Predictable AI infrastructure costs, as you scale
Volume pricing, dedicated support and custom SLAs for teams processing serious traffic.
$1K – 5K
per month
Volume discount starts here
Rate reduction on your highest-volume models, applied automatically.
$5K – 20K
per month
Deeper discount + priority support
Named contact, priority routing, and capacity reserved ahead of burst traffic.
$20K+
per month
Custom contract + dedicated SLA
Negotiated rates, contractual uptime and latency targets, invoicing on your terms.
What you get beyond the rate
The parts that matter once this stops being an experiment and starts being a dependency.
- Contractual uptime and latency SLAs
- Dedicated capacity, reserved ahead of your burst traffic
- Named technical contact, not a shared queue
- Net-30 invoicing, PO and procurement paperwork supported
- Zero-retention inference available on request
- 30 days' notice before any rate change affects you
On compliance: we serve open-weight models under their own licenses, so there is no third-party resale exposure in your supply chain. We’ll share our current certification status and subprocessor list during evaluation rather than claiming more than we hold.
Tell us your volume
How the threshold and the process work
Written out because a procurement evaluation stalls on unanswered mechanics far more often than on price.
The threshold is $1,000 a month
Below that, the public pricing already is our answer — every rate is published per model with the official rate beside it, there is no platform fee to negotiate away, and there is no minimum spend to be released from. A volume conversation starts to be worth both sides' time at about $1,000 of monthly usage, which is the first band above.
What we need from you is a range, not a forecast
The form asks for company, an estimated monthly band, a work email and anything else worth knowing. The band is the only qualifier that matters at this stage; the model mix and traffic shape are what refine the rate afterwards, because a workload that is mostly repeated prefix prices very differently from one that is mostly output.
A reply within 24 hours, with indicative rates
The first response carries indicative rates for the volume you described, not a calendar link. If the numbers work, the paperwork follows: DPA, PO, net-30 invoicing and any regional processing requirement. If they do not, you have lost one email rather than a discovery call.
What changes as you move up a band
The first band is a rate reduction on your highest-volume models, applied automatically. The second adds a named contact and capacity reserved ahead of burst traffic. The third is a negotiated contract: rates, contractual uptime and latency targets, and invoicing on your terms rather than ours.
What we are not claiming
Constraints worth knowing before an architecture is drawn around us.
- Pay-as-you-go accounts carry no contractual uptime or latency target. Those exist only inside a negotiated agreement, and we would rather state that than publish an SLA nothing obliges us to meet.
- There is no multi-provider failover. Each model is served by a single upstream, and when it degrades the request returns an error with a machine-readable code rather than a substituted model — which is the right default for a licensed, priced dependency, but it does mean redundancy is your side of the line.
- The catalogue is deliberately narrow: open-weight models only, so no GPT, Claude or Gemini. That is a licensing position, not a roadmap gap — reselling closed-weight access frequently breaches the origin provider's terms, and that is not exposure we will put in your supply chain.
- We will share our current certification status and subprocessor list during evaluation rather than listing badges here. Anything we have not completed, we will say so.