compute dedicated blackwell GPUsReserve dedicated NVIDIA B200 and B300 GPUs — complete 8-GPU HGX nodes, single-tenant and NVLink-connected. Reserved capacity, prepaid per term, rate locked, provisioned in under 48 hours. One month to three years: your call.
Powering inference for teams at Bank of America, Boeing, Google, Intel, Salesforce and more.
Complete 8-GPU HGX nodes, one payment for the whole term. Rate locked from day one — no metering, no surprise bills.
12 months — prepaid once, locked for the full term.
Effective rate: $4.32/GPU·hr · 730 hr/mo basis
8× B200 · 1440 GB HBM3e · NVLink 1.8 TB/s domain
128 vCPUs · 2048 GB host RAM · 61 TB NVMe · per node
Effective rate: $4.95/GPU·hr · 730 hr/mo basis
8× B300 · 2304 GB HBM3e · NVLink 1.8 TB/s domain
192 vCPUs · 4096 GB host RAM · 61 TB NVMe · per node
Sold as complete 8-GPU HGX nodes — one node per online order; multi-node fleets via sales. All rates are per GPU, per hour, prepaid for the full term (730 hr/GPU/month basis), plus applicable sales tax/VAT where applicable. Prices as of Oct 2026, subject to change — your reserved rate locks at order. Available in North America (US) only. SXM modules on HGX baseboards; host specs illustrative, confirm against NVIDIA HGX datasheets.
Machine-readable price feed (prices + stock, JSON): avian.io/compute/prices.json
Volume & multi-node. Reserving multiple nodes or building a multi-node cluster with fabric interconnects? Talk to sales for volume pricing, priority provisioning, and rack-level scheduling (InfiniBand / RoCE). Get a quote →
Need inference instead of raw GPUs? Run DeepSeek V4.1, MiniMax M3, GLM-5.3, Kimi K3, MiMo-V2.6 and more on Avian's per-token API — from $0.1275/M tokens, no reservation required. See inference pricing →
The same Blackwell hardware, without the hyperscaler meter running.
Pay once for the whole term — there is no hourly bill to watch and no surprise invoice at month-end.
Whatever the market does to GPU prices mid-term, your rate doesn't move. Budget with certainty.
B200-class GPUs, speculative decoding, 489 tok/s in production. You get the fleet we trust with our own API.
12-month+ terms support invoiced orders (NET-30 / PO) — one predictable line item for the finance team.
From reservation to a running node in under two days.
B200 or B300 — complete 8-GPU HGX nodes, from one node to a full rack. 1 month to 3 years, one payment for the whole term, rate locked on day one.
Your GPUs are racked, imaged, and handed over — vLLM, SGLang, or your own stack. Single-tenant and isolated: no noisy neighbors, ever.
Serve models, train, fine-tune, or run your own inference stack. What you were quoted is what you pay — for the entire term.
The same standards that run our inference API, applied to your dedicated hardware.
Everything you need to know about Avian GPU Compute
You pay once, upfront, for the entire term of your 8-GPU node — 1, 3, 6, or 12 months, or 3 years. There is no hourly metering and no recurring invoice. Your rate is locked the day you reserve and never moves, no matter what the market does. 12-month+ terms also support invoiced orders (NET-30 / PO) for procurement teams.
Anything. The hardware is single-tenant and entirely yours: pick a standard inference image (vLLM, SGLang), or bring your own stack — custom CUDA, training and fine-tuning frameworks, speculative decoding, RL rollouts. It's dedicated metal, not a shared endpoint.
No. Ingress and egress are free on reserved nodes — move datasets in and results out without watching a second meter.
None. One prepaid payment covers the entire term — hardware, power, cooling, rack, and networking included. The number you see is the number on the invoice.
Nodes ship as complete 8-GPU HGX systems — one node minimum — so you get the full NVLink domain and predictable throughput. For sub-node workloads, the per-token API is usually the better deal: no commitment, from $0.1275/M tokens. See inference pricing.
Standard node reservations are provisioned in under 48 hours from payment. Multi-node clusters with fabric interconnects are typically ready within 5 business days.
Hardware is covered by a replacement SLA — we handle the RMA and swap the module, and your reservation is covered by our 99.9% uptime SLA with service credits. Your prepaid term is not consumed by our hardware failure.
Yes. Additional nodes join at your current locked rate tier for the remainder of the term. For multi-node additions, talk to sales about interconnect options and scheduling.
Reservations are currently available in North America (US) only. Talk to sales if you need capacity elsewhere — we prioritize US deliveries today.
Yes — that's the Avian you already know. Run DeepSeek V4.1, MiniMax M3, GLM-5.3, Kimi K3, MiMo-V2.6 on our OpenAI-compatible API, billed per token from prepaid credits, from $0.1275/M tokens. See inference pricing. Many teams run both: the API for spiky workloads, dedicated GPUs for baseline traffic and training.