The NVIDIA B200 is the flagship of the Blackwell generation: two reticle-limited dies fused into one 208-billion-transistor GPU, 180 GB of HBM3e at up to 8 TB/s, a second-generation Transformer Engine with FP4 and FP8, and fifth-generation NVLink at 1.8 TB/s per GPU. Against an H100 it roughly doubles training throughput and, with FP4 inference, can deliver several times the tokens per second on frontier-sized models.
packet.ai sells the B200 on demand at $3.75/hr on Dynamic and $6.99/hr on Dedicated, one card at a time, billed hourly, with no reservation and no eight-GPU minimum. Multi-GPU nodes, bare metal and InfiniBand clusters up to 1,024+ GPUs are available at wholesale rates around 30% below retail.
Why rent the B200 on packet.ai
Lowest published B200 rate
$3.75/hr Dynamic is the cheapest on-demand B200 tracked by GetDeploying as of September 2026, below Verda's spot tier and roughly a quarter of AWS and Google Cloud's per-GPU rate. Dedicated at $6.99/hr is within twenty cents of RunPod and Lambda and below Nebius, CoreWeave and every hyperscaler.
180 GB of HBM3e
A 70B model fits in FP16 on one card. A 405B model fits in FP4 on one card. Long-context serving, large KV caches and big batch sizes stop being a memory-planning exercise.
FP4 Transformer Engine
The second-generation Transformer Engine adds FP4 with micro-tensor scaling, and NVIDIA's TensorRT-LLM, vLLM and SGLang all support it. For inference, FP4 doubles effective throughput again over FP8 with quality that holds on most benchmarks.
8 TB/s bandwidth and 1.8 TB/s NVLink
Memory bandwidth is roughly 2.4× the H100's and NVLink 5 doubles inter-GPU bandwidth, so tensor-parallel serving across 2, 4 or 8 cards scales close to linearly.
One card, or a thousand
Start with a single Dynamic B200 for a fine-tune. Move to Dedicated for production serving. Grow into 8× bare-metal nodes and InfiniBand clusters on the same account. Typical 64-node B200 clusters land at $2.80–3.20/GPU-hr on 12-month terms.
B200 pricing on packet.ai
Three ways to buy the same card. Mix them on one account: Dynamic for development and bursty work, Dedicated for production, Clusters when a single node is not enough.
Dynamic
Shared card, scheduler-enforced isolation. Same peak performance and VRAM as Dedicated.
- Hourly billing, no commitment
- SSH-ready in under 5 minutes
- Never pre-empted
Dedicated
Whole card, exclusively yours. Zero scheduler interference, predictable p99.
- 99% SLA
- Single-tenant hardware
- 5–10 minute provisioning
Clusters
Multi-node, InfiniBand fabric, custom storage, named TAM.
- Bare-metal nodes and 8× configurations
- Up to 1,024+ GPUs
- 2–6 week provisioning
How it compares
Published on-demand starting rates for the same GPU at other providers, USD per GPU-hour. packet.ai's reference rate is $3.75/hr Dynamic.
| Provider | B200 rate |
|---|---|
| packet.ai | $3.75/hr Dynamic Dedicated $6.99/hr |
| Verda (spot) | $3.06/hrlower by 18% |
| Hyperstack | $6.00/hrpacket.ai −38% |
| Verda (on-demand) | $6.11/hrpacket.ai −39% |
| Modal | $6.25/hrpacket.ai −40% |
| RunPod Secure Cloud | $6.79/hrpacket.ai −45% |
| Lambda (1×) | $6.99/hrpacket.ai −46% |
| Nebius | $7.15/hrpacket.ai −48% |
| CoreWeave (8× node) | $8.60/hrpacket.ai −56% |
| AWS p6-b200 (8× node) | $14.24/hrpacket.ai −74% |
| Google Cloud a4 (8× node) | $16.11/hrpacket.ai −77% |
Specifications
| Feature | Value |
|---|---|
| GPU architecture | NVIDIA Blackwell (dual-die, 208B transistors, TSMC 4NP) |
| Memory | 180 GB HBM3e (192 GB on-package) |
| Memory bandwidth | Up to 8 TB/s |
| FP4 Tensor (2nd-gen Transformer Engine) | 18 PFLOPS sparse / 9 PFLOPS dense |
| FP8 Tensor | 9 PFLOPS sparse / 4.5 PFLOPS dense |
| FP16 / BF16 Tensor | 4.5 PFLOPS sparse / 2.25 PFLOPS dense |
| TF32 Tensor | 2.2 PFLOPS sparse / 1.1 PFLOPS dense |
| FP64 / FP64 Tensor | 40 TFLOPS |
| Interconnect | NVLink 5 · 1.8 TB/s per GPU · PCIe 6.0 |
| Multi-instance GPU | Up to 7 MIG instances |
| Decompression engine | Yes (800 GB/s) |
| Confidential computing | Yes |
| Board power | Up to 1,000 W (HGX B200 SXM) |
| Form factor | SXM, 8-GPU HGX baseboard |
What you can run on 180 GB
Where the B200 is the right card, where it is comfortable, and where you should step up.
