Workload price snapshot:Checking the declared provider-SLA boundary.Valid through ; historical after that instant.
Pinned benchmark × bounded price snapshot
deepseek-r1 Server cost
Reprice this verified deepseek-r1 Server result with an exact Nebius AI Cloud B300 price deadline, formula and pinned evidence. The benchmark result and cloud price keep independent provenance, so a price refresh never rewrites the measured performance.
Reproducible workload economics
Scale the measured result
The published throughput is converted into elapsed cluster-hours per one million measured units, then multiplied by the matched price-snapshot cluster rate. Change only the number of equivalent runs or output batches; the evidence-backed performance stays fixed. The primary estimate uses the lowest compatible observed rate; the range shows sensitivity across all compatible published rates and does not imply capacity.
Validate the workload boundary
Check this estimate against your workload and quotes
A human review tests benchmark transferability, workload assumptions, capacity evidence and non-compute costs while keeping unknowns explicit.
Measured boundary
Exactly what was benchmarked
Nebius B300 n1 (8x B300-SXM-270GB, TensorRT) reported 60,413.4 Tokens/s for deepseek-r1 in the Server scenario. The quality target was exact_match: 81.58614402917047 TOKENS_PER_SAMPLE: 3721.7894257064722. Software was TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.0. This page does not extrapolate to a different model, implementation, accelerator count or quality target.
Interpretation guardrails
Use this estimate inside its evidence boundary
Comparable workload
The cost applies to this named benchmark, scenario, software stack, quality target and published accelerator system. A production workload can produce different throughput or time.
Serial time only
The displayed time scales benchmark-equivalent runs or output batches serially on the measured system. It excludes queueing, provisioning and any parallel execution strategy.
Compute line only
Storage, networking, data preparation, queueing, engineering time, failed jobs, support and taxes are excluded. Add them in a complete deployment budget.
Capacity separate
The price source proves a published rate, not immediate capacity. Availability remains a separate signal and must be confirmed before scheduling work.
- The result applies only to the published model, software, scenario and 8-GPU system.
- The current Nebius normalized per-GPU rate is joined separately and multiplied by the benchmark GPU count; identical topology and simultaneous capacity are not asserted.
- Storage, data transfer, orchestration, queueing, support, taxes and discounts are excluded.
- A benchmark result is not a guarantee of performance for a different workload or configuration.
Related evidence
Change one decision variable
Move to another scenario or inspect the exact provider and accelerator pricing behind this result.
workload cost
gpt-oss 20b pretraining time to train to quality target gpu cost
3 published official-source rows. Open the page to inspect its own price range and validity deadline.
Open decision pageworkload cost
llama 3.1 8b pretraining time to train to quality target gpu cost
3 published official-source rows. Open the page to inspect its own price range and validity deadline.
Open decision pageworkload cost
deepseek-r1 offline gpu cost
3 published official-source rows. Open the page to inspect its own price range and validity deadline.
Open decision pageworkload cost
gpt-oss-120b offline gpu cost
3 published official-source rows. Open the page to inspect its own price range and validity deadline.
Open decision pageworkload cost
gpt-oss-120b server gpu cost
3 published official-source rows. Open the page to inspect its own price range and validity deadline.
Open decision pageworkload cost
qwen3-vl-235b-a22b offline gpu cost
3 published official-source rows. Open the page to inspect its own price range and validity deadline.
Open decision page