NVIDIA B200: Specs, Price per GPU-Hour, and B200 vs H200 (October 2026)

NVIDIA B200 Specs

NVIDIA publishes B200 numbers at the system level, for the eight-GPU DGX B200. The per-GPU column divides those totals by eight. NVIDIA marks its FP4 figure as sparse and dense and its FP8 figure as sparse only, with dense at half.

NVIDIA B200 Price: Cloud Rental per GPU-Hour (October 2026)

These are the published on-demand prices for one B200, read from each provider's pricing page on October 6, 2026. Modal bills per second and charges CPU and RAM separately; RunPod and Lambda bundle host CPU, RAM, and disk into the hourly rate.

To buy, NVIDIA does not list a price for a single B200. Northflank reported OEM quotes of $45,000 to $50,000 per GPU and complete 8x B200 servers above $500,000 in August 2025. At $6.79 per hour, $45,000 buys about 6,600 GPU-hours of rental, or nine months of one GPU running around the clock, before power, cooling, and the host server.

B200 vs H200 vs H100

The H200 and H100 share Hopper compute and differ in memory. The B200 moves to Blackwell, adds FP4, and raises memory and bandwidth again. For more on the Hopper pair, see H100 vs H200 and the NVIDIA H200 guide.

H200 rental prices are Modal ($0.001261/s) and RunPod ($4.59). H100 prices are Modal ($0.001097/s) and Lambda's 8x and 4x instances ($4.09 and $4.19).

B200 Bandwidth per Dollar: Why It Wins on Decode

Each decode step reads the model weights and the KV cache from HBM, so tokens per second on a fixed model scales with memory bandwidth more than with FLOPS. Dividing the hourly price by bandwidth gives a rough cost of decode capacity. At the RunPod and Lambda prices above, the B200 is the cheapest of the three.

The ratio breaks down for small models at low concurrency, where none of the three GPUs is bandwidth-limited and the cheaper hour wins. Measured B200 and B300 serving numbers for open models are on the dedicated inference benchmarks page.

What Fits on 8x B200 (1,440 GB)

FP8 weights take about one byte per parameter, FP4 about half a byte. GLM-5.2 at 753B parameters needs roughly 755 GB in FP8. On an 8x B200 node that leaves about 685 GB for KV cache and activations; on an 8x H200 node (1,128 GB) it leaves about 373 GB, and it does not fit on an 8x H100 node (640 GB). A dense 70B model in FP8 fits on one B200 with about 110 GB left for context.

GB200 NVL72 and B300: The Other Blackwell Options

GB200 NVL72 is a liquid-cooled rack of 36 Grace CPUs and 72 Blackwell GPUs in a single NVLink domain. NVIDIA lists 130 TB/s of GPU-to-GPU bandwidth and 720 PFLOPS of dense NVFP4 for the rack. Each GB200 Superchip pairs one Grace CPU with two Blackwell GPUs. The rack exists for models too large or too communication-heavy for an eight-GPU node.

The B300 raises memory to 288 GB of HBM3e per GPU. It rents for $7.10 per GPU-hour on Modal ($0.001972/s) and $7.89 on RunPod, a 14% to 16% premium over the B200 for 1.6x the memory. NVIDIA's next generation is Rubin; its HGX page projects 10x the token throughput of HGX B200 for HGX Rubin NVL8 and labels that projection subject to change.

Rent, Buy, or Use a Dedicated B200 Endpoint

Renting by the hour suits bursty training and evaluation. Buying needs a site that can deliver 14.3 kW per DGX and enough utilization to beat about $6.79 per GPU-hour. For serving an open model in production, a dedicated endpoint removes the serving stack: Morph runs reserved B200 and B300 capacity behind an OpenAI-compatible API at $9.98 per B200-hour. See dedicated inference for models and regions.

Frequently Asked Questions

Related Resources

• NVIDIA H200: VRAM, Bandwidth, and How It Compares to the H100
• H100 vs H200
• NVIDIA H100: Specs, Price, and Comparisons
• A100 vs H100
• LLM Inference
• Dedicated Inference on B200 and B300