A
Low latency
Real-time responses for voice agents, copilots, and interactive products, with capacity that is never shared.
Dedicated capacity for production workloads
Reserved GPU capacity behind the same OpenAI-compatible endpoint, tuned to your model and traffic, isolated from public load. Live in under 24 hours, backed by a 99.99% SLA.
99.99%
SLA uptime, with service credits
<24h
from signed order to live endpoint
3
regions: us-east, us-west, eu-central
100%
capacity isolation, no shared GPUs
What you get
A dedicated deployment runs the same router and the same API surface as the public endpoint, 71µs p50 routing overhead and 235µs p99 included. What changes is who else is on the GPUs: nobody.
B
High throughput
Coding agents, batch pipelines, and parallel generation without queueing behind public traffic.
C
Predictable at scale
Stable latency and uptime you can plan launches against, backed by an SLA.
D
Tuned to the workload
Serving configuration matched to your model, traffic pattern, and context lengths.
Your own serving stack
Dedicated capacity is a separate deployment: reserved GPUs, its own router pool, its own request queue. A traffic spike on the public endpoint changes nothing about your latency, and your reserved headroom is never lent out to other tenants.
The API surface is identical to nyxprovider.com/v1, including byte-for-byte streaming and an immediate 429 at your reserved ceiling instead of silent queueing. Cutover is a base URL change, not a migration, and you can route back to the public endpoint at any time.
Read the cutover guide →Built for security review
The controls your security team will ask about, stated plainly. Compliance documentation ships with the order form.
SOC 2 Type IIReport available under NDA
Zero data retentionPrompts and completions never stored
US data residencyus-east-1 primary
Private networkingPrivateLink on request
How it lands
Three steps, and most of the work is ours. Typical time from the first call to the first production token is under 24 hours.
1
Scope the workload
A short call on your model, traffic shape, and latency targets. Bring a traffic sample if you have one.
2
Capacity comes up
We reserve and tune the GPUs, then hand you an endpoint that drops into your existing configuration.
3
Cut over when ready
Same API surface as the public endpoint, so the cutover is a base URL change, not a migration.
Common questions
Anything not covered here is a short email away: hello@nyxprovider.com.
What is the minimum commitment?
Dedicated capacity bills monthly, no annual contract required. Most teams start with a single replica and scale from there, and volume discounts start at a three month term.
Which models can we run?
Any open-weight model, including your own fine-tunes. If the weights run on our serving stack, and every model in the public catalog does, we can serve them. Custom weights come up on the same timeline: under 24 hours from a signed order.
What does the SLA cover?
99.99% monthly uptime on the dedicated endpoint, measured at the router. When we miss it, service credits apply to the next invoice automatically, no ticket required.
How does billing work?
A reserved GPU-hour rate for the capacity, plus per-token rates discounted below the public prices published at /v1/models. Both lines appear on one invoice, itemized per deployment.