GPU × use-case guide · Information extraction

Is the GB200 NVL72 (per GPU) a good GPU for information extraction?

The GB200 NVL72 (per GPU) is a blackwell NVIDIA GPU with 192 GB HBM3e (8000 GB/s), 2250 TFLOPS BF16 / 4500 TFLOPS FP8, and a 1200 W TDP. Information extraction workloads care most about structured-output throughput at scale; latency matters less than tokens-per-dollar. Here's how the GB200 NVL72 (per GPU) measures up.

What models fit on a single GB200 NVL72 (per GPU)?

Weights only, reserving ~25% of the 192 GB for KV cache, activations and fragmentation. ✓ = fits on one card.

Largest single-card fit: Qwen 2.5 72B at BF16, Qwen 2.5 72B at FP8, Qwen 2.5 72B at INT4. Bigger models need tensor-parallel across 72 cards.

GB200 NVL72 (per GPU) for information extraction, specifically

Information extraction is throughput-bound rather than VRAM-bound, so the GB200 NVL72 (per GPU)'s 8000 GB/s of bandwidth and 2250 TFLOPS matter more than raw capacity. Because it runs offline, batch aggressively to push tokens-per-dollar down. Size it precisely on the calculator.

GB200 NVL72 (per GPU) pricing across providers

Verdict

With 192 GB, the GB200 NVL72 (per GPU) is a data-center-class card that comfortably handles information extraction for models up to Qwen 2.5 72B at full precision on a single card — a strong pick if your budget supports ~$8.5/hr.