About Paranine

Building Pakistan's sovereign inference layer.

From gateway to token factory.

Mission

Sovereign, accessible, locally accountable AI infrastructure.

Paranine exists to make AI infrastructure sovereign, accessible, and locally accountable in Pakistan and across the AMEA region.

Every enterprise, developer, and freelancer in Pakistan currently depends on foreign APIs to access AI. Every API call exports sensitive data to US or EU servers, bills in dollars, and can be cut off overnight by a policy change or export control. Paranine replaces that dependency with sovereign infrastructure: open-source models served on hardware inside Pakistan, billed in rupees, governed per request, and controlled by the customer.

The inference thesis

The real bottleneck on commercialising AI is no longer training. It is inference.

The systems that will consume tokens at scale are not chatbots: they are agents, enterprise copilots, autonomous workflows, and real-time retrieval. Every one of them needs inference infrastructure that is stable, low-latency, and scalable.

NVIDIA declared the training era over at GTC 2026, recasting the data centre as an AI factory whose product is tokens and whose profitability is measured in tokens per watt. The weights are free and improving. The scarce resource is the factory that turns electricity into tokens, and the gateway that bills for them.

Paranine is that gateway for Pakistan.

What we built

The Paranine inference engine

An NPU-GPU agnostic inference engine built by Paranine to work on both Huawei Ascend NPUs and Nvidia GPUs.

First NPU-native LLM gateway

Deploys pre-optimised open-source models and runs inference on clusters of Huawei NPUs. Up to 3x more tokens/second than existing model API platforms on the same frontier open-source models.

CUDA to CANN bridge

A unified interface to deploy models on both Huawei NPUs and Nvidia GPUs. Customers use inference on Huawei NPUs without knowing CANN. No learning curve.

15 MB Docker image

Complex gateway features in a footprint smaller than most log shippers. About 2 ms added per request at p50.

4,520 req/s sustained throughput

Benchmarked 15x faster than LiteLLM proxy on the same host.
What we built

The P/9 Gateway

OpenAI-compatible API

One endpoint across every open-source model Paranine serves. Drop-in replacement.

Per-key governance

Every request inspected. Anomaly alerts per key. Cost attribution per agent, per feature, per business unit.

PKR billing

First platform to localise LLMs. Invoiced locally in rupees. No forex approvals.

Intelligent routing

P/9 Router reads each prompt and routes to the best-fit model. Quality preserved, cost driven down.

Dedicated inference

Single-tenant model endpoints the customer owns and controls. Deploy, configure, and manage models in production through the UI.

Toolhub

Agentic AI development tools (web search, agent loops, function calling, memory, SQL connectors, document parsing) selectable per API key.
The market

Pakistan is missing a local LLM API platform.

Pakistan consumes approximately 5-7 trillion LLM tokens per month across 600,000 registered developers, 25,000+ enterprises, and 3 million freelancers. That number is growing exponentially month over month, with projections reaching 100 trillion tokens/month by 2028.

$25-45 million per month currently leaves the country for foreign AI APIs. Pakistan has a 42% developer account growth rate on GitHub. Agentic AI and LLM-RAG services contributed an additional $1.4 billion in software exports over the past 12 months.

The comparable in India, Sarvam AI, reached a $1.5 billion valuation processing 2 million API calls a day within two years. Pakistan’s developer landscape is no different: it is missing a local LLM API platform, priced in local currency via a local paywall.

The roadmap

Gateway to token factory

Paranine is executing a four-phase plan to build Pakistan's inference layer, and to exit into it.

Phase I: Confirm product-market fit (2026-2027)

  • Rent two bare-metal servers from Sky47 (rev-share already signed and live), reducing CapEx.
  • Token-based pricing: freemium model, pay as you go.
  • Askari Bank as the lighthouse deployment. Successful enterprise deployments across Askari Bank, HBL, Mahmood Group, IDC, 10+ software houses, 1,000+ direct developers.
  • Target: 100 billion tokens/month routed. ~$0.9M ARR.
  • Funded by: $500K pre-seed.

Phase II: Scale on Sky47 bare metal (2027-2028)

  • Add nodes across Sky47's footprint, including the 8.5 MW Islamabad campus (built to scale beyond 50 MW).
  • Still asset-light: Sky47 owns the metal, Paranine owns the demand.
  • Own factory: one CloudMatrix-class node, 384 Ascend chips, 559 kW, 1.44T output tokens/month.
  • Monetised share rises to 55%. Value-added revenue begins.
  • Target: 1.5T tokens/month routed. ~$8.7M ARR.
  • Funded by: ~$4M equity + $6.2M asset finance.

