AWSBee

Open models · NVIDIA GPUs · Your cloud

Private GPU inference, in your own cloud account.

AWSBee deploys Llama, Qwen, Mistral and NVIDIA Nemotron models on NVIDIA GPU instances you own, then gives you one OpenAI-compatible API with usage analytics and cost per token.

No credit card. You pay your cloud provider for GPU hours directly.

one endpoint for every deployment
curl https://awsbee.com/api/v1/chat/completions \
  -H "Authorization: Bearer $BEE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "support-llama",
    "messages": [{"role": "user", "content": "Summarize ticket #4521"}],
    "stream": true
  }'
13
curated open models
16
NVIDIA GPU instance types
6
GPU generations sized

How it works

From an empty account to a serving endpoint in three steps.

  1. 01

    Connect your cloud account

    Create a scoped access key with the policy we provide. AWSBee checks your GPU quotas per region before you launch anything.

  2. 02

    Pick a model and a GPU

    The sizing engine computes weights, KV cache and overhead for your context length and recommends the cheapest NVIDIA instance that fits.

  3. 03

    Call one API

    AWSBee launches a CloudFormation stack running vLLM or NVIDIA NIM. Every deployment is reachable through one OpenAI-compatible endpoint.

Built for production

Everything between “we picked a model” and “it is serving traffic”.

Product tour

Your account, your data

Weights, prompts and outputs stay on instances you own. AWSBee never stores request bodies.

GPU-aware sizing

Memory math for every model × instance pair: T4, A10G, L4, L40S, A100 and H100, with tensor parallelism.

vLLM or NVIDIA NIM

Serve open models with vLLM, or run NVIDIA NIM containers with your NGC key — same API either way.

OpenAI-compatible gateway

Swap the base URL in any OpenAI SDK. Streaming, chat, completions and embeddings pass straight through.

Cost per token

Latency, time to first token, tokens per second and spend per million tokens, per deployment.

Locked-down by default

Model ports open only to the AWSBee gateway, IMDSv2 enforced, encrypted volumes, secrets passed as NoEcho.

Model catalog

Open-weight models, sized for NVIDIA GPUs.

All models

Already running a model server?

Register any OpenAI-compatible endpoint — your own vLLM or Triton cluster, or a hosted NVIDIA API catalog model — and get the same keys, playground and analytics without launching anything.

GPU prices shown in the product are estimates; your cloud provider bills you directly.