7 min readparanine
AI infrastructure in Pakistan: from NPU racks to an inference factory
AI infrastructure in Pakistan today: Ascend NPUs serving models inside the country, and paranine's published plan from rented racks to an inference factory.
Pakistan's AI infrastructure today is mostly a credit card pointed at a server abroad. The exception is a small but real fleet of Huawei Ascend NPUs serving open-source models inside the country, and a published plan to grow that fleet into an inference factory. This post describes both and keeps them clearly apart.
Everything here about plans is from paranine's own published roadmap, and is described as a plan.
What runs in the country now
paranine's PK endpoint runs on Huawei Ascend 910B NPUs, 64GB accelerators racked in Pakistan rather than rented by the hour in a foreign region. Paranine built the serving stack for these chips and operates the fleet; the team that wrote the stack is the team on call for it.
The routes served there today include gpt-oss 120B, Qwen3.8 27B, DeepSeek-OCR 2, Laya Decision and Qwen3 Embedding 8B. Prompts and completions for those routes are processed in Pakistan, and the gateway stores neither.
Why two kinds of chip
Every major inference provider runs on one chip vendor. For a country, that is one supply chain and one export decision away from losing access. paranine's serving layer spans Huawei Ascend NPUs and Nvidia GPUs, so the same route ID can be served from either and the calling code does not know which.
If a model is CUDA-native, Paranine translates its code to run on CANN, Huawei's compute stack, and serves it on NPUs.
The published plan, phase by phase
The figures below are targets in a plan, not measurements. What exists today is phase I.
| Phase | What it adds | Target routed |
|---|---|---|
| I (2026 to 2027) | Two bare-metal servers rented from Sky47, rev-share signed and live | 100 billion tokens a month |
| II (2027 to 2028) | More nodes across Sky47's footprint, plus one owned CloudMatrix-class node: 384 Ascend chips, 559 kW | 1.5 trillion a month |
| III (2027 to 2030) | An inference factory at Port Qasim, Karachi: 4 nodes, 2.24 MW | 3 trillion a month |
| IV (2030 onward) | Tokens as an export to the Gulf and wider AMEA | 30 trillion a month |
Why Port Qasim
An inference factory turns electricity into tokens, so power and connectivity decide where it goes. Port Qasim is grid-adjacent to Bin Qasim generation, on industrial land, and near the cable landing stations where SMW-6, Africa-1 and 2Africa reach Karachi. K-Electric, through KE Ventures, is the counterparty for power, land and grid.
What this means for a team building today
You do not need to wait for phase III to keep a workload in the country. The PK routes serve now, through the same OpenAI-compatible API as everything else, billed in rupees. The infrastructure plan matters mostly as a signal of how much capacity there will be, and of who will be accountable for it.
What a sovereign AI gateway is, and how to tell a real one from a badge
Common questions
- Is there AI inference infrastructure in Pakistan?
- Yes. paranine serves open-source models on Huawei Ascend 910B NPUs racked inside Pakistan, through an OpenAI-compatible API. Its first phase runs on bare-metal servers rented from Sky47 under a signed revenue share.
- Why NPUs rather than GPUs for AI in Pakistan?
- Supply and independence. A platform built only on one vendor's GPUs inherits that vendor's export exposure. paranine serves the same route from Huawei Ascend NPUs or Nvidia GPUs through one serving layer, so the hardware an AI workload depends on is not a single point of failure.
- Where will Pakistan's AI data centre be built?
- paranine's published roadmap places an inference factory at Port Qasim, Karachi, grid-adjacent to Bin Qasim generation and near the cable landing stations, with KE Ventures and AsiaPak building the facility. It is planned for 2027 to 2030 and is a plan, not a facility that exists today.