Product spec

Each product publishes USD per GPU-hour (for each vote and for the product, as the median of the vote medians) and USD per dense BF16 PFLOP-hour for the product. The conversion is USD per GPU-hour times 1,000, divided by the product's dense BF16 TFLOPS.

Throughput is the datasheet dense (no structured sparsity) BF16 tensor rate with FP32 accumulation. Arithmetic uses 12-decimal fixed point, divisions round half up, and the median of an even count is the half-up mean of the two middle values. Values are never rescaled across methodology versions.

List asking price

Inputs are the on-demand prices providers currently show. They are not completed rentals, fills, or realized rates.

On-demand, USD

Spot, interruptible, reserved, and committed prices do not match. Quotes flagged as unconverted currency or daily-resolution FX are excluded.

Peak, not sustained

The divisor is the datasheet peak dense rate. Clocks, power, cooling, interconnect, and workload are not measured.

Memory-blind

The unit prices BF16 throughput only. Memory enters solely through exact product identity, so an H100 80 GB never mixes with an H100 NVL 94 GB.

Per GPU as sold

Host CPU, RAM, storage, network, and region are recorded but never adjust the price or gate eligibility. A quote with an incomplete bundle still votes. Egress, taxes, and discounts are not added or removed.

Total price required

A provider that publishes only a GPU-hour component, with no total price, is rejected for that quote.

The panel is the nine products below. Each product is published on its own round. The composite is the equal-weight mean of the normalized values of the usable products, with no weighting by volume, spend, or votes.

A composite needs at least 7 of the 9 products usable, and averages whichever 7, 8, or 9 are. Below 7 no composite is produced. A product counts only if its publication identity (methodology, policy, hardware registry, publication policy, composition version) matches the composite's.

1. IdentityCanonical manufacturer, model, and variant equal the registry entry, and the rental type is on-demand.
2. GPU countExactly 1 GPU, except H200 SXM5, which also admits 8.
3. AvailabilityStatus must be available or limited. Unknown quotes never vote. An unavailable quote votes only as a sold-out vote: a direct provider's current posted price counts for up to 6 hours (21,600 seconds) after that provider last had a counted in-stock vote for the product in a published valid round, and only while the provider has no accepted quote for it. Marketplaces, aggregators, and bid sources never get sold-out votes.
4. QualityQuotes flagged currency-unconverted, possible-duplicate, hardware-time-mismatch, fx-daily-resolution, or hardware-variant-unknown are excluded. The variant flag excludes only products that have a variant, so the four GeForce products ignore it.
5. SourceAggregator quotes need a verified named upstream supplier. Direct providers must not name a different upstream. Marketplaces need a complete order book, one quote per machine, verified hosts, and at least 5 machines on 3 hosts, apart from named exceptions.

H200 SXM5 eight-GPU division

An eight-GPU H200 SXM5 quote is admitted only when its price proves to be a USD server price: an instance, virtual machine, node, or server hour whose raw amount equals 8 times the per-GPU price. The voting price is the server price divided by 8.

RunPod queries 8-GPU node-hour prices for Secure and Community Cloud, takes the uninterruptible price as a node total, and divides by 8. One-GPU quotes use the published per-GPU price unchanged. All tiers and counts feed RunPod's single vote median.

How votes form

A direct provider or marketplace is one vote, however many listings it has. An aggregator is one vote per verified upstream supplier. A vote's price is the median of its quotes. Votes that share an overlap group are collapsed to one, with direct supply taking priority over marketplaces, then aggregators. E2E Networks and Jarvis Labs count once, as do DigitalOcean and Paperspace.

Quorum

A product round needs 3 independent counted votes, 3 distinct collecting providers, and at least 1 in-stock vote. They are separate tests: an aggregator can add several votes but is one collector, and a round made only of sold-out votes has no value. Every round shows how many of its counted votes were in stock and how many were sold out.

Provider freshness

Each provider contributes its latest complete snapshot within its own freshness limit, between 300 and 2,700 seconds, listed below.

Staleness and retention

If a new attempt fails quorum, the last valid round is served while it is under 1,800 seconds old and its identity matches. At 1,800 seconds the product is stale and leaves the composite.

Same for every product

Availability, quorum, freshness, retention, and composite rules do not vary by product.

Throughput comes from the hardware registry dense-bf16-v0.6.1. The last column is the price of one PFLOP-hour at a given $1 per GPU-hour.

Only providers marked for the canonical index can vote: 28 are marked today, but CUDO, Northflank, Scaleway, Shadeform, and Vast never vote at this version, so 23 can. Eligibility does not mean a provider quotes every product; that depends on its catalog at the time. Freshness is the number of seconds before a provider's snapshot is ignored. Any product can sit at exactly the quorum at any time; the live counts are in GET /v1/products/current as data[].currentRound.voteCount and collectingProviderCount.

Coverage only, never voting: GrabGPU, Novita, OctaSpace, Gcore, and Salad.