🧠 AI Inference at the Network Edge

AI that responds in
milliseconds, not seconds

Run large language models directly at Internet Exchanges — where Starlink, satellite, and rural broadband traffic naturally flows. Sub-5ms AI inference that makes real-time voice agents and interactive AI practical for any user, anywhere.

View Pricing See API Docs ↓
<5ms
Inference Latency at IX
7
Internet Exchange POPs
100%
OpenAI-Compatible API
0
Hyperscaler Dependency
Pricing

Dedicated GPU inference.
No per-token cloud markup.

Flat monthly pricing on dedicated NVIDIA hardware. No surprise bills. No rate limits per request. Your tokens, your models, your SLA.

🚀
Starter
$299/mo
10M tokens/mo · 3 model slots
  • ✦ 10 million tokens per month
  • ✦ 3 concurrent model deployments
  • ✦ OpenAI-compatible REST API
  • ✦ Llama 3, Mistral, Phi included
  • ✦ Sub-5ms inference at nearest IX
  • ✦ HTTPS endpoint + API key auth
  • ✦ Standard support (email)
Get Started
⭐ Most Popular
⚡
Growth
$799/mo
50M tokens/mo · 10 model slots
  • ✦ 50 million tokens per month
  • ✦ 10 concurrent model deployments
  • ✦ Dedicated GPU partition
  • ✦ All Starter models + DeepSeek, Qwen
  • ✦ Custom model fine-tunes supported
  • ✦ Streaming & batch endpoints
  • ✦ Usage dashboard + analytics
  • ✦ Priority support (4-hour response)
Get Started
🌐
Scale
$1,999/mo
200M tokens/mo · Unlimited models
  • ✦ 200 million tokens per month
  • ✦ Unlimited model deployments
  • ✦ Dedicated NVIDIA GPU node
  • ✦ Multi-POP deployment option
  • ✦ Custom model fine-tuning service
  • ✦ HIPAA-ready data handling
  • ✦ SLA with uptime guarantee
  • ✦ 24/7 priority support
Get Started
🏛️
Enterprise
Custom
Unlimited · Dedicated node · SLA
  • ✦ Fully dedicated GPU hardware
  • ✦ Multi-region redundancy
  • ✦ Private network connectivity (BGP)
  • ✦ Air-gapped deployment option
  • ✦ Government / CMMC compliance
  • ✦ Custom SLA with penalties
  • ✦ Dedicated account team
  • ✦ On-site deployment available
Contact Sales
How It Works

AI at the peering point,
not in the cloud

Traditional cloud AI routes your request from the edge, across the internet, to a hyperscaler data center, and back. We intercept it at the IX — where your traffic already terminates.

200ms+
Cloud AI
Round-Trip
<5ms
Peering Edge
Inference
🛰️
Your ApplicationAPI call from Starlink / rural / satellite
↓
📡
ISP Ground StationTraffic terminates at the POP
↓
⚡
Peering Edge IX NodeNVIDIA GPU runs inference at peering point
↓
🔁
vLLM Inference EngineLlama / Mistral / DeepSeek model runs locally
↓
✅
Response ReturnedSub-5ms total processing time
Available Models

Open-source models,
production-optimized

All models served through an OpenAI-compatible API. Your existing SDKs and code work without changes — just point to our endpoint.

Meta AI
Llama 3.1 70B
Meta's flagship open model. Strong reasoning, coding, and instruction following. Ideal for general-purpose applications and chatbots.
70B params · Q4 quantized
Meta AI
Llama 3.2 11B Vision
Multimodal Llama with vision capabilities. Process images alongside text — perfect for document analysis, OCR, and visual Q&A.
11B params · Vision enabled
Mistral AI
Mistral 7B Instruct
Blazing-fast inference. Mistral's 7B model punches above its weight class for summarization, classification, and structured output tasks.
7B params · Fastest latency
Mistral AI
Mixtral 8x7B MoE
Mixture-of-Experts architecture. Near-GPT4-level performance on knowledge tasks with efficient inference for multi-turn conversations.
MoE · 8x7B · High throughput
DeepSeek
DeepSeek-R1 14B
China's top-performing reasoning model. Exceptional at math, logic, and coding tasks. Often outperforms larger models on structured problems.
14B params · Reasoning focused
Microsoft
Phi-3 Medium
Small but mighty. Microsoft's Phi-3 is optimized for edge deployment — maximum capability at minimum GPU memory. Great for voice AI latency.
14B params · Edge optimized
Alibaba
Qwen 2.5 72B
Top-tier multilingual model with exceptional coding capabilities. Strong across English, Chinese, and many other languages. Excellent for enterprise use.
72B params · Multilingual
Custom
Your Fine-Tune
Bring your own fine-tuned model or let us fine-tune a base model on your data. Proprietary knowledge, domain-specific behavior, custom personas.
Growth + Scale tiers
Features

