Paranine wants AI to speak rupees.

Consider the invoice. An AI assistant can answer beautifully and still create an awkward purchasing decision. Which currency pays for it? Where did the customer’s words travel? Who can stop a runaway application before its next request? Paranine, an AI infrastructure company in Pakistan, has built its proposition around those less glamorous questions. Its OpenAI-compatible gateway serves open-weight models, metered against a prepaid rupee balance. Local accountability is what Paranine sells.

That distinction explains the company better than the phrase “sovereign AI factory,” which it adopted when Intelligent Inference became Paranine in September 2026. Sovereignty here has several moving parts: the gateway, its records, the hardware executing the model and the money paying for the output. A local API address does not make all of them local.

Two chips. One front door.

The engineering ambition is to bring Huawei Ascend NPUs and Nvidia GPUs under one orchestration layer. The accelerator families use different software environments. Nvidia applications commonly use CUDA; Ascend uses Huawei’s CANN stack. Paranine says it translates model code for CANN and manages the serving machinery behind a common interface. Developers avoid learning a second accelerator toolchain.

Its Pakistan endpoint describes Ascend 910B units with 64GB of memory, racked in the country. The idea: make available hardware useful to people who do not want to become hardware specialists. This puts the company between model developers and businesses buying inference, with an additional role as an operator of local compute.

The public rollout has a clear boundary. Global pass-through routes are live; the Pakistan NPU-native pool is marked “Coming soon” in the catalogue and “Opening shortly” in the residency documentation. Global routes execute through upstream providers outside Pakistan, although the gateway, telemetry and billing records remain in Pakistan. A workload that requires local inference needs an available PK route, not simply a locally managed account.

The useful number is Rs 75.20

On its pricing page, Paranine lists gpt-oss 120B at Rs 30.08 per million input tokens and Rs 180.47 per million output tokens. A million tokens read, followed by a quarter-million tokens generated, therefore costs about Rs 75.20 at those published rates. Separate the reading from the writing, price both, then add them.

The gateway is prepaid and pay as you go, with no subscription or seat licence. New organisations receive Rs 1,000 in starter credits. Enterprise capacity and terms are quoted separately. For dedicated inference, the meter changes: customers reserve Ascend capacity and pay for the hours they hold it. Heavy, steady traffic can make that attractive; a machine sitting idle still costs money. Paranine’s dedicated page explicitly recommends shared capacity when the reservation would go underused.

There is a third revenue stream: embedded engineering, sold through monthly PKR retainers. A named engineer helps with integration, routing, monitoring and optimisation; larger engagements add engineers and procurement documentation. The service helps turn an API response into a dependable feature.

A key that knows when to stop

For a developer, the gateway can power a support assistant, document summariser, coding tool or AI feature inside an existing product. For the person paying, the more interesting object may be the API key. Keys can carry a model scope, an expiry and a lifetime rupee cap. Spend can be attributed by key, project and model. Experiments need not consume the organisation’s entire balance.

The controls have teeth. An expired key returns 401, a reached credit cap returns 402 and a request outside the permitted model scope returns 403. A key can be restricted to PK routes, preventing accidental access to global routes. The company says its request log retains operational metadata, including tokens, timings and cost, while excluding prompt and completion text.

Budget protection also creates a practical trap. Paranine’s OpenCode guide explains that an explicit maximum output allowance can trigger a worst-case budget reservation. A coding agent may receive a 402 even when a short test request succeeds. The fix is to check the key’s remaining headroom against the requested output allowance. A connection test cannot establish that every agent request fits.

Read the small number carefully

Paranine publishes a gateway result of 4,520 requests per second and a median latency through the gateway of 2.28 milliseconds. The test uses a deterministic mock upstream on a two-vCPU host. It measures the layer that authenticates, checks budgets and forwards requests. It does not measure how quickly a real language model writes an answer. Keep those measurements separate.

The same care applies to reliability. The launch material describes 99.99% as a target. The shared-service terms offer no uptime SLA; dedicated commitments require a separate agreement. Readers comparing Paranine with OpenRouter, Together AI or DeepInfra should therefore start with their own requirements: route availability, residency, price, workload performance and the contract they can actually obtain.

Why the engine became a factory

CEO Ahad Khurshid links the new name to work across Huawei and Nvidia hardware, sovereign compliance and banking customers. In the announcement, he says the team had begun “serving Pakistan’s largest banks.” That is the company’s account of its enterprise work. Its public developer offer also addresses smaller teams who want model access, local billing and controls without building an inference department.

The team’s careers page supplies a useful glimpse of the operating culture: engineers own kernels, batching and runbooks, carry problems across the serving stack and dashboard, and take the calls when shipped software pages them. Documentation belongs in the repository. An easy interface requires that operational discipline.

“Inference was where we started”Ahad Khurshid · CEO

Copy the questions, then test the answers

A practical evaluation starts with a small set of real tasks. Paranine’s Urdu guide recommends testing Urdu script, Roman Urdu and mixed-language messages in the proportions customers actually use. Score names, amounts and register separately from fluency. Its local tokenizer examples show why an English token budget cannot simply be assumed to fit Urdu. The useful lesson travels well beyond Pakistan: measure the traffic you have.

Then check the route’s location, give each project a budget, compare input and output costs, and read the availability terms before making a promise to a customer. Paranine’s opportunity lies in making those decisions manageable through one interface. The winning model is a beginning. Does the arrangement fit the people who operate it and pay for it?