The GPU rental market hosts dozens of platforms, each with differing page structures, units, and billing models: some quote full-server rates while others quote per-GPU; some bill hourly, others per minute; some mix spot and pay-as-you-go pricing in the same view. The result is—The same H100 can cost up to twice as much across different websites, not because the difference is genuinely double, but becauseDifferent Metrics。

This specification does not require any platform to change its prices—only thatClarify the Definition: Within a single quote, clarity is required on what exactly the price covers, when it was captured, and whether it remains reliable. Only then does cross-platform price comparison, AI-driven referencing, and historical tracking become meaningful.

• Per GPU per Hour: All prices are normalized to RMB per GPU per hour. For monthly bare-metal rentals: total monthly fee ÷ GPU count ÷ 720 hours.
• Separate Pay-as-You-Go and Spot Pricing：ondemand and spot These are two products with entirely different risk levels—spot prices must not be mixed in the same ranking, as spot instances may be reclaimed at any time.
• GPU Model Naming Convention: Only identically named and generation-matched GPUs are considered equivalent. Modified variants—e.g., RTX 4090 48G—must be explicitly labeled and excluded from standard model price slots.
• Exchange Rate: All foreign-currency prices are converted using fixed rates: 1 USD = 7.2 CNY, 1 EUR = 7.8 CNY, and noted in billing_notes lists original currency and original price for verification.
• Timestamp：fetched_at Records the actual time when the price was obtained, not the page generation time.

• Open: The specification and implementation are fully open-sourced—no patents claimed, no licensing fees.
• Neutral: Ranking is strictly price-based; paid promotion slots are separately marked and never included in rankings.
• Falsifiable: This site openly publishes its data collection success rate, coverage duration, and stale-quote ratio—any claim of “full real-time coverage” lacking verifiable coverage metrics is untrustworthy.
• Correction channel: Report pricing errors for correction; verified updates appear within 24 hours with full audit trail.

A specification earns the title of standard only when others actually implement it. Thus, GPUPLN’s adoption mechanism is intentionally lightweight:No application required, no authorization needed, no need to contact us：

1. Replace /.well-known/gpupln.json Copy it to the same path under your own domain (the file itself is the template).
2. Replace publisher with your name, homepage, and contact email; replace endpoints.prices with your own pricing feed URL.
3. Make your feed pass gpupln/schema.json Validation (JSON Schema draft 2020-12 — readable directly by any language’s validation library).
4. Updated each time data is refreshed updated_at。

Complete these four steps, and your data is declared compliant with GPUPLN v0.1 — automatically discoverable and aggregatable by rentgpu.link and other compliant parties. rentgpu.link’s own feed has passed the above schema validation (tested on 2026-10-28: all 112 price entries validated successfully).

How price data is published and what payment method buyers use are two separate matters. GPUPLN standardizes only the former; the latter resides in an optional payment In the section,Pluggable— Implementers may use any payment protocol—or none at all.

• payment.protocols：array, values x402 / mpp / ap2 / other, multiple entries allowed.
• payment.network、payment.asset、payment.price_per_call：billing track, assets, and unit price.
• Not filled payment The section does not affect compliance with the attestation—free and publicly available data sources may legitimately omit a payment section entirely.

rentgpu.link’s own attestation document is currently active x402(Base mainnet-native USDC, $0.05 per use), and authorization The section declares support for AP2 mandate as the authorization layer. We avoid protocol lock-in: the spec defines only how to list, not which protocol must be listed.