross-border SaaS payment and US LLC compliance math for non-US founders, with the full 3-year cost ledger. |
Long-form, arithmetic-first research. Each page shows the full cost model, the crossover point, and where the "cheap" option stops being cheap.
GPU & LLM inference
- The H100 Rental Trap: Why $2.13 per GPU-Hour Is Not the Price You Pay — catalogue rate vs effective cost of a useful GPU-hour after bundle minimums, utilization gaps and restart overhead.
- RunPod vs Vast.ai (2026): Real H100 & A100 Effective Cost After Cold-Start, NVMe Storage & Bandwidth Billing — a 150 GB two-hour test costs 63% more on the "25% cheaper" marketplace; a 24/7 node flips back above 5.4 TB/month.
Scraping & anti-bot
- The 403 Tax: What Anti-Bot Protection Really Does to Your Scraping Bill — credit multipliers, retry amplification, and the metric that survives production.
- Smartproxy (Decodo) vs ScraperAPI (2026): Raw Residential Pool vs Managed Anti-Bot Unblocker TCO — $3.50/GB vs $299/month are not competing products; three worked scenarios plus the 30-line hybrid router.
- The $0.49 Proxy Illusion: How a 200-Line CLI Exposes Real Scraping Costs — instrumenting a real burst to convert advertised price into cost per usable page.
- Why Rotating Residential Proxies Still Get 403s: Sticky Session TTL, JA3 Fingerprints and ASN Economics — a rotating IP is not a trusted IP; how clearance binding creates the re-run tax.
Cross-border payments & compliance
- The 5% MoR Illusion: Exact US LLC + Stripe Breakeven for Non-US SaaS Founders (2026) — break-even at $3,205 / $5,376 / $15,152 MRR across three real card mixes, compliance bill included.
- Firstbase vs Stripe Atlas vs Doola (2026): True 3-Year US LLC Cost for Non-US Founders — formation fee is not the price: $876 vs $1,600 vs $6,177 over three years, plus the $25,000 Form 5472 penalty exposure.
- Reproducible by design. Static HTML + vanilla JS, or a single stdlib-only Python module. Read the source, re-run the numbers, argue with the arithmetic.
- Vendor-neutral arithmetic. Calculators include links to commercial providers, some of which are affiliate links. Commission never changes a formula, a default or a recommendation — the models are also published as standalone HTML and JSON you can audit.
- Planning models, not quotes. Published rates are snapshots and can change; verify live pricing with the provider before you commit money. Model output is arithmetic, not a measured benchmark or a guarantee.