Dedicated endpoint
GPUs held for one workload.
Yours alone.
Dedicated Inference Endpoints
One Tenant. One Workload.
Throughput That Holds.
The single-tenant GPU platform for production LLM traffic, running your own weights or open models behind one OpenAI-compatible API, built for guaranteed throughput and predictable tail latency.
- Single-tenant GPUs
- US and APAC
- OpenAI-compatible
Single-Tenant GPUs, Reserved
for One Workload
Same API. Hardware That Stays Yours.
One tenant per GPU
No shared queue. No neighbor in your r99.
No cold starts
Weights stay resident for every hour held.
Uptime in contract
Availability and response terms, in writing.
Shared inference is priced per token because the capacity is shared. Yours is reserved.
Production Traffic That Needs
Its Own GPUs
Agents, voice, RAG at scale, regulated deployments.
Coding agents
Agents call in sequence, so the wait is the sum of every step. One warm endpoint for the whole fleet. No cold start mid-demo.
Real-time voice
Voice breaks on variance, not on averages. Sessions stay open. Streaming on warm capacity, pinned to a single region.
RAG at scale
Millions of queries a day, long context, tail latency that shows. Throughput bounded by hardware you hold, not by platform load.
Compliance-gated
Finance, healthcare, public sector. Procurement moves slower than the build. A deployment shape you can put in writing before anyone signs.
A shared tier costs less per token and more per incident.
Reserved Capacity, Burst
Above It
You set the floor. Everything above it scales on demand.
Throughput you set
The floor you reserve is the floor you get.
No rate limits
No platform ceiling on sustained calls.
Size to your floor
Reserve an ordinary day, not your worst hour.
Overflow served
Spikes run on demand, billed as they occur.
Resize a running endpoint
Replica counts change without a redeploy or a new URL.
Pooled inside your region
Overflow draws from the same region, so residency holds during a spike.
06:00
Util served
69%
Reserved capacity
82%
Headroom
13 pts
Headroom available in VPC
Bring the Model. The Runtime
Comes Tuned.
Fine-tunes, custom weights, and the serving stack to match.
Your own weights
Proprietary checkpoints in, a live endpoint out.
Fine-tuned models
Your checkpoint, behind an OpenAI-compatible URL.
Open source LLMs
Open models on our hardware, not brokered out.
Managed vLLM
We run the serving stack. You get an endpoint.
Matched hardware
GPU class chosen by memory and context length.
Long-tail models get the same runtime work as the popular ones.
The Open-Weight Catalog,
on GPUs You Hold
Language, reasoning, code, embeddings, speech, vision.
AILLM
anthropic/claude-fable-5
✦LLM
qwen-image-2.0-2026-03-03
AILLM
anthropic/claude-fable-5
AILLM
anthropic/claude-fable-5
Every model here runs on Inferel hardware, not brokered out to someone else's.
Where Per-Token Pricing Stops Winning
The more of the reservation you use, the less each token costs.
Crossover utilisation = Reserved cost ÷
(capacity in millions × your per-token price)
Above it, reserve. Below it, stay on per-token.
Two inputs you cannot estimate. The hourly rate, and sustained tokens per second for your model on the hardware you would hold. We measure both in the pilot.
TALK TO SALES →Dedicated or Serverless, Side by Side
Pick by the shape of your traffic.
DEDICATED ENDPOINT
Your traffic has a floor
Reserve hardware around the load you already sustain.
- Volume that never drops to zero, including overnight
- One or two models carrying most of the calls
- Your own weights, or a model nobody else serves
- A latency budget a shared queue would break
- A security review that asks about tenancy
INFERENCE SERVICE · PER-TOKEN
Your traffic has no floor yet
Start per token and move when usage becomes predictable.
- Bursty, seasonal, or still unknown volume
- Many models, swapped often, none dominant
- Evaluating models before committing to one
- Nothing measured yet to size a reservation against
- Below your crossover point, this is the cheaper option
Built for the Security Review
Regions, tenancy, retention, and the paperwork behind them.
Regions
US and APAC. Inference stays in the region you pick, overflow included.
Compliance
Security documentation is available during customer review.
Retention
Configured per deployment, including no storage, no logs, no training.
Frequently Asked Questions
A private, single-tenant deployment where reserved GPUs serve only your workload through an OpenAI-compatible endpoint.
Serverless capacity is shared and billed per token. Dedicated capacity is held for you, keeping weights warm and throughput predictable.
It usually wins when your sustained usage crosses the reservation’s utilisation threshold. We calculate that crossover with your model during the pilot.
Yes. Inferel can deploy proprietary or fine-tuned checkpoints and expose them behind the same API used by the model catalog.
Language, reasoning, code, embedding, speech, and vision models can be matched to hardware according to memory and context needs.
Your reserved capacity has no platform rate limit. The physical throughput of the hardware becomes the available floor.
Overflow can burst into pooled capacity in the same region and is billed on demand while your reserved floor remains available.
The GPU class is selected from model memory, context length, latency, and sustained throughput targets.
Inferel operates capacity in the United States and APAC. Exact regional availability is confirmed during sizing.
The service combines owned capacity, a tuned runtime, an OpenAI-compatible API, and contracted operating terms in one deployment.
Usually only the base URL and credentials change. Existing OpenAI-compatible clients can continue using their current request schema.
From Traffic Profile to Endpoint
Scope, measure, pilot, size. Then the reservation fits.