CoreWeave Earns NVIDIA Exemplar Cloud Status for GB200 NVL72 Inference

What the validation covers

NVIDIA introduced the Exemplar Cloud program in 2025 as a standard benchmark that lets cloud providers validate the performance of their infrastructure. According to CoreWeave, every participating provider goes through a comprehensive evaluation meant to reflect the needs of customers running complex and demanding AI workloads.

To qualify, a provider must show high performance and resiliency across a suite of open, workload-specific benchmarking recipes covering inference, fine-tuning and scaled pretraining. The result is a comparison of performance validated with the same criteria for everyone.

The inference validation follows an earlier milestone: CoreWeave says it was also one of the first cloud providers to achieve NVIDIA Exemplar Cloud for Training on NVIDIA GB200 NVL72. The company argues that holding both validations shows its platform can train models and also serve them efficiently and reliably in production.

How the inference tests worked

NVIDIA's inference benchmarks use DeepSeek-R1, Llama 3.3 and GPT-OSS models in single-node and multi-node configurations. They measure throughput and latency for common agentic use cases. NVIDIA specified the number of GB200 NVL72 GPUs and the backend, either TRT-LLM or SGLANG. The multi-node throughput test also used NVIDIA Dynamo.

Each scenario covered five phases of inference, with input and output context lengths in parentheses:

• Reasoning (1k/1k): long prompts and completions reflecting Chain-of-Thought processing.
• Chat (128/128): responsiveness of interactive applications, prioritizing ultra-low latency and high user concurrency.
• Summarization (8k/512): the I/O and memory bandwidth needed to ingest large prompts before producing a concise output.
• Generation (512/8k): throughput and efficiency while producing a high volume of continuous tokens.
• Disaggregation (8k/1k across nodes): efficiency when prompt processing and token generation run on different GPU nodes.

Throughput was measured in tokens per second per GPU (TPS/GPU), and latency as Time-to-First-Token (TTFT) in milliseconds. Throughput tests ran DeepSeek-R1, Llama 3.3 and GPT-OSS on single nodes with one to four NVIDIA Blackwell GPUs, and on multi-node setups with NVIDIA Dynamo using 32 NVIDIA Blackwell GPUs. Latency tests used DeepSeek-R1 and Llama 3.3 on single nodes with four GPUs. CoreWeave says it met or exceeded each scenario across all five phases in both test types.

The company stresses latency because, as it puts it, a single user request in agentic AI might trigger ten sequential model calls, which makes responsiveness a primary constraint.

The stack behind the results

CoreWeave attributes the results to a vertically integrated stack spanning compute, storage, networking and orchestration. It says it paired GB200 NVL72 with high bandwidth networking and optimized memory architectures to remove bottlenecks common in generalized cloud environments.

CoreWeave Mission Control served as the central dashboard during testing, with observability from hardware telemetry such as GPU temperature and NVLink bandwidth to application metrics such as time-to-first-token. The team also used CoreWeave GPU Straggler Detection, a recently launched Mission Control feature. According to CoreWeave, it provides real-time collective metrics including Bus Bandwidth, and automatically detects hardware lockups and pinpoints their root cause.

On the serving side, the company points to optimized runtimes and scheduling, with batching, request routing and GPU utilization built in. It also says its high-performance object storage and fast provisioning pipelines significantly reduce model load times, which improves time-to-first-token and scaling responsiveness.

What it means for customers

CoreWeave says the shared benchmark gives customers predictable AI workload performance validated through joint testing with NVIDIA, confidence in a co-engineered and tuned stack, and objective data for choosing a cloud, grounded in application performance measurements rather than vendor claims.

For teams comparing providers for production inference, the validation offers a common yardstick across reasoning, chat, summarization, generation and disaggregated workloads. The source does not publish the underlying throughput or latency figures in the text provided, so the claim that CoreWeave met or exceeded targets rests on the company's own account of the results.