MARKET PULSEH100 $4.18/GPU-HR 3PH200 $4.47/GPU-HR 1PMI300X $2.59/GPU-HR 1PL40S $1.62/GPU-HR 2PB300 $16.0/GPU-HR 1POBSERVED 2026-09-03 · 1H CACHE
COMPUTELEDGER.FOR AGENTS ↗

OUR STANDARD / VERSION 2.2

Trace every
number.

Compute is too fragmented for false precision. We publish the unit, observation time, check time, provenance, rights and limitations behind each signal.

Primary sources first

We prefer licensed or openly reusable machine feeds, exchange notices, index-provider methodologies, regulatory filings, investor-relations releases and direct product documentation. Secondary reporting is used for context and clearly labeled.

Two clocks, not one

Checked at records when Compute Ledger last queried a source. Observed at records when the upstream measurement was produced. The public market API is refreshed on demand with an hourly cache, while its current CC-BY price source publishes daily. Calling that data “real-time” would be misleading, so the interface exposes both clocks and marks delayed or fallback states.

Units before headlines

Public GPU plan prices are normalized to US dollars per individual GPU-hour. When a plan contains multiple GPUs, its hourly total is divided by the source’s reported GPU count. Source-plan totals include the complete upstream catalog; normalized-offer totals include only rows with an explicit GPU model, count and hourly price. We do not silently combine public list prices, negotiated contracts, spot capacity or executed-transaction indices.

WHAT THE MEDIAN MEANS

A family median is calculated across the currently observed qualifying plans after per-GPU normalization. Different memory, interconnect, region and service levels remain basis risk, so the minimum, maximum and provider count are shown beside it.

Azure Spot retail tape

The Azure tape queries Microsoft's unauthenticated Retail Prices API for Consumption meters containing “Spot” on two exact Linux VM SKUs: Standard_NC40ads_H100_v5 and Standard_NC24ads_A100_v4. Windows, Dev/Test, Low Priority, reservation and non-Spot rows are excluded. For each SKU and region, the latest effective tariff that is not future-dated is retained.

Microsoft documents those VMs as containing one H100 NVL 94GB and one A100 PCIe 80GB respectively. The published VM-hour retail price is divided by that documented GPU count and rounded to six decimal places. Observed at is our API query time; effective start date is Microsoft's tariff date. Neither field proves that capacity can be provisioned.

Microsoft also states that net-new capacity is currently being deployed only for the NCads H100 v5 generation. The A100 tariff is therefore published with an explicit legacy, capacity-constrained warning; a listed tariff does not establish available A100 supply.

CAPACITY IS NOT A QUOTE

Azure Spot pricing varies by region and SKU and remains subject to quota, capacity and interruption. The tape is not executable, tradeable, availability-guaranteed or a record of transactions. It covers two representative single-GPU SKUs rather than the full Azure GPU catalog.

CL-FCR: Frontier Compute Rate

CL-FCR is the median of every eligible normalized offer in the current public-cloud snapshot across the H100 family, H200, MI300X and B300. Each plan total is divided by its reported GPU count before the offers are ordered; the middle value, or mean of the two middle values, becomes the published USD-per-GPU-hour rate.

It is a current-snapshot reference rate, not a time-series return index. It is not an investable or transaction-based benchmark, an executable quote, or the settlement price of any derivative. Hardware mix, geography, availability, service terms and changing provider coverage create basis risk. The API therefore publishes the range, provider count, offer count, constituents, eligible model strings and observation time beside the value.

Rights are part of the data

GPU Cloud Compare’s open dataset is embedded under CC-BY 4.0 with attribution. ORNN’s hourly transaction-based index is a valuable institutional source, but ORNN states that third-party redistribution requires a commercial agreement. ORNN is therefore documented but not connected or republished. Silicon Data’s API likewise requires an eligible subscription.

News versus analysis

Official NVIDIA, AMD and AWS RSS feeds are checked through an hourly cache. Intel and Google Cloud remain listed editorial references but are not described as connected feeds because reliable official RSS endpoints are unavailable. We filter for compute-market relevance and display only the headline, publication date, source and canonical link; article bodies are not copied. Publication recency is calculated separately from connector health. Compute Ledger analysis is separately labeled. Planned products remain pending until their venue and regulator confirm launch.

Corrections

Material changes update the timestamp and preserve a clear correction note. The public corrections standard and current submission status are maintained on the about page.

Machine use

The JSON endpoint uses stable field names and ISO 8601 timestamps. llms.txt documents canonical resources; the source registry exposes rights and cadence; source health reports degraded feeds; RSS orders updates. Agents should retain attribution, license, source URL, observed time, checked time and freshness status when citing the dataset.