deploybase

GB200 on CoreWeave: Pricing, Specs & How to Rent

Deploybase · October 7, 2025 · GPU Pricing

Contents

GB200 GPU Specifications

GB200 is NVIDIA's Grace Blackwell superchip: one Grace CPU (72 ARM Neoverse V2 cores, 480GB LPDDR5x) paired with two B200 GPUs, connected via NVLink-C2C at 900 GB/s. The GB200 NVL72 rack-scale system contains 36 Grace CPUs and 72 B200 GPUs. Each B200 has 192GB HBM3e memory at ~8 TB/s bandwidth. Built for scale.

Specs (per B200 GPU):

  • Memory: 192GB HBM3e
  • Memory Bandwidth: ~8 TB/s
  • NVLink-C2C (Grace-to-Blackwell): 900 GB/s
  • Peak FP8: ~9 PFLOPS
  • Precision: FP4, FP8, FP16, BF16, FP32
  • Grace CPU: 72 Neoverse V2 cores, 480GB LPDDR5x

The memory bandwidth and capacity are what matter for large-model inference. Memory-bound token generation bottlenecks on HBM throughput, not compute.

CoreWeave GB200 Pricing

CoreWeave offers the GB200 NVL72 configuration: 4x GB200 at $42/hour ($10.50 per GPU). Fixed pricing. For larger scale, Azure offers 4x GB200 at $108.16/hour.

For context, 8x H100 costs $49.24/hour on CoreWeave and 8x B200 runs $68.80/hour. GB200 NVL72 is priced separately due to its integrated Grace CPU and unified memory architecture. The memory gain (192GB HBM3e per GPU) and NVLink-C2C interconnect justify GB200 for mega-scale inference. For most workloads, H100s or H200s are fine.

How to Rent GB200 on CoreWeave

  1. Create CoreWeave account.
  2. Request GB200 access (limited availability).
  3. Verify identity.
  4. Pick cluster size, region, OS image.
  5. Choose Kubernetes or bare metal.
  6. Deploy the workload.
  7. Access via SSH or kubeconfig.
  8. Scale up/down as needed.

CoreWeave handles drivers, networking, multi-GPU sync. Deploys in minutes after approval.

GB200 vs. H100 & H200

Understanding GB200's advantages relative to earlier GPU generations informs rental decisions.

GB200 vs. H100:

  • Memory: 192GB vs. 80GB (2.4x larger on GB200)
  • Throughput: Similar peak TFLOPS, but GB200's memory advantage matters more for large models
  • Cost: GB200 approximately 40% more expensive
  • Use: GB200 for single-model inference of 200B+ parameters, H100 for training efficiency

GB200 vs. H200 specifications:

  • H200 offers 141GB HBM3, bridging gap between H100 and GB200
  • H200 cost intermediate between H100 and GB200
  • GB200 has superior multi-GPU interconnect through NVLink C2C
  • H200 preferred for most use cases, GB200 for maximum scale

Vs. B200:

  • B200 represents NVIDIA's next-generation single-GPU peak
  • B200 has same memory as one GB200 GPU (192GB HBM3e); GB200 adds integrated Grace CPU and unified memory
  • B200 typically deployed in smaller clusters
  • GB200 more common in production as of March 2026

GB200 Performance & Use Cases

GB200 shines in specific workload categories where its advantages matter most.

Large language model inference:

  • LLaMA 405B: 10-15 tokens/second per GPU (unquantized)
  • GPT-4 scale models: 5-10 tokens/second per GPU
  • Multi-GPU throughput: near-linear scaling to 8 GPUs
  • Context window support: full sequence length without optimization

Code generation workloads:

  • Code Llama 70B: 40-60 tokens/second per GPU
  • Function generation: sub-500ms latency
  • Multi-turn conversations: efficient memory management

production document processing:

  • PDF extraction: 100+ documents/second with OCR
  • Information retrieval: 1000s concurrent embeddings
  • Multi-modal analysis: image + text simultaneously

Scientific computing:

  • Climate modeling: extreme throughput for ensemble runs
  • Molecular dynamics: superior memory bandwidth for simulation state
  • Computational physics: real-time analysis of massive datasets

Cost-effectiveness depends on model size and batch requirements. Smaller models don't justify GB200's premium; larger models amortize cost across more inference requests.

FAQ

Should I rent GB200 or build a cluster? Rental through CoreWeave is optimal unless expecting 6+ months sustained usage. Capital equipment costs exceed rental in most scenarios.

What's the minimum GB200 deployment? CoreWeave typically requires 8-GPU clusters minimum. This scales inference for most production scenarios.

Can GB200 run multiple models simultaneously? Yes, with careful memory partitioning. 192GB supports several 70B models or one 200B+ model with headroom.

How does GB200 latency compare to H100? Latency is similar (both 10-50ms range). GB200's advantage is throughput and memory, not latency.

Does CoreWeave offer trial periods for GB200? Contact CoreWeave sales directly. Early-access customers sometimes receive trial allocations.

Sources