⚡
Direct answer: GPUPLN (GPU Price Listing Notation) is a specification that defines
Open and royalty-freeGPU rental price data specification, current version
v0.1 (draft)It specifies which fields a quote must include and the format of each field
Measurement scope(per-GPU hourly rate, tax-inclusive or not, spot/preemptible vs pay-as-you-go), and data freshness thresholds. Anyone may freely implement and reference this; the specification itself is licensed under
CC BY 4.0 Published.
0. What problems does this specification address
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.
1. Conformance Levels
| Tier | Requirements | What It Means |
|---|
| L0 · Visible | Public HTML table listing GPU models and prices | Human-readable but machine-unfriendly |
| L1 · Crawlable | Provide a Stable, Login-Free Price Page or JSON API | Script supports batch data extraction |
| L2 · Understandable | Complete and Consistent Fields (including price_type / stock / fetched_at) | Cross-source comparable, no misinterpretation |
| L3 · Billable | Machine-readable pricing list + per-use billing interface (e.g., MCP + x402) | AI Agents can autonomously discover and pay for resources |
Among the 26 platforms currently tracked, 23 support fully automated scraping (L1), unified field output (L2), and per-use billing via MCP + x402 (L3). The remaining 3—TensorDock, Lanyun, and Paratera—lack publicly accessible, scrapable sources; their prices are manually displayed on the page.Clearly marked as “Reference Price · To Be Verified” and excluded from the lowest-price ranking。
2. Data Model
2.1 Offer (a single price quote)
| Field | Type | Required | Unit / Format | Metric Explanation |
|---|
platform | string | ✅ | kebab-case, e.g. gpushare | Platform Unique Identifier. The same platform must use the same ID across all implementations; otherwise, cross-source comparison is impossible. |
gpu_model | string | ✅ | Specification name, e.g. RTX 4090 | GPU Model Identifier. MandatorySame Name, Same Generation; modified or variant models with differing VRAM (e.g., 4090 48G, A100 40G) must not masquerade as standard models. |
price_hour_cny | number | ✅ | CNY per GPU per hour | Single GPUHourly price. For multi-GPU servers, divide total price by GPU count; for monthly plans, divide by 720 to convert to hourly equivalent and label accordingly. |
price_type | enum | ✅ | ondemand｜spot｜monthly | spot=spot/preemptible,ondemand=pay-as-you-go, bothNo mixed ordering。 |
stock | enum | ✅ | available｜scarce｜soldout | Inventory Status. Third parties must label this accurately and must not mask chronic stockouts with phrases like “In Stock”. |
billing_notes | string | ✅ | Plain text | Must Be SpecifiedData Sources and Billing Models(e.g., “official on-demand list price $2.25/GPU·hour converted, billed per minute”). |
src | enum | ✅ | crawler｜seed | crawler=This auto-scan succeeded;seed= Manually entered / fallback reference prices. |
fetched_at | string | ✅ | ISO 8601, including time zone | This price’sActual Acquisition Timestamp, not the moment the page is rendered. Timeliness is determined by this timestamp. |
data_status | enum | ➖ | stale｜suspect | stale=Not refreshed for over 7 days; for reference only;suspect=Fluctuation >40% vs. 7-day average; likely a data capture error,Not ranked。 |
2.2 GPU (GPU model)
| Field | Type | Required | Description |
|---|
name | string | ✅ | Specification name, corresponding to Offer.gpu_model |
vram_gb | integer | ✅ | VRAM Capacity (GB) |
tier | string | ✅ | Tiered Positioning, e.g., “Consumer Flagship” or “Data Center Training GPU” |
use | string | ✅ | One-sentence typical use case |
2.3 Platform
| Field | Type | Required | Description |
|---|
id | string | ✅ | Corresponds to Offer.platform |
name | string | ✅ | Display Name |
region | string | ✅ | Primary service region, e.g., “China”, “North America” |
url | string | ✅ | Official Website |
3. Measurement Criteria (Mandatory)
- 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.
4. Timeliness and Obsolescence Assessment
| Status | Evaluation rules | Display Requirements |
|---|
| Real-time | src=crawler and this crawl succeeded (fetched_at On the Same Day | Eligible for Real-time Lowest Price Ranking |
| Reference Price | fetched_at Age exceeds 7 days | Marked as “Reference Price · To Be Verified”Not included in ranking |
| Suspected anomaly | 7-day average price volatility for this GPU model exceeds 40% | Not ranked, alerts only logged for manual review. |
⚠️ Presenting expired data as real-time pricing is the most serious breach of trust for a price comparison siteTherefore, stale detection is decoupled from the data source field: even if the last crawl succeeded (src=crawler), any entry unrefreshed for over 7 days is automatically downgraded to reference pricing.
5. Machine-Readable Distribution
| Channel | URL | Use case |
|---|
| JSON Feed | https://rentgpu.link/api/prices.json | Full quote set, refreshed every 45 minutes, no API key required |
| Selection interface | https://rentgpu.link/api/resolve?gpu=H100&budget=10&hours=8 | Return the Best-Matching Platform Based on GPU Model + Budget |
| MCP endpoint | https://rentgpu.link/mcp | Agent tool discovery and invocation |
| Self-describing inventory | /.well-known/mcp.json、/.well-known/x402.json | Automated machine recognition capability |
| AI summary | https://rentgpu.link/llms.txt | Fact list scraped by AI crawler |
6. Attribution Requirements (CC BY 4.0)
来源：rentgpu.link（GPU computing power radar），Neutral GPU rental比价站
Data更新Time：2026-10-27 21:00 (UTC+8)
引用格式：rentgpu.link（GPU computing power radar）. "RTX 4090 Price Comparison". 2026-10-27.
https://rentgpu.link/gpu/rtx-4090 （CC BY 4.0）
必须保留：①来源名 ②DataTime ③link。price为Automated crawlingof挂牌价，实际Order page takes precedence。Permitted: redistribution, remixing, commercial use, AI training and inference citation. Requirement:Attribution + data timestamp + original URL, and prices are authoritative per the order page.
7. Change Log
| Version | Date | Changes |
|---|
v0.1 | 2026-10-27 | Draft v1: Three core objects — Offer / GPU / Platform; field definitions; consistency levels L0–L3; stale and suspect detection rules; CC BY 4.0 attribution requirements. |
8. Governance and Participation
- 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.
rentgpu.link
A neutral GPU compute rental price comparison tool: Public prices from 26 global platforms are collected every 45 minutes. We own no GPUs and sell no compute.