H100 from $0.65/hr, H200 from $2.25/hr, B200 from $3.75/hr. No contracts, no spot/reserved maze, no vendor lock-in. Here’s what packet.ai is and how it works.
packet.ai is a GPU cloud built on hosted-ai infrastructure, giving developers on-demand access to NVIDIA B200, H200, A100, and RTX 6000 Pro GPUs at prices typically 50% below hyperscaler rates. SSH-ready in under 5 minutes. No contracts, no minimums, per-hour billing.
Key takeaways
GPU cloud pricing is broken in a specific way: the gap between what datacenters pay for GPU compute and what developers pay to access it is enormous. Hyperscalers charge 3-5x wholesale rates. Most neoclouds are closer but still carry significant margin. The hardware is the same NVIDIA silicon everywhere.
packet.ai is built on hosted-ai infrastructure, which gives us access to wholesale datacenter economics that third-party GPU cloud providers cannot match. We pass that through to pricing. The B200 at $3.75/GPU-hr on packet.ai is not a promotional rate - it is what the hardware actually costs to run at our infrastructure layer, with a sustainable margin on top.
packet.ai currently offers five NVIDIA GPU SKUs on demand:
H100 SXM and H200 are in the pipeline. GPU waitlist at packet.ai/gpu-waitlist.
Every GPU SKU on packet.ai is available in two tiers:
Dynamic is scheduler-managed capacity. The GPU runs your workload, but the scheduler can place it on different physical hardware across the fleet. Same peak performance and VRAM as Dedicated - the scheduling model is different, not the hardware. Lowest rate. Suited for batch inference, fine-tuning with checkpointing, experiments, and workloads where an interruption is recoverable.
Dedicated pins the GPU to your account for the duration of the session. No scheduling interference. 99.99% uptime SLA. Suited for production inference APIs with latency SLAs, long training runs, and compliance workloads that require pinned hardware.
The pricing difference between Dynamic and Dedicated reflects the scheduling model, not the hardware. Both tiers run on the same NVIDIA silicon.
Token Factory is packet.ai's OpenAI-compatible inference API. Instead of provisioning a GPU, installing a model, and running vLLM yourself, you call the Token Factory API with the same request format you use with OpenAI.
Current model support: Llama 3 (8B and 70B), DeepSeek R1 and V3, Qwen 2.5, Mistral 7B. Per-token billing. Scale to zero between calls.
Token Factory is the right choice when you want inference without GPU management. For teams that need custom models, fine-tuned weights, or specific inference configurations, provisioning a GPU directly gives you more control.
Pixel Factory is packet.ai's image and video generation API. Text-to-image, image-to-image, ControlNet, and video generation - Flux.1, SDXL, Stable Diffusion 3.5, and video models - with per-image billing. No GPU management, no model loading, no VRAM sizing.
Pixel Factory is launching soon. Early access at packet.ai/pxl.
packet.ai runs on hosted-ai infrastructure. hosted-ai operates large-scale GPU and compute facilities across the UK, Netherlands, Germany, and the US. The provisioning agent that allocates GPU capacity to packet.ai customers is the same agent that hosted-ai has deployed across its enterprise cloud fleet.
This matters for two reasons. First, we have operational history before packet.ai launched - the provisioning stack is not new. Second, our cost structure reflects wholesale datacenter economics, not third-party reselling margins.
Sign up at dash.packet.ai. The onboarding flow walks through adding an SSH key, selecting a GPU, and launching your first instance. From account creation to SSH access takes under 5 minutes for most users.
If you want inference without GPU management, Token Factory and Pixel Factory are accessible immediately after signup with no instance provisioning required.
For questions, help@packet.ai. No sales process, no demo call required to access any tier.
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start Building →
