Open models should be as easy to run as an API call, and stay in your own account.

Teams adopting AI hit the same wall: hosted APIs are easy but send data to a third party, while self-hosting open models means learning GPU memory math, drivers, inference servers, networking and cost tracking before the first request is served.

AWSBee closes that gap. It sizes open-weight models for NVIDIA GPUs, launches them in the customer's own cloud account with a hardened, reproducible template, and gives developers one OpenAI-compatible endpoint with keys, a playground and per-token analytics. The data plane — GPUs, model weights, prompts and outputs — never leaves the customer's account.

Team

• MKMarwan KhanfarFounder, AWSBeeLinkedIn

What we are building next

• Autoscaling across multiple instances per deployment, driven by queue depth and KV-cache pressure.
• Fine-tuning jobs (LoRA) on the same GPU fleet, with adapters served side by side.
• TensorRT-LLM engine builds for NVIDIA Hopper and Ada GPUs, benchmarked against vLLM per model.
• Spot capacity with automatic fallback to on-demand.