Benchmark-backed workload costs

GPU workload cost evidence

Move from hourly price to completed-workload economics. Each page joins a pinned peer-reviewed benchmark result to a compatible current cloud price without blending their evidence clocks.

Browse verified pages Publishing methodology Full interactive database

Canonical decision inventory

Every page passes the evidence gate

The inventory is rebuilt from official-source records. A page needs enough configurations, regional or provider diversity and distinct price points for its exact intent. It is not published merely because a keyword can be formed.

Training economics

Use time-to-quality profiles to estimate compute cost per benchmark-equivalent training run at current compatible rates.

workload cost2 rates

gpt-oss 20b pretraining time to train to quality target gpu cost

$7.850–$7.850 per GPU-hour · 2 regions.

Pinned benchmark result: 5,008.8865 seconds · Time to train to quality target.

Completed-workload cost range: $87.3772 per benchmark-equivalent run on the pinned 8× NVIDIA B300-SXM-270GB system across the current compatible observed rates.

Serial benchmark-equivalent time: 83.48 min per benchmark-equivalent run serially on the pinned 8× NVIDIA B300-SXM-270GB system.

Deployability evidence: 2/2 rows (100%). This is not a live-stock claim.

Open evidence sheet Inspect current NVIDIA B300-SXM-270GB rates Inspect benchmark source ↗

workload cost2 rates

llama 3.1 8b pretraining time to train to quality target gpu cost

$7.850–$7.850 per GPU-hour · 2 regions.

Pinned benchmark result: 4,320.74175 seconds · Time to train to quality target.

Completed-workload cost range: $75.3729 per benchmark-equivalent run on the pinned 8× NVIDIA B300-SXM-270GB system across the current compatible observed rates.

Serial benchmark-equivalent time: 72.01 min per benchmark-equivalent run serially on the pinned 8× NVIDIA B300-SXM-270GB system.

Deployability evidence: 2/2 rows (100%). This is not a live-stock claim.

Open evidence sheet Inspect current NVIDIA B300-SXM-270GB rates Inspect benchmark source ↗

Inference throughput economics

Use serving throughput profiles to estimate cost per one million measured outputs while keeping the pinned benchmark system boundary visible.

workload cost2 rates

deepseek-r1 offline gpu cost

$7.850–$7.850 per GPU-hour · 2 regions.

Pinned benchmark result: 69,318.9 Tokens/s · Offline.

Completed-workload cost range: $0.2517 per 1M tokens on the pinned 8× NVIDIA B300-SXM-270GB system across the current compatible observed rates.

Serial benchmark-equivalent time: 14.43 s per 1M tokens serially on the pinned 8× NVIDIA B300-SXM-270GB system.

Deployability evidence: 2/2 rows (100%). This is not a live-stock claim.

Open evidence sheet Inspect current NVIDIA B300-SXM-270GB rates Inspect benchmark source ↗

workload cost2 rates

deepseek-r1 server gpu cost

$7.850–$7.850 per GPU-hour · 2 regions.

Pinned benchmark result: 60,413.4 Tokens/s · Server.

Completed-workload cost range: $0.2888 per 1M tokens on the pinned 8× NVIDIA B300-SXM-270GB system across the current compatible observed rates.

Serial benchmark-equivalent time: 16.55 s per 1M tokens serially on the pinned 8× NVIDIA B300-SXM-270GB system.

Deployability evidence: 2/2 rows (100%). This is not a live-stock claim.

Open evidence sheet Inspect current NVIDIA B300-SXM-270GB rates Inspect benchmark source ↗

workload cost2 rates

gpt-oss-120b offline gpu cost

$7.850–$7.850 per GPU-hour · 2 regions.

Pinned benchmark result: 106,885 Tokens/s · Offline.

Completed-workload cost range: $0.1632 per 1M tokens on the pinned 8× NVIDIA B300-SXM-270GB system across the current compatible observed rates.

Serial benchmark-equivalent time: 9.36 s per 1M tokens serially on the pinned 8× NVIDIA B300-SXM-270GB system.

Deployability evidence: 2/2 rows (100%). This is not a live-stock claim.

Open evidence sheet Inspect current NVIDIA B300-SXM-270GB rates Inspect benchmark source ↗

workload cost2 rates

gpt-oss-120b server gpu cost

$7.850–$7.850 per GPU-hour · 2 regions.

Pinned benchmark result: 100,437 Tokens/s · Server.

Completed-workload cost range: $0.1737 per 1M tokens on the pinned 8× NVIDIA B300-SXM-270GB system across the current compatible observed rates.

Serial benchmark-equivalent time: 9.96 s per 1M tokens serially on the pinned 8× NVIDIA B300-SXM-270GB system.

Deployability evidence: 2/2 rows (100%). This is not a live-stock claim.

Open evidence sheet Inspect current NVIDIA B300-SXM-270GB rates Inspect benchmark source ↗

Batch throughput economics

Use offline and batch throughput profiles to estimate cost per one million measured outputs at the current compatible cloud rate.

workload cost2 rates

qwen3-vl-235b-a22b offline gpu cost

$7.850–$7.850 per GPU-hour · 2 regions.

Pinned benchmark result: 78.2775 Samples/s · Offline.

Completed-workload cost range: $222.8539 per 1M samples on the pinned 8× NVIDIA B300-SXM-270GB system across the current compatible observed rates.

Serial benchmark-equivalent time: 3.55 h per 1M samples serially on the pinned 8× NVIDIA B300-SXM-270GB system.

Deployability evidence: 2/2 rows (100%). This is not a live-stock claim.

Open evidence sheet Inspect current NVIDIA B300-SXM-270GB rates Inspect benchmark source ↗

workload cost2 rates

qwen3-vl-235b-a22b server gpu cost

$7.850–$7.850 per GPU-hour · 2 regions.

Pinned benchmark result: 45.1513 Queries/s · Server.

Completed-workload cost range: $386.3553 per 1M queries on the pinned 8× NVIDIA B300-SXM-270GB system across the current compatible observed rates.

Serial benchmark-equivalent time: 6.15 h per 1M queries serially on the pinned 8× NVIDIA B300-SXM-270GB system.

Deployability evidence: 2/2 rows (100%). This is not a live-stock claim.

Open evidence sheet Inspect current NVIDIA B300-SXM-270GB rates Inspect benchmark source ↗

Autonomous, not indiscriminate

How this index stays useful

Publish

A new URL becomes indexable only after its current official data passes family-specific breadth, freshness, source and uniqueness thresholds.

Update or noindex

Changed evidence triggers an update. A formerly qualified route that fails a cycle is preserved for users but immediately removed from search indexing.

Retire

After three consecutive failed cycles the route is retired. Canonical migrations are merged explicitly, preventing uncontrolled duplicates and crawl waste.