Measured workload economics
Verified GPU cost benchmarks
Convert exact MLPerf 6.0 results into current cloud economics. See cost per million reported tokens, cost per batch unit and cost to reach a published training quality target—without transferring the result to a different model or hiding the assumptions.
- Official MLCommons source paths pinned to full commits
- 20 raw training runs retained by SHA-256 hash
- Exact Nebius 8× B300 system and current official price match
Transparent repricer
Change the rate, never the published result
The benchmark throughput or time stays locked. Change only the per-GPU price and optional target volume to audit the arithmetic.
Scenario summary is ready to copy.
Exact published system
Published reference economics
Every result below uses the same Nebius eight-GPU B300 system and is re-priced with the current official on demand $7.85 per-GPU-hour catalog rate, equal to $62.80 per hour for the exact eight-GPU reference system. This is a transparent reference calculation, not a promise for another workload.
batchOffline
qwen3-vl-235b-a22b
$0.2229 / 1K samples
- Published result
- 78.28 Samples/s
- Reference time
- 12.78 s / 1K samples
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 448.72K samples
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 1 verified result
- Source revision
4d3916ac9cf4
F1_HIERARCHICAL: 0.7880721729948121
Inspect source evidence ↗batchServer
qwen3-vl-235b-a22b
$0.3864 / 1K queries
- Published result
- 45.15 Queries/s
- Reference time
- 22.15 s / 1K queries
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 258.83K queries
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 1 verified result
- Source revision
4d3916ac9cf4
F1_HIERARCHICAL: 0.7865237675409595
Inspect source evidence ↗inferenceOffline
deepseek-r1
$0.2517 / 1M reported tokens
- Published result
- 69,318.9 Tokens/s
- Reference time
- 14.43 s / 1M reported tokens
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 397.37M reported tokens
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 1 verified result
- Source revision
4d3916ac9cf4
exact_match: 81.58614402917047 TOKENS_PER_SAMPLE: 3722.1116681859617
Inspect source evidence ↗inferenceServer
deepseek-r1
$0.2888 / 1M reported tokens
- Published result
- 60,413.4 Tokens/s
- Reference time
- 16.55 s / 1M reported tokens
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 346.32M reported tokens
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 1 verified result
- Source revision
4d3916ac9cf4
exact_match: 81.58614402917047 TOKENS_PER_SAMPLE: 3721.7894257064722
Inspect source evidence ↗inferenceOffline
gpt-oss-120b
$0.1632 / 1M reported tokens
- Published result
- 106,885 Tokens/s
- Reference time
- 9.36 s / 1M reported tokens
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 612.72M reported tokens
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 1 verified result
- Source revision
4d3916ac9cf4
exact_match: 82.959
Inspect source evidence ↗inferenceServer
gpt-oss-120b
$0.1737 / 1M reported tokens
- Published result
- 100,437 Tokens/s
- Reference time
- 9.96 s / 1M reported tokens
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 575.75M reported tokens
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 1 verified result
- Source revision
4d3916ac9cf4
exact_match: 83.337
Inspect source evidence ↗trainingTime to train to quality target
GPT-OSS 20B pretraining
$87.3772 / reference run
- Published result
- 5,008.89 seconds
- Reference time
- 83.48 min to published quality target
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 1.14 reference runs
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 10 verified runs
- Source revision
eabf23a07b2a
C4 · 3.34 log perplexity
Inspect source evidence ↗trainingTime to train to quality target
Llama 3.1 8B pretraining
$75.3729 / reference run
- Published result
- 4,320.74 seconds
- Reference time
- 72.01 min to published quality target
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 1.33 reference runs
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 10 verified runs
- Source revision
eabf23a07b2a
C4 · 3.3 log perplexity
Inspect source evidence ↗Evidence boundary
What this proves—and what it does not
These are results for a published Nebius system with eight NVIDIA B300 accelerators, a named model, scenario, software stack and quality result. The economics layer multiplies the current official per-GPU price by eight and applies only the published throughput or time.
Machine-readable evidence