GPU rental prices

GPU opens GPU Rental Prices. GPU H100 or GPU B200 selects a model. Pro unlocks the full board, dated history and Changes. Free access, including signed out, previews three latest rows; history filters cannot unlock older preview values.

The Board shows published rates in USD per GPU-hour, their basis, availability where stated, and observation time. History charts real dated observations, beginning at the first point. Step lines connect source observations; markers show actual readings. Archive captures are marked archived page, reconstructed, official offer versions are official published history, and live collections are live observation. Changes records price moves, explicit availability transitions and changes in the providers included in a list median. Equities places related shares beside a GPU's seven-day move; Enter opens a ticker and t opens The Buildout (TBO). The standard pane refresh reads the latest stored board; it does not trigger another scrape.

Methodology

History groups these indices under Reference, with muted, distinct chart colours. They are context, not Gloom list medians or completed rental prices, and never enter our price baskets. Reference A exposes a rolling three-month daily history for five public GPU types; Reference B exposes seven-day anonymous chart cards across seven GPU series. Older reference data requiring login is not accessed. Pro unlocks full stored reference history; free returns up to three latest reference values. Collection is disabled by default. Source hosts and fetch timestamps remain private server-side evidence, never public links.

Changes distinguishes moves between Gloom observations from published transitions. For example, Nebius's page explicitly provides the old and new rates and their October 1, 2026 effective date. That announced change can appear on the first collection day, tagged as published, without creating observations for days before collection began.

All rates are normalized to USD per GPU-hour. A node's hourly price is divided by its stated GPU count; an eight-GPU node therefore contributes one eighth of the node rate. SXM and PCIe remain separate, as do memory variants such as A100 40GB and A100 80GB. AWS p4d is the 40GB variant. OCI H100 and H200 are mapped to their SXM capacities of 80GB and 141GB. Missing H100 SXM memory is 80GB across feeds; H100 NVL remains a separate 94GB PCIe variant. Unambiguous missing fields also use A10 PCIe 24GB, L40/L40S PCIe 48GB, MI300X OAM 192GB and B200 SXM. Explicit source capacities remain unchanged, and an H200 label without a form factor stays separate from SXM. Node prices include the host's bundled CPU, RAM and interconnect where the provider bundles them.

One rate per provider, GPU variant and basis contributes to the board and list median. Flagship eight-GPU SXM nodes take precedence where available; list comparisons use the lowest-priced US region among the regions sampled. Global rate cards retain their global region. US region coverage is limited by the public endpoints and SKU availability, not exhaustive geographic coverage. Live AWS list collection uses small Linux on-demand price maps. The explicit history importer streams official monthly EC2 offer versions for us-east-1, discards bulk payloads and keeps only eligible shared-tenancy Linux on-demand GPU rows. Each observation retains the file publication timestamp, original product SKU and rate evidence. Historical regional coverage is narrower than the current US sample. AWS spot uses the median across regions reporting a given instance, with region count, minimum and maximum retained. Different AWS instances with the same GPU are reduced to a single provider observation for that variant and basis.

AWS list samples Northern Virginia, Ohio, Northern California and Oregon. Azure selects among the US regions returned by its paged API. The H100 board and hyperscaler list median use the RDMA-equipped ND96isr_H100_v5 family (including isrf), comparable with AWS p5 and GCP a3-highgpu-8g. The sampled node rate is $98.32, or $12.29 per GPU-hour. Cheaper non-RDMA ND96is, noIB and flex H100 variants are omitted from both list and spot comparisons. GCP uses the page's Iowa (us-central1) selection, verified alongside its headers. CoreWeave uses its North America rate card; other unregionalized pages remain unspecified or global rather than being presented as a particular US location. Azure's GB200 VM has four GPUs per node, according to its published VM specification.

Hyperscaler and neocloud list medians are separate. Each includes its constituent providers, count, minimum and maximum. When membership changes, the series is chained using the continuing providers so adding or removing a provider does not create a price jump; a membership event records the change. A chained value can differ from the current unadjusted median, which is also retained. Marketplace ask quantiles are separate from these list series. Aggregator offers use the underlying cloud when it is identified.

GCP columns are matched against the page's visible headers: Price (USD) or On-Demand (USD) is list, Current Spot pricing is spot. DWS Flex-start, DWS Calendar Mode and committed-use discounts are different columns. An unavailable list cell stays unavailable. Together uses the GPU Clusters table rather than dedicated inference pricing. Nebius selects the rate column whose stated effective date has arrived, using UTC dates.

