Sovereign AI in Pakistan

The P/9 PK Endpoint: AI APIs with Pakistani data residency.

Access open-source models through one API while your prompts, requests, logs and billing records are processed in Pakistan, on hardware built for it.

Talk to our team

Live on the PK NPU pool today: gpt-oss 120B, Qwen3.8 27B, DeepSeek-OCR 2, Laya Decision, Qwen3 Embedding 8B. The gateway, the keys, the ledger and the global pass-through routes are live on this endpoint too.

Same OpenAI-compatible APINPU-native routing in PakistanData stays in Pakistan

PK data residencyHuawei Ascend 910BBilled in PKRPer-key scope and capsEncrypted in transitKeys stored as hashes

NPU-native inference in Pakistan

Huawei NPUs are going to get better and be everywhere.

There is a learning-curve challenge to using Huawei's Ascend NPUs.

But Pakistan, along with other countries that face Nvidia import challenges, will mostly have NPUs in the country.

We are here to bridge that gap. We give you highly optimised open-source model endpoints, with a sovereign gateway that works on both Huawei NPUs and Nvidia GPUs.

Dedicated support, and accountability for uptime, on Huawei NPUs.

Huawei Ascend 910B: 64GB NPUs, racked in country rather than rented by the hour in a foreign region.
One gateway, two kinds of silicon: The same route id serves from Ascend NPUs or Nvidia GPUs. Your code does not know which, and does not need to.
Accountable uptime: The people who wrote the Ascend serving stack are the people on call for it. Dedicated support, not a ticket queue.

Already built on CUDA

CUDA

CANN

CUDA to CANN translation.

If your model is CUDA-native and already prepared for deployment on GPUs, that is not a blocker. We translate the model code to run on CANN, Huawei’s Ascend compute stack, and serve your inference on NPUs with the performance numbers to show for it.

What the PK Endpoint is

Nothing crosses an international border.

What sovereign data residency means when you call the P/9 API from inside Pakistan.

When you call a model through a PK NPU-native route, your request is served on Paranine’s own Huawei Ascend hardware inside Pakistan. The prompt, the completion, and all intermediate state are processed on sovereign infrastructure. Nothing crosses an international border.

The gateway stores no prompts and no completions on any route. A request log keeps the model, key, status, timings, tokens, and cost: nothing of what was said.

What sovereignty means in practice

Residency, privacy, billing and compliance, in Pakistan.

Data residency

All inference runs on hardware physically located in Pakistan, operated by Paranine. Your data is processed under Pakistani jurisdiction.

No prompt storage

The gateway does not store the content of your requests or responses. The audit log records metadata: which key, which model, how many tokens, what it cost, how long it took.

SBP-aligned path

For regulated workloads in banking and financial services, the PK Endpoint is the only approvable path for LLM inference that keeps data inside the country.

Local billing

Billed in PKR on local rails. No foreign card, no dollar invoice, no SBP friction.

Full regulatory compliance

Pakistani jurisdiction end to end. No GDPR cross-mapping, no reliance on foreign data processing agreements.

Infrastructure

Huawei Ascend 910B NPUs, racked in country.

The PK Endpoint runs on Huawei Ascend 910B NPUs: 64GB accelerators racked in country rather than rented by the hour in a foreign region. Paranine built the inference serving stack for these chips and operates the fleet directly.

Dual-chip architecture

The same route ID serves from Ascend NPUs or Nvidia GPUs. Your code does not know which and does not need to.

CUDA to CANN translation

If your model is CUDA-native, Paranine translates the model code to run on CANN, Huawei's compute stack, and serves inference on NPUs.

Accountable uptime

The team that wrote the Ascend serving stack is the team on call. Dedicated support, not a ticket queue.

Capacity

Phase I rents two bare-metal servers from Sky47 (rev-share already signed and live). Phase II adds owned CloudMatrix-class nodes: 384 Ascend chips, 559 kW, 1.44T output tokens/month per node.

How the PK Endpoint differs from global routes

Paranine serves two kinds of routes through the same API.

PK NPU-native routes: served on Paranine’s own hardware inside Pakistan. Data residency guaranteed. Models include DeepSeek-OCR 2 and other open-weight models optimised for Ascend inference.

Global pass-through routes: served on partner infrastructure outside Pakistan (Google Cloud KSA, Groq, etc.). Faster time to model availability. No data residency guarantee. Clearly marked in the model catalogue.

Both route types are called through the same API, the same keys, and the same billing. The model catalogue marks which pool serves each route, so you choose the residency posture per workload.

Who needs the PK Endpoint

Teams whose data has to stay in Pakistan.

Banks and fintechs

Processing KYC, loan documents, compliance reports, or customer data through LLMs.

Government and defence

Workloads where data cannot leave Pakistani jurisdiction under any circumstances.

Software houses

Whose overseas clients require a data residency certificate as part of vendor qualification.

Any enterprise

Whose legal or compliance team has flagged foreign API data export as a risk.

Learn more

Residency architecture and compliance posture.

Full technical documentation on the PK Endpoint, including residency architecture and compliance posture: paranine.com/pk-endpoint, this page, with the data residency guide in the docs.

Questions: info@paranine.com

Residency, control, accountability

Built for teams that need to say where their AI ran, and in what currency it was paid for.

