Cloud GPU price-comparison site with Flatkey as its preferred commercial recommendation. The site is branded RackRate through NEXT_PUBLIC_SITE_NAME; the brand and final domain can be changed without touching page code.
- Filterable GPU-offer table by GPU, provider, region, availability, and rental type
- Indexable GPU and provider detail pages
- Transparent methodology page
- Sitemap and robots metadata
- Standalone Next.js build and Docker image
- Versioned dataset with source links, observation dates, and a fail-closed official-page collector
- Flatkey retains a separate commercial service profile and quote-request CTAs carrying per-page UTM attribution
- Sourced offer rows remain sorted by price; Flatkey has no invented price or inventory records
RackRate consumes the reusable GPU Price Monitor Python library, pinned to 0.24.0, for ten tracked providers: Nebius, Crusoe, Lambda, UCloud, AutoDL, RunPod, Microsoft Azure, DigitalOcean, Verda and Vast.ai. Anonymous sources need no provider credentials. The explicit authorized mode also collects signed-in catalogs and account estimates; it requires approved sessions, with Nebius account access enabled separately or explicitly excluded. Neither mode starts GPU instances. See daily publication operations for current source limitations and cloud validation status. The library only fetches and returns quotes; RackRate owns the schedule, storage, history and presentation. Flatkey is a separate commercial recommendation, with no competitor price or inventory records.
The GitHub Actions workflow starts once daily at 10:30 Asia/Shanghai (02:30 UTC). GitHub may queue scheduled runs. Tests, typechecking, a production build and a check that the deployed website supports the new dataset must succeed before data is committed. A failed core collection retains the last verified dataset; optional Azure/DigitalOcean/Verda/Vast.ai failures publish no rows for that source and remain explicitly failed. The comparison board and GPU directory include only fixed hourly prices with verified country/region applicability. This version's provider directory lists all ten configured sources, including failed sources and those with audit-only catalogs. Vast.ai is a bounded search of concrete offers, always labelled partial; country-only offers stay in the audit. Earlier observations stay in dated history and are not carried into a new current dataset.
The running website reads the verified public main/data/offers.json snapshot at request time. After committing data, GitHub Actions warms the public status endpoint and verifies both its SHA-256 and the homepage's snapshot marker against the exact committed file bytes. Once a compatible application version is deployed, a successful daily refresh needs no Railway deployment token, CLI login, desktop computer or daily application deployment. Requests check for updates at most once per 60 seconds per server process; GitHub/CDN propagation can take several minutes, so verification retries for up to six minutes. If an upstream read fails or is invalid, the server keeps its last verified snapshot or bundled fallback and exposes that state instead of claiming a current publication.
Daily collection is deduplicated after the 10:30 cutoff. A manual workflow run can explicitly enable force_refresh; a scheduled run cannot force a duplicate. Runs that skip collection still verify the live cloud snapshot. Codex's 10:45 follow-up is additional supervision and can retry a missing or failed workflow; it is not needed for normal cloud publication. Code changes still require an application deployment. See the daily runbook.
Rollout status, 2026-10-23: the nine-source cloud run collected prices successfully, then stopped because old tests depended on changing daily data. Those tests now use fixed fixtures. The new application deployment remains blocked by Railway upload/source-connection authorization errors; a successful ten-source live publication has not been verified. A runtime capability failure preserves the existing website/GitHub dataset, while the workflow still uploads the collection artifact if tests and the build passed.
Each successful refresh saves the publication-safe source result to data/public-prices.json, adapts it into data/offers.json, saves European Central Bank reference rates in data/fx-rates.json, and appends a compact snapshot to data/price-history.json. Prices display in USD with their original amount and currency alongside. Every conversion includes its rate, publication date and official source. Monthly prices stay USD per GPU-month; they are not divided by an assumed number of hours. The archive retains 180 days of original and converted amounts, units, FX metadata and source status.
Each validated refresh also produces one bilingual CSV per provider at data/exports/{provider}.csv and data/exports/manifest.json. The CSVs contain fixed hourly quotes with verified country/region applicability, preserve original-currency amounts, and add usd_gpu_hour_amount and usd_instance_hour_amount; pricing_conditions identifies the FX conversion. The manifest records source/export/exclusion counts, observation times, FX provenance and file hashes. A source with no qualifying quotes has a header-only CSV and an explicit reason in the manifest. All other source observations remain in the JSON audit. See CSV exports and publication checks.
bun install
bun run devOpen http://localhost:3000.
bun run typecheck
bun run build
PORT=3000 node .next/standalone/server.jsOr build and run the included Dockerfile.
The bundled fallback snapshot is read from its original JSON file and validated at server startup. outputFileTracingIncludes copies data/offers.json into the standalone image. Keep this file read as raw UTF-8: importing it as a JSON module can change the last binary digit of an FX number during compilation and invalidate its exact provenance.
Copy .env.example to .env.local and set:
NEXT_PUBLIC_SITE_URL: canonical production originNEXT_PUBLIC_SITE_NAME: final public brandNEXT_PUBLIC_FLATKEY_WEBSITE_URL: request-based Compute service overview (defaults tohttps://flatkey.ai/compute); current quote CTAs use this pageNEXT_PUBLIC_FLATKEY_COMPUTE_URL: console entry, reserved for console-specific links (defaults tohttps://console.flatkey.ai/compute/market)
Keep immutable GPU specifications separate from time-series offers. Each production offer should preserve its provider SKU, source currency, billing unit, original total price, normalized USD/GPU-hour, availability status, source URL, and observation time.
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-collector.txt
.venv/bin/python -m playwright install chromium
bun run data:refresh
bun run test
bun run typecheck
bun run buildOn Linux, install Chromium system dependencies with .venv/bin/python -m playwright install --with-deps chromium. On Windows, use .venv\Scripts\python.exe and set GPU_PRICE_MONITOR_PYTHON to that executable if needed. An existing installation can be selected with GPU_PRICE_MONITOR_PYTHON=/path/to/python. No .env is required for public collection.
The generated dataset preserves contact-sales entries, starting prices, membership requirements, Spot rates, component-only prices and offline promotions. The formal comparison excludes unknown geography, starting prices, monthly packages and component-only charges; those records remain in the source audit. Prices are converted to USD for display with original currency, hardware configuration and price basis preserved. Exact same-configuration member prices are preferred and labelled as requiring membership; ordinary prices remain in the original source data. Provider catalogs do not establish live inventory. See DATA.md and the daily runbook.
The library's price_scope, compute_countries and location_source survive as priceScope, computeCountries and locationSource in RackRate offers. A platform footprint does not become a country-specific price; regional comparisons require a quote explicitly tied to that region or zone.
See DATA.md for the freshness, normalization, and failure policy.
MIT
RackRate uses GA4 only. Set NEXT_PUBLIC_GA4_MEASUREMENT_ID to the web stream ID (G-...) before building; Next.js embeds public environment variables at build time. Analytics is loaded only after the visitor grants analytics consent. Events include page_view, price_filter_changed, flatkey_console_clicked, flatkey_website_clicked, and external_source_clicked. No Mixpanel requests are sent. Page URLs retain campaign parameters but exclude search text and other query parameters. Enhanced Measurement is disabled for this stream; the application sends page views, filter changes, and external-link events itself. Measurement ID: G-PP87NKBGK5.