for teams watching the AI invoice
Cut your AI bill, not your AI.
Most of what you pay a cloud API for is bulk traffic that never needed a frontier model. Wide Area Intelligence runs that baseline on GPUs you already own — no per-token fees — and bursts to the cloud only for the genuinely hard requests. Recurring OpEx becomes predictable CapEx, and your apps don't change a line.
/// where the money goes
You're paying frontier rates for non-frontier work
Look at an itemized AI bill and the pattern is almost always the same: a small slice of genuinely hard requests, and a long tail of autocomplete, classification, summarization, internal RAG, and bulk generation — work an open-weight model on commodity hardware handles fine. Every token of that tail is billed at metered cloud rates anyway.
Wide Area Intelligence lets you draw the line. Send the bulk to hardware you own, keep the hard part on the cloud model of your choice, and set the split per category. The baseline stops costing per token; you pay the cloud only for the burst it actually catches.
/// who this is for
Four ways the bill gets cut
The team with a bill-shock problem
Your combined Claude / OpenAI spend quietly hit five figures a month and finance flagged it. Route the cheap, high-volume traffic to hardware you own and keep frontier APIs for the hard 20% — cut the line item without cutting capability.
Inference is your COGS
Every customer prompt is a variable cost paid to a provider, and it's eating your gross margin. Moving steady inference onto owned nodes converts per-token OpEx into fixed CapEx — the exact unit-economics metric your investors watch.
High-volume batch work
Hundreds of thousands of prompts a month for copy, classification, and content. Run the bulk on the workstations your team already owns — on a schedule, overnight if you like — instead of metering every token to a cloud provider.
Outgrowing per-seat API credits
Paying thousands per engineer in API credits adds up fast. One capable server running an open-weight model can absorb the baseline and pay for itself in weeks — the intelligence stays in-house instead of evaporating into a monthly invoice.
/// how the routing works
The bill drops at every tier. Cache is free, your hardware has no token fees, cloud is burst-only.
EDGE CACHE
An identical request you've served before is returned straight from Cloudflare's edge — the prompt never reaches a GPU. Per-account toggle, your own TTL.
YOUR HARDWARE
A one-line install turns any machine with a GPU into a node. It opens a secure Cloudflare Tunnel — no port forwarding, no static IP — and serves llama.cpp behind your gateway with no per-token fees.
CLOUD FAILOVER
Node busy, offline, or timed out? The same request silently re-routes to the cloud through our managed gateway, billed only against prepaid credits — burst only, never the baseline.
/// drop-in integration
Same code, smaller invoice.
Change one line.
Nothing in your stack changes. Point your existing OpenAI-compatible client at the gateway and the baseline starts running on hardware you own — the cloud only catches what overflows. The first time you see the bill, the line you changed is the base URL.
- Change one line: the base URL.
- Works with the OpenAI SDK, LangChain, agents, curl — anything OpenAI-compatible.
- Bring your own gateway key; routing, caching, and failover are automatic.
from openai import OpenAI client = OpenAI( base_url="https://wideareaai.com/api/v1", api_key="wai_sk_…", ) resp = client.chat.completions.create( model="llama-3.1-8b-instruct", messages=[…], )
/// go deeper
Put a number on it
/// faq
Cost questions
- How does running AI on my own hardware lower the bill?
- A cloud API meters every token, forever. When a request is served by a GPU you already own through Wide Area Intelligence, there are no per-token fees — your only marginal cost is electricity. Repeated requests are answered from an edge cache for free. You keep paying metered cloud rates only on the burst traffic that overflows your hardware, not on the steady-state baseline.
- Don't I still need a frontier model for the hard requests?
- Yes, and you keep it. The point isn't to replace frontier models — it's to stop paying frontier rates for the 80% of traffic that is autocomplete, classification, summarization, internal RAG, and bulk generation. That bulk runs on open-weight models on your own hardware; the genuinely hard 20% still routes to the cloud model of your choice. You decide the split per category.
- What is the CapEx-vs-OpEx argument?
- Cloud API spend is recurring operating expense that vanishes every month with nothing to show for it. A workstation or a refurbished server is a one-time capital expenditure that depreciates predictably and keeps serving inference for years. For steady, high-volume workloads the hardware often pays for itself within weeks, then the marginal cost of inference drops to electricity.
- Do I have to rewrite my apps?
- No. Wide Area Intelligence is OpenAI-compatible, so any client that already speaks the OpenAI API — the OpenAI SDK, LangChain, agent frameworks, coding tools, or curl — works by changing one line: the base URL. Routing, caching, and cloud failover are automatic.
- What happens when my hardware is busy or offline?
- The same request silently fails over to the cloud through the managed gateway, billed only against prepaid credits. You never get a hard outage because a node was asleep or maxed out — the cloud is burst-only insurance, never the baseline.
/// initialize
Stop renting the baseline. Own it, and burst the rest.
no credit card · 2 nodes free, forever · openai-compatible