A production facility
for model output.
Reserved compute, run as a stable production system, sold as the output it produces. The unit of sale is the token; the unit of risk stays with the operator.
The machine is not the product. The output is.
A token factory is an operating model, not a piece of hardware. Compute is contracted as a dedicated pool, the inference stack is run as a production system, and delivery is measured in tokens rather than in machine hours.
That distinction matters because the two are sold very differently. Capacity is bought ahead of demand and paid for whether or not it is used. Tokens are counted after the fact and billed against what the application actually consumed.
Four layers, one boundary.
Each layer is operated by Xpert. The customer touches the last one and the meter that feeds it.
Capacity
Compute is contracted as a dedicated pool and physically isolated from every other workload. The pool is sized for production traffic, not for benchmarks.
Runtime
The inference stack is ours to run: scheduling, continuous batching, cache management, autoscaling and model versioning. Upgrades happen on our schedule, without a migration on the customer side.
Metering
Requests are counted at the boundary of the pool and split into input and output. Consumption is reported per tenant, per model and per billing period.
Delivery
Output leaves through an API or a private network path, in a request shape the application already speaks. Nothing needs to be installed to start.
Two ways to buy the same silicon.
The difference is not technical. It is a difference in who carries which risk.
| Dimension | Buying compute | Buying tokens |
|---|---|---|
| Cost structure | Fixed, paid whether used or not | Variable, follows output |
| Utilisation risk | Carried by the customer | Carried by the operator |
| Scaling | Procure, install, integrate | Raise the ceiling on the agreement |
| Operations | Serving stack is the customer's | Operated on our side |
| Time to first output | Weeks to months | Days |
| Reporting | Utilisation of machines | Consumption by model and tenant |
Three ways in.
API endpoint
An HTTPS endpoint with token authentication. Requests follow a common chat-completion shape, so existing client code ports with a base URL and a key.
Private endpoint
A dedicated network path between the customer environment and the pool. No public ingress, address allowlisting, and the same meter behind it.
Embedded
The model surface embedded inside a customer application, with metering handled at runtime and reported against the same account.
What we do, and what we do not.
We do
- Reserve and operate dedicated capacity
- Serve, monitor and upgrade models
- Meter consumption and report it per tenant
- Integrate output into existing applications
- Support delivery inside ASEAN markets
We do not
- Resell bare metal or rent GPU hours
- Co-locate other tenants inside the pool
- Commit to volumes we have not reserved
- Quote a per-token price before scoping
- Act as a model vendor or reseller
Before a first conversation.
What counts as a token?
A token is the unit of text the model reads or writes. Input and output are counted separately and reported separately, so the split between prompt volume and generated volume is always visible on the invoice.
Is the capacity shared with other customers?
No. The pool is reserved, physically isolated and dedicated to a single tenant. There is no shared tenancy inside the boundary, which is what makes a stable latency profile possible.
What is the commitment term?
Reserved capacity is held under a 60-month term. Delivery is then scoped against that reservation — volume bands, models, endpoints and reporting are agreed in a separate order.
Where is the capacity located?
In a designated data centre, contracted through the Singapore entity. Delivery into ASEAN markets is handled from that same contracting point.
Can we run our own model?
Open-weight models can be deployed on the pool. Licensing and provenance are reviewed before deployment, and the same metering applies.
Tell us what you intend to run.
Volume, models and timeline are enough to start. We will come back with a capacity and delivery plan.
Request access