Paranine Launches as Pakistan's First Sovereign AI Factory, Running Frontier Models on Dual-Chip Infrastructure at 4,520 Requests Per Second.

Intelligent Inference is now paranine (P/9), launching as Pakistan's first sovereign AI factory: a full-stack platform that produces tokens from frontier open-source models on infrastructure based entirely in Pakistan, billed in Pakistani rupees. For anyone already building on it, the rebrand changes nothing about the products or the services, only the identity around them.

What follows is the launch announcement, published from Islamabad in September 2026. Where it quotes a figure, this post says who measured it and what has not been published yet, because a number without its conditions is not a benchmark.

the release as issued, with the media contact

What a sovereign AI factory means here

A sovereign AI factory is infrastructure that processes data inside one country's borders, serves open-weight models rather than a proprietary ecosystem, and bills in the local currency. paranine is not a research lab, a wrapper around a foreign API, or a consulting practice: it is a production inference platform reached through an OpenAI-compatible API.

The launch places Pakistan alongside France, the UAE, Saudi Arabia, India and Singapore, each of which has built or funded national AI infrastructure over the past two years.

what a sovereign AI gateway is, and what it has to be able to prove

The numbers being published today, and what they measure

Two things in that table need reading carefully. The first two rows are gateway figures: they measure how quickly the routing layer accepts and forwards a request, not how quickly a model generates tokens, and the model, sequence length, