Rent GB200 NVL72
Compare live on-demand and spot rental prices across 97+ cloud providers. Rack-scale Blackwell system. 72 GPUs + 36 Grace CPUs in one NVL rack. Highest throughput available.
Choosing the right billing model for GB200 NVL72
Provision and terminate at any time. Ideal for development, short experiments, and workloads with unpredictable duration.
Instances can be reclaimed when demand spikes. Best for fault-tolerant batch jobs, training with checkpointing, and preprocessing.
Lock in a rate for 1–3 months. Right for sustained production inference or long training runs where cost predictability matters.
Best use cases
- Frontier LLM training
- Trillion-param models
- Rack-scale AI
GB200 NVL72 — Specs & Benchmarks
Performance bars, compute tiers (FP32/FP16/BF16/FP8/INT8), memory specs, LLM model size guidance, and related GPU comparisons.
GB200 NVL72 Rental Guide
Renting the GB200 NVL72 makes sense when your workload requires 192GB of HBM3e memory and 4500 TFLOPS of FP16 compute. The most common use cases are Frontier LLM training, Trillion-param models, Rack-scale AI. Before committing to a rental, verify that your model and batch size fit within 192GB — a 70B parameter model requires approximately 140GB at FP16, which would require two GB200 NVL72 instances with tensor parallelism.
For fault-tolerant batch workloads — preprocessing, offline inference, or training with checkpointing — spot instances typically save 40–70% vs on-demand rates. The GB200 NVL72's 192GB VRAM makes it well-suited for long-running batch jobs where interruption recovery is manageable. On-demand instances give you full control with no commitment — ideal for development, short experiments, and workloads with unpredictable duration. Most providers bill per second or per minute, so short jobs are not penalized by hourly minimums.
For sustained production inference serving, reserved instances (1–3 month commitments) typically offer 20–35% savings vs on-demand. The GB200 NVL72's high throughput makes it cost-effective for high-volume inference where per-token cost matters. When comparing providers, look beyond the headline hourly rate: check region availability (latency matters for interactive inference), spot interruption frequency, and whether the provider offers per-second billing. Use the GPU cost calculator to model total cost across different billing models and utilization rates before choosing a provider.
Frequently Asked Questions
How much does it cost to rent a GB200 NVL72?
GB200 NVL72 on-demand rental prices vary by provider and region. On-demand rates typically range based on availability and provider margins — use the comparison table above to see current live rates across all providers. Spot instances are generally 40–70% cheaper than on-demand but can be interrupted. Monthly cost estimates (hourly rate × 730 hours) are shown in the table for sustained workloads.
Which cloud provider has the cheapest GB200 NVL72?
The cheapest GB200 NVL72 provider changes as providers update their pricing. The comparison table above shows live rates sorted by price, so the cheapest option is always at the top. Factors beyond headline price include region (latency to your users), availability (high/medium/low), and billing granularity (per-second vs per-hour minimums).
What can I run on a GB200 NVL72?
With 192GB of HBM3e, the GB200 NVL72 can run LLM models up to approximately 96B parameters at FP16, 192B at INT8, or 384B at INT4/GGUF quantization. Common workloads include: Frontier LLM training, Trillion-param models, Rack-scale AI. Rack-scale Blackwell system. 72 GPUs + 36 Grace CPUs in one NVL rack. Highest throughput available.
Should I use on-demand or spot pricing for GB200 NVL72?
Spot instances save 40–70% vs on-demand but can be interrupted when the provider needs capacity back. Use spot for: batch inference jobs, training runs with checkpointing, preprocessing pipelines, and any workload that can tolerate interruption and restart. Use on-demand for: production inference serving, interactive workloads, and jobs that cannot be interrupted. Most providers bill per second, so short on-demand jobs are not penalized by hourly minimums.
How does the GB200 NVL72 compare to the H100 for cloud rental?
The H100 80GB delivers 1,979 TFLOPS FP16 with 3,350 GB/s HBM3 bandwidth, compared to the GB200 NVL72's 4500 TFLOPS FP16 and 8000 GB/s bandwidth. The H100 is significantly more expensive — typically $2.50–$5.00/hr vs lower rates for the GB200 NVL72. For workloads that fit within 192GB and don't require FP8 precision, the GB200 NVL72 often delivers better cost-per-token than the H100.
What is the memory bandwidth of the GB200 NVL72 and why does it matter?
The GB200 NVL72 has 8000 GB/s of memory bandwidth. For LLM inference, memory bandwidth is often more important than raw TFLOPS — each autoregressive token generation reads the full model weight matrix from VRAM, so bandwidth directly determines tokens-per-second throughput. Higher bandwidth means faster inference for the same model at the same batch size. For batch inference (processing many requests simultaneously), compute throughput becomes more important.
Can I use the GB200 NVL72 for Stable Diffusion or image generation?
Yes — the GB200 NVL72 is well-suited for Stable Diffusion and image generation workloads. Image generation is primarily FP32 and FP16 compute-bound, and the GB200 NVL72's 90 TFLOPS FP32 throughput determines images-per-second. The 192GB VRAM fits SDXL (requires ~6GB) and most ControlNet pipelines. For high-throughput image generation at scale, compare cost-per-image across providers using the GPU cost calculator.