| Workload | Fit | Notes |
|---|---|---|
| 70B LLM inference at FP16 | Single card with large KV cache | No quantisation, long contexts |
| 405B LLM inference | Single card at FP4; 2 cards at FP8 | Llama 3.1 405B, DeepSeek-V3 class MoE across a node |
| Frontier MoE serving (DeepSeek, Kimi, Qwen-Max class) | 4–8 cards with NVLink | Tensor + expert parallel on one HGX node |
| Full fine-tuning to 70B | 8-card node | FP8 training with Transformer Engine |
| LoRA fine-tuning to 405B | 2–4 cards | Hourly Dynamic keeps 20-hour runs cheap |
| Pre-training and large-scale RL | InfiniBand cluster | Wholesale quote, 2–6 week provisioning |
| 7–13B inference or SDXL | Overkill | Use RTX 4090, L40S or RTX 6000 Pro |
Deploy in minutes
- Create an accountSign up at dash.packet.ai. No credit card required to start; billing is usage-based.
- Pick the B200 and a tierChoose Dynamic or Dedicated, hourly or monthly billing, and the region. Multi-GPU and bare metal are on request.
- DeployFrom the dashboard, the API, or the packet CLI. Dynamic instances are typically SSH-ready in under five minutes.
- SSH in and runUbuntu, current NVIDIA drivers and CUDA are in place, with local NVMe attached. Bring your containers or install directly.
# after the instance shows Running in the dashboard ssh ubuntu@<instance-ip> nvidia-smi # confirm the B200 and driver # example: serve a model with vLLM pip install vllm python -m vllm.entrypoints.openai.api_server --model <hf-model-id> --port 8000
CLI and API reference, including how to launch and stop instances programmatically, is in the packet.ai docs.
FAQ
What does a B200 cost on packet.ai?
$3.75 per GPU-hour on Dynamic and $6.99 per GPU-hour on Dedicated, billed hourly with no commitment. At 730 hours that is about $2,738 and $5,103 per month respectively; monthly commits take up to 20% off. Clusters on 12-month terms typically land at $2.80–3.20 per GPU-hour.
Is there a waitlist or minimum?
No. The B200 is live in US East (Virginia) on both Dynamic and Dedicated, and you can rent a single card. Hyperscalers sell it in 8-GPU shapes, often reservation-only; packet.ai does not.
Dynamic or Dedicated for the B200?
Dynamic for fine-tuning, evaluation, batch inference and development: same peak performance and the full 180 GB, roughly half the price. Dedicated for production inference where you need single-tenant hardware, predictable p99 and the 99% SLA. Most teams run both on one account.
How does the B200 compare to the H100 and H200?
Roughly 2.25× the H100's dense FP16 throughput, 2.4× the memory bandwidth, and 2.25× the memory (180 GB vs 80 GB). Against the H200 (141 GB, 4.8 TB/s) it adds about 28% more memory, 67% more bandwidth and the FP4 Transformer Engine. On packet.ai the B200 is also the only one of the three with a published rate today; H100 SXM and H200 are on the notify list.
Can I get an 8× B200 node with NVLink?
Yes. Bare-metal HGX B200 nodes (8 GPUs, NVLink 5 and NVSwitch) and multi-node InfiniBand clusters up to 1,024+ GPUs are available through the wholesale quote form. Single-node bare metal typically provisions in days; clusters take two to six weeks for InfiniBand cabling and validation.
What software stack works on it?
CUDA 12.x with Blackwell support, PyTorch, JAX, TensorRT-LLM, vLLM and SGLang all run natively, including FP4 kernels. Instances ship with Ubuntu, current drivers and CUDA; bring your own containers or install directly.
What are the hidden costs?
None beyond the rate card. No platform fees, no ingress, egress at $0.04/GB, up to 2 TB local NVMe included, object storage from $0.018/GB-month.
How fast can I deploy?
Dynamic B200 instances are typically SSH-ready in under five minutes; Dedicated in five to ten. No quota approval.
Same silicon. Smarter economics.
B200 from $3.75/hr, billed hourly, live in US East · Virginia. No quota approval, no credit card to start.