The NVIDIA RTX 6000 Pro (RTX PRO 6000 Blackwell Server Edition) is the card that changed the maths for mid-sized inference in 2026. It puts 96 GB of ECC GDDR7 behind a full GB202 Blackwell die: 24,064 CUDA cores, 752 fifth-generation Tensor Cores with native FP4 and FP8, and 1.79 TB/s of memory bandwidth. That is more VRAM than an H100 at a fraction of the hourly price.
On packet.ai the RTX 6000 Pro rents from $0.66/hr on Dynamic capacity, billed hourly with no commitment. Dedicated single-tenant cards are launching soon. Instances are SSH-ready in under five minutes from the dashboard, API or packet CLI, with local NVMe attached and the CUDA stack in place.
Why rent the RTX 6000 Pro on packet.ai
96 GB on one card
A 70B-parameter model in 4-bit fits with 15–20 GB of headroom for KV cache. Llama 3.1 70B, Qwen 72B and DeepSeek distills run on a single GPU without tensor parallelism, which removes a whole class of latency and cost.
Blackwell Tensor Cores
Fifth-generation Tensor Cores add native FP4 alongside FP8 and BF16. For inference that means roughly 2× the throughput of the Ada-generation L40S at the same batch size, and FP4 quantised models that keep quality while halving memory again.
1.79 TB/s GDDR7
Nearly double the bandwidth of the L40S and within 10% of the A100's HBM2e. Token generation is memory-bound; bandwidth per dollar is the number that matters, and this card leads the field on it.
Lowest published rate
$0.66/hr on packet.ai Dynamic is 33% below the next-cheapest published rate for this card and 64–78% below RunPod, Hyperstack, Verda and fal.ai. Same silicon, same VRAM, same clocks.
Built for the data centre
The Server Edition is passively cooled, ECC-protected and supports MIG partitioning into up to four 24 GB instances, plus 4× NVENC and 4× NVDEC for video pipelines. It is a workstation-class die packaged for 24/7 rack duty.
RTX 6000 Pro 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 $0.66/hr Dynamic.
| Provider | RTX 6000 Pro rate |
|---|---|
| packet.ai | $0.66/hr Dynamic |
| Vast.ai (marketplace) | $1.00/hrpacket.ai −34% |
| Hyperstack | $1.85/hrpacket.ai −64% |
| Verda | $1.89/hrpacket.ai −65% |
| RunPod Secure Cloud | $2.09/hrpacket.ai −68% |
| CoreWeave (reserved) | $2.50/hrpacket.ai −74% |
| fal.ai (dedicated) | $2.99/hrpacket.ai −78% |
Specifications
| Feature | Value |
|---|---|
| GPU architecture | NVIDIA Blackwell (GB202) |
| CUDA cores | 24,064 |
| Tensor Cores | 752 (5th generation, FP4 / FP8 / BF16 / TF32) |
| RT Cores | 188 (4th generation) |
| Memory | 96 GB GDDR7 with ECC |
| Memory interface | 512-bit |
| Memory bandwidth | 1,792 GB/s |
| FP32 compute | ~117 TFLOPS |
| AI performance | Up to ~3,750 TOPS (FP4, sparse) |
| Interconnect | PCIe 5.0 ×16 (no NVLink) |
| Multi-instance GPU | Up to 4 × 24 GB MIG partitions |
| Media engines | 4 × NVENC (9th gen), 4 × NVDEC (6th gen), AV1 encode |
| Board power | 400–600 W configurable (Server Edition) |
| Form factor | Dual-slot, passive cooling |
What you can run on 96 GB
Where the RTX 6000 Pro is the right card, where it is comfortable, and where you should step up.
| Workload | Fit | Notes |
|---|---|---|
| 70B LLM inference (4-bit) | Single card, 15–20 GB headroom for KV cache | Llama 3.1 70B, Qwen 2.5 72B, DeepSeek-R1 distills |
| 30–34B LLM at FP16 | Single card, comfortable | Full-precision serving without quantisation |
| 8 × 7–8B models on one node | 8 cards, 768 GB total VRAM | Multi-model serving at $5.28/hr for the node |
| FLUX.1 / SDXL image generation | Full precision plus LoRA stacks | FP4 pipelines roughly double images per hour |
| Video generation (Wan, Hunyuan, CogVideo) | Long contexts fit in 96 GB | AV1 hardware encode for output |
| LoRA / QLoRA fine-tuning to 70B | Single card for QLoRA; 2–4 cards for LoRA at 70B | Hourly billing suits 4–40 hour runs |
| Pre-training or 405B inference | Not the right card | Use B200 or a cluster |
Deploy in minutes
- Create an accountSign up at dash.packet.ai. No credit card required to start; billing is usage-based.
- Pick the RTX 6000 Pro and a tierChoose Dynamic (Dedicated launching soon), 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 RTX 6000 Pro 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 it cost to rent an RTX 6000 Pro on packet.ai?
$0.66 per GPU-hour on Dynamic capacity, billed hourly, no minimum term. That is about $482 per month at 730 hours; monthly commits take up to 20% off the hourly rate. Dedicated single-tenant RTX 6000 Pro is launching soon; join the notify list on the pricing page.
What is Dynamic and how is it different from Dedicated?
Dynamic is a multi-tenant card with scheduler-enforced isolation. You get the same peak performance and the full 96 GB of VRAM; packet.ai's scheduler co-locates workloads that stress different GPU dimensions, and publishes latency within roughly 2–5% of Dedicated. Dedicated is the whole card, exclusively yours, with a 99% SLA.
How does the RTX 6000 Pro compare to an H100 for inference?
It has more memory (96 GB vs 80 GB) and much less bandwidth (1.79 TB/s vs 3.35 TB/s) with no NVLink. For single-card 30–70B inference at moderate batch sizes it is the more cost-efficient card by a wide margin. For latency-critical serving that needs sub-400 ms time-to-first-token at high concurrency, or multi-GPU tensor parallelism, the H100 or B200 is the better fit.
Can I rent multiple RTX 6000 Pro GPUs in one instance?
Yes. Multi-GPU configurations up to 8 × RTX 6000 Pro (768 GB total VRAM) are available on request, and bare-metal nodes and clusters are quoted through the wholesale form. Note the card uses PCIe 5.0, not NVLink, so inter-GPU bandwidth is lower than SXM parts.
Which region is it in?
US West (California) today. Use the region filter on the pricing page for the live list; European capacity for this SKU is on the roadmap.
What software is on the instance?
Ubuntu with current NVIDIA drivers and CUDA, plus SSH access and local NVMe. Bring your own containers or install vLLM, SGLang, TensorRT-LLM, ComfyUI, diffusers or any framework directly. There is no template marketplace; you get a clean machine.
How fast can I get one?
Dynamic instances are typically SSH-ready in under five minutes from the API call. No quota approval, no credit card to start.
Are there hidden fees?
No platform fees, no ingress & no Egress.
Same silicon. Smarter economics.
RTX 6000 Pro from $0.66/hr, billed hourly, live in EU West. No quota approval, no credit card to start.