Phase III: The inference factory at Port Qasim (2027-2030)

  • A scale AI data centre at Port Qasim, Karachi: grid-adjacent to Bin Qasim generation, on industrial land, near cable landing stations.
  • KE Ventures / AsiaPak embark on building the facility. Paranine becomes the inference layer on top.
  • 4 nodes, 2.24 MW, 5.76T output tokens/month capacity.
  • Growth from local AI ventures born on the gateway, not from existing enterprises alone.
  • Target: 3T tokens/month routed. ~$20M ARR.
  • Funded by: ~$10M equity + $30M asset finance.

Phase IV: Export (2030 onward)

  • Factory becomes an export business. Tokens are the ideal Pakistani export: no shipping, no congestion, no perishability.
  • Sold into the Gulf and wider AMEA on the corridor the Pak-Saudi channel already opens.
  • Subsea capacity: SMW-6, Africa-1, and 2Africa now landed at Karachi.
  • EBITDA recycled as deposits compounds capacity to 30T tokens/month.
  • Target: 30T tokens/month routed. ~$166M ARR.
  • Exit range: $600M to $1B, a 66% to 79% discount to comparable US inference infrastructure companies.
Founder

Ahad Khurshid, Founder & CEO

Three years of extensive R&D on AI inference while delivering state-enterprise-level private LLM deployments. Generated $1.25M+ in government AI revenue through custom AI deployments across KSA government entities. Built and delivered sovereign private LLMs for multiple enterprises, on both NPUs and GPUs.

Previous affiliations include Pure Group, Afiniti, Saudi Aramco’s National Technology Development Program, Qualcomm, and IoT Squared.

Partners and investors

Who Paranine builds with.

Infrastructure partners

  • Sky47: supplies Huawei Ascend NPUs on a rev-share basis. Signed and live. Starting with 8-16 NPUs, scaling across Sky47's multi-campus footprint.
  • K-Electric (KE) / KE Ventures: the only privately owned utility in Pakistan that generates, transmits, and distributes. One counterparty for power, land, and grid. Provides the facility, land, power, and grid backstop for the Port Qasim inference factory.
  • FrontierOne ASGARD: security and assurance layer for regulated workloads.

Investors and strategic partners

  • AsiaPak Investments: investor and strategic partner for the Port Qasim data centre build.
  • Mahmood Group: strategic partner.
  • SPARS Group: strategic partner.
  • Daewoo: strategic partner.

Distribution partners

  • Nayatel, PTCL, Zong 4G, Huawei, Transworld Home: distribution and affiliate partners for mass developer reach.

Developer distribution

  • Influencer cohort already in coordination: Hisham Sarwar (2M subscribers), Azad Chaiwala (3M subscribers), Usman Asif (500K subscribers), Daniyal Sheikh (264K subscribers).
  • 100 hand-picked notable freelancer-developers. White-glove onboarding, free credits, direct line to the team. Target: 1,000 happy, active developers acquired by word of mouth.
Enterprise customers

In production with Pakistani banks.

  • JS Bank / Zindigi: in production.
  • Askari Bank: lighthouse deployment. LOI signed. Dedicated endpoints live. Billing active.
  • Target pipeline includes HBL, IDC, 10+ software houses (including NetSol, Arbisoft, and peers).
Competitive position

No foreign inference platform can match what Paranine offers.

  • Only NPU-native LLM gateway outside China.
  • Only inference platform with in-country residency for SBP-regulated workloads, billed in PKR on local rails.
  • Only platform offering dual-chip orchestration (Huawei NPU + Nvidia GPU) in a unified serving layer.
  • Only gateway with per-key governance, cost attribution per agent and feature, and anomaly alerting built into the request path.
  • Gateway footprint of 15 MB and ~20 ms added latency: lighter and faster than OpenRouter, LiteLLM, or Portkey.

Fireworks AI ($17.5B), Together AI ($8.3B), and Baseten ($13B) serve more tokens in a day than Paranine plans to serve in a month. Paranine is not competing on global throughput. It is competing on access, architecture, and jurisdiction: capabilities no NVIDIA-bound platform can replicate.

Contact

From Pakistan, for the AMEA region.

info@paranine.com · paranine.com

Cost-effective inference for every leading open model.