Four services,
one meter.
Production tokens, managed serving, integration and regional delivery — all reported against the same consumption account.
Everything is measured in the same unit.
Services are not sold as separate products with separate bills. A workload that comes in through integration and runs on managed inference is reported against the same token meter as a direct API call.
That keeps the commercial conversation simple: what was consumed, at what rate, over which period.
What each service covers.
Token production
Inference billed against consumption. Requests run on dedicated capacity inside the isolated pool; output is counted at the boundary and reported per model and per tenant.
- Input and output tokens counted separately
- Volume bands agreed per order
- Reporting per model and billing period
Managed inference
Models deployed, served and kept current on our stack. Scheduling, batching, autoscaling and versioning sit with us, so the customer does not run a serving layer.
- Deployment and version management
- Latency and throughput monitoring
- Upgrades without customer-side migration
Integration
Retrieval, tooling and application wiring, so output lands inside the system that needs it rather than beside it.
- Retrieval over customer corpora
- Tool and function calling
- Evaluation before rollout
Regional access
Delivery into ASEAN markets from a Singapore contracting entity, with one point of accountability across markets.
- Single contracting entity
- Local delivery support
- Consistent metering across markets
How an order comes together.
Scope
A working session on workloads, models, expected volume and any residency constraints.
Capacity plan
We map the workload onto the reserved pool and return a delivery plan with bands and endpoints.
Order
Volume bands, models, endpoints and reporting are fixed in an order against the reservation.
Integrate
Endpoint credentials, a private path if required, and a reporting feed from the first billing period.
Why there is no rate card.
A per-token rate depends on the model, the context length, the latency target and the volume band. Publishing one number would mean quoting an average that fits no particular workload.
Rates are set after scoping, against the reservation that already exists. Once set, they hold for the term of the order.
- Rates follow the model, not the customer
- Volume bands step down as commitment rises
- Consumption is reported before invoicing
- Reserved capacity and delivered volume are billed separately
Start with the workload.
Describe what you intend to run. The commercial shape follows from it.
Request access