The PK Endpoint is for organisations that need guarantees about where AI data is processed, how a caller is constrained, and how the money moves. Each guarantee below is a property of the deployment as it stands, not a setting someone has to remember to switch on.

PK data residency

On the PK NPU-native routes, prompts, requests and outputs are processed on Ascend NPUs racked in Pakistan. Request telemetry, the audit log and the ledger live on infrastructure in Pakistan for every route.

A sovereign gateway

Keys, budgets and model scope are enforced on the gateway, in country. A key can be scoped to PK routes alone, so a caller cannot reach a pass-through route by accident.

Billing residency

Prices are quoted per million tokens in rupees, debited from a prepaid balance and settled in the same currency. No foreign card, no dollar invoice, no exchange-rate surprise.

Security and control

Traffic is encrypted in transit. P/9 stores only a SHA-256 hash of each API key, revocation takes effect on the next request, and every administrative change is written to an append-only audit log you can export.

Built-in guardrails

Support Pakistani AI. Keep your data in Pakistan.

The endpoint is built to grow NPU-native inference in Pakistan without cutting you off from global models when you need them. The goal is simple: more of the country’s AI served in the country, with residency you can enforce rather than hope for.

01
NPU-native first
The catalogue highlights the routes that run on Ascend in Pakistan and tags them plainly, so a team choosing where its tokens are served can see it at a glance.
02
Pass-through, labelled
Global routes are relayed to partner capacity outside Pakistan and are marked Global everywhere they appear: the catalogue, the playground picker, the dashboard's model card.
03
Your choice, per key
Scope a key to PK routes only and the gateway answers 403 for anything else. Residency becomes a property of the key, enforced on every request, not a checkbox someone has to remember.

Contact sales

FAQ

Questions we are usually asked.

Read the docs

What does sovereign AI mean in Pakistan?

AI that runs under the country's own control instead of being sent to a foreign provider. On P/9 that means a pool of routes executed on Huawei Ascend NPUs in Pakistan, marked PK NPU in the catalogue; routes relayed to partner capacity outside the country marked Global wherever they appear; billing in rupees against a prepaid balance; and API keys that can be limited to the models you choose.

What is the P/9 PK Endpoint?

It is P/9's OpenAI-compatible API, served through a sovereign gateway that runs in Pakistan, with a pool of routes that execute on Huawei Ascend NPUs racked in country. It is for teams that need to know where their AI requests are processed and in what currency they are billed.

What is the endpoint URL?

https://sovereign.paranine.com/v1 for every route. Chat completions are at https://sovereign.paranine.com/v1/chat/completions.

Are the PK NPU-native routes available today?

Live today: gpt-oss 120B, Qwen3.8 27B, DeepSeek-OCR 2, Laya Decision, Qwen3 Embedding 8B. Every route shares the same keys and the same billing, so moving between pools is a change of model string.

How is this different from calling a foreign API?

Three ways. The gateway and its records (keys, budgets, audit log, ledger, telemetry) run in Pakistan. The PK NPU pool executes the model in Pakistan. And the bill is in rupees against a prepaid balance, not a dollar invoice on a foreign card.

What happens when I call a Global route?

The gateway relays the request to partner capacity outside the PK NPU pool and bills it in rupees like every other route. Those routes are marked Global wherever they appear. If a caller must never reach one, scope its key to PK routes and the gateway refuses the rest with a 403.

Which models are on the PK Endpoint?

Live on the PK NPU pool: gpt-oss 120B, Qwen3.8 27B, DeepSeek-OCR 2, Laya Decision, Qwen3 Embedding 8B. Live on the global pass-through pool: gpt-oss 120B, Kimi K3, GLM-5.3, GLM-5.3 Fast, DeepSeek V4.1 Flash. The full list, with per-million-token prices in rupees, is on the Models page and in the dashboard.

Do I need to change my integration?

Only the base URL and the model string. The request and response shapes are OpenAI-compatible, streaming included, so an existing client library works as it is.

Can I limit what a key can spend or reach?

Yes. Each key can carry an expiry, a lifetime credit limit in rupees and a list of allowed models. The gateway enforces all three: 401 after expiry, 402 at the cap, 403 outside scope. The rest of the organisation's balance is untouched by any one key hitting its limit.

What does the gateway record about my requests?

Request telemetry for your own analytics and billing: timings, token counts, the route and the key, kept in Pakistan and shown back to you on the dashboard. Retention terms for request content are published in the documentation.

What currency am I billed in?

Pakistani rupees, end to end. Rates are quoted per million tokens in PKR, the balance is prepaid in PKR, and the gateway reports the exact rupee cost of each request in its debug metrics.

Can I reserve NPUs for my organisation alone?

Yes, as dedicated inference: Ascend NPUs held for one organisation, billed per hour of reserved capacity and quoted per organisation. See Dedicated inference.

Who is the PK Endpoint for?

Teams with residency, procurement or currency constraints: anyone handling customer data, internal knowledge or regulated workflows who needs to say where the model ran, and anyone whose finance team would rather not settle an AI bill in dollars.

Start building with AI while your data stays in Pakistan.

One OpenAI-compatible API, a sovereign gateway with per-key scope and caps, NPU-native routing in Pakistan, and a bill in rupees. Get a key now, or be first to know when the PK NPU pool opens.

Same API. NPU-native routing. Data stays in Pakistan. Questions: info@paranine.com