NVIDIA B200 nodes, reserved from $3.42 per GPU-hour

As of October 2026, a dedicated 8-GPU NVIDIA B200 HGX node from Avian costs $4.32 per GPU-hour on a 12-month term ($25,229 per node per month) and $3.42 per GPU-hour on a 36-month term. You pay once for the whole term, the rate is locked from the day you reserve, and the node is live in under 48 hours.

B200 price by term

Per GPU-hour and per 8-GPU node, prepaid once for the whole term. Monthly figures use 730 hours; prices in USD, before applicable tax.

TermPer GPU-hourPer node, per monthWhole term, prepaidReserve
1 Month $5.40 $31,536 $31,536 Sold out
3 Months $4.95 $28,908 $86,724 Sold out
6 Months $4.77 $27,857 $167,141 Sold out
12 MonthsSave 20% $4.32 $25,229 $302,746 Reserve
24 MonthsSave 28% $3.87 $22,601 $542,419 Reserve
3 YearsSave 37% $3.42 $19,973 $719,021 Reserve

Savings are against the 1-month rate. Need more than one node? Request a cluster and sales will quote volume pricing.

What's in a B200 node

A complete HGX system, single-tenant: all 8 GPUs and the full NVLink domain are yours for the whole term.

GPUs8× NVIDIA B200 (Blackwell), SXM on an HGX baseboard
GPU memory180 GB HBM3e per GPU, 1,440 GB per node
Memory bandwidth8 TB/s per GPU
FP4 compute9 PFLOPS dense NVFP4 per GPU
GPU interconnectNVLink, 1.8 TB/s per GPU
Power1,000 W per GPU
Host memory2 TB+ per node
RegionNorth America (US)
ProvisioningUnder 48 hours from payment

What one B200 node can run

With 1,440 GB of GPU memory in a single NVLink domain, one node holds models that would otherwise need to be split across machines:

  • A model of about 600 billion parameters in FP8 (around 600 GB of weights), with roughly 840 GB left for KV cache.
  • A 400B-class dense model in FP8, or a 70B-class model in BF16 with plenty of room for long contexts and high concurrency.
  • Fine-tuning and training of mid-size models across all 8 GPUs over NVLink.

Not sure your model fits? Work it out with the GPU memory calculator.

B200 or B300?

The B200 is the lower-cost node when your model fits comfortably in 1,440 GB. The B300 adds memory (288 GB per GPU against 180 GB) and FP4 compute for the largest models and longest contexts.

B200 reservation FAQ

12 months is the shortest term on sale right now (1 month, 3 months, 6 months terms are sold out). Terms run up to 36 months, and the longer the term, the lower the rate: $3.42 per GPU-hour on a 36-month term.

One payment for the whole term, with no hourly metering. Send a request and we'll confirm availability and invoice you (NET-30 or PO on 12-month and longer terms), or pay for a single node online by PayPal or card.

Standard reservations are provisioned in under 48 hours from payment. Multi-node clusters with fabric interconnects are typically ready within 5 business days.

Yes. Online payment covers one node; for two or more, send a request with the node count and sales will quote volume pricing, with InfiniBand or RoCE between nodes.

No. Each node is single-tenant: all 8 GPUs and the full NVLink domain are reserved for you for the whole term, and the rate never changes during it.

Reserve a B200 node

From $3.42 per GPU-hour, prepaid and rate-locked. Live in under 48 hours.