Everything you need for
production AI inference

🔌

OpenAI-Compatible API

Drop-in replacement for OpenAI's API. Change one URL and one key — your existing app, SDK, LangChain, or LlamaIndex integration works immediately.

🔒

Data Sovereignty

Your prompts and completions never leave our hardware or transit through hyperscaler infrastructure. Processing stays at the IX — no data sharing, no training on your data.

🏎️

vLLM Inference Engine

Industry-leading throughput with PagedAttention and continuous batching. Maximizes GPU utilization so your tokens cost less and arrive faster.

🌊

Streaming Responses

Server-Sent Events streaming for real-time token delivery. Build voice AI, typing indicators, and progressive UIs without waiting for the full completion.

📊

Usage Analytics

Per-key token tracking, latency histograms, error rates, and throughput graphs. Know exactly what you're using and how your applications perform.

📞

Voice AI Ready

Sub-5ms processing enables natural voice conversations. Integrate directly with FreeSWITCH, Asterisk, FusionPBX, or any SIP system for real-time phone AI.

Quick Start

Zero code changes.
Just change the endpoint.

Any application using OpenAI's Python SDK, Node.js SDK, or REST API works with Peering Edge immediately. Switch one URL, one key.

Python — OpenAI SDK
from openai import OpenAI client = OpenAI( # Change just these two lines: base_url="https://api.peeringedge.com/v1", api_key="pe_your_api_key_here", ) response = client.chat.completions.create( model="llama3.1-70b-instruct", messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Explain edge AI in one paragraph."}, ], max_tokens=512, stream=True, ) for chunk in response: print(chunk.choices[0].delta.content, end="")
cURL — Direct API
curl -X POST \ https://api.peeringedge.com/v1/chat/completions \ -H "Authorization: Bearer pe_your_api_key" \ -H "Content-Type: application/json" \ -d '{ "model": "mistral-7b-instruct", "messages": [ { "role": "user", "content": "What is edge AI inference?" } ], "max_tokens": 256, "stream": false }' # Response in <5ms from nearest IX # Same JSON format as OpenAI # Drop-in replacement, zero refactoring
FAQ

Common questions about
Edge AI Inference

Is this really a drop-in replacement for OpenAI? +
Yes. We implement the same OpenAI REST API spec — /v1/chat/completions, /v1/completions, /v1/embeddings. Change base_url and api_key in your client, update the model name, and your application works. LangChain, LlamaIndex, AutoGen, and most AI frameworks support custom base URLs natively.
How does the latency compare to OpenAI / Anthropic? +
For users in or near our IX locations, we're typically 40–50x faster to first token than OpenAI. For Starlink users in particular, cloud AI adds 200–400ms of round-trip network latency before inference even begins. Our IX placement eliminates that entirely — your first token arrives in under 5ms of processing time.
What happens to my data? +
Your prompts and completions are processed only on our hardware at the Internet Exchange. We do not log completions by default, do not use your data to train models, and do not share data with third parties. Enterprise customers can request air-gapped deployments with zero network egress logging.
Can I run my own fine-tuned model? +
Yes, on Growth and Scale tiers. Upload your GGUF or safetensors fine-tune and we'll deploy it to a model slot on our GPU infrastructure. We support LoRA adapters on top of base models. We also offer a fine-tuning service if you have training data but not the technical resources to do it yourself.
How does billing work? Is it really flat rate? +
Yes — flat monthly rate. No per-token billing surprises. Your 10M / 50M / 200M token allowance resets monthly. If you need more, we can discuss an overage rate or upgrade. Most customers find their allowance covers normal usage with headroom. Tokens are counted as input + output combined.
Ready to Deploy?

Go live in under an hour.

We'll provision your API endpoint, configure your model slots, and get you a working key. From sign-up to first inference — under 60 minutes.

Start Free Trial See Voice AI →