Collection and history

Reads are sequential with a declared User-Agent. Failures, rate limits and server errors back off. Each source has an independent enable switch. A failed parse leaves the last good observations intact and reports the failure; observations are marked stale after 48 hours without success. Disabling a source removes it from newly served results.

Gloom stores at most one live observation per source/SKU/hour and retains a daily observation even when prices do not change. Historical imports retain every distinct source/SKU/exact observation timestamp. Repeating an import is idempotent; a corrected historical point supersedes the current value while retaining its previous revision. Late imports recompute neighboring Changes events, including the transition into live history. Changes suppresses differences smaller than USD 0.000001 per GPU-hour, avoiding false moves when live and bulk endpoints round the same rate differently. A missing SKU is not evidence that capacity became unavailable: availability events require two explicit, differing source states.

The 1D, 7D and 30D columns compare against real earlier observations at or before the target time and within two hours of it. Imported observations can supply those anchors, but a sparse monthly archive cannot establish a daily price. Gaps stay -; no calendar days, prices or historical aggregates are invented. Archive rows never enter a live aggregate merely because a capture is recent. History steps connect real observations and do not create intermediate records.

A provider's effective date does not establish that Gloom observed its price on that date. Azure's API-provided effective dates remain in effectivePoints, separately from observed history, with official-history provenance. They cannot reconstruct intervening prices or past cheapest-region choices.

The explicit Azure meter importer enumerates NC, ND, NV and NG virtual-machine catalog entries across regions and keeps verified Linux meter/SKU/region combinations separate. On-demand, Spot and Low Priority meters retain their native identities; Windows and licensed-software prices are excluded. Unsupported hardware allocations are rejected. These are current meter rates with published effective dates, not a historical revision tape. Fractional GPU allocations retain their documented shares and allocated GPU memory. Pro REST and MCP callers can request includeFractional=true to include them in effectivePoints; default responses contain whole-GPU rates for compatibility with existing clients. This option does not change observed history or free previews.

The archive CLI queries one provider URL and calendar year at a time, selects at most one capture per month, and limits each CDX result to 50 entries. It covers the existing Lambda, Nebius, CoreWeave, Crusoe, DigitalOcean, Together, GCP and OCI adapters, including Lambda's old domain. Failed annual queries fall back to monthly Archive availability lookups and exact-timestamp CDX digest checks. Raw pages are fetched only for previously unseen digests. Retries use persisted exponential backoff with jitter, capped at five minutes; longer server Retry-After instructions are honored. Providers rotate so one failing source does not starve the others. Robots restrictions remain enforced.

Every imported point uses the exact capture date and retains the original URL and raw archive evidence URL. Layouts that cannot be parsed yield no observations. Saved pages can be reparsed offline after an adapter correction; compact reviewed seeds load without fetching remote pages. Archive failures and rate limits retain a resumable cursor and explicit source coverage; an unavailable archive never becomes a zero price. GCP's official catalog history requires API caller credentials, and OCI's current catalog update time is not a GPU-price effective date; neither is used to invent historical observations.

The AWS list CLI selects one official archived offer version per catalog month back to 2023, with actual publication dates on points. There are no purchases or authenticated calls for list history. The separate AWS DescribeSpotPriceHistory importer requires GPU_AWS_SPOT_HISTORY_ENABLED=1 and explicit credentials, keeps instance/availability-zone series separate, and has no keyless fallback. It is never scheduled and was not run in this validation. The historical importers exclude Vast, RunPod, Shadeform, Hyperbolic, SF Compute, reference indices and the keyless AWS spot feed. Reference indices use their own anonymous daily collector.

Imports are explicit commands, limited to one request per second, with a hard total cap per resumable cursor (300 archive requests with a 200-request default, 120 AWS requests, 60 Azure meter requests). Dry runs never connect to a database. Writes require an explicit scratch database URL or both remote-write acknowledgements. Cursors are locked against concurrent runs; after a crash, verify the importer has stopped before removing its .lock directory. Raw source pages and offer records remain outside the database. No automatic compaction or retention drops historical evidence; corrections are retained.

These are published rates and asks, with differences in region, hardware, service terms and availability. They do not establish realized utilization, transaction volume or rental proceeds. No availability is inferred from a low price.

Commands

API, MCP and headless outputs preserve pricing basis, observation/effective dates and provenance labels. Provider observations include evidence URLs; anonymised reference indices never include source URLs or private evidence. Provider rate cards and offer files establish published prices, not completed rentals. Sources that cannot supply real dated prices contribute no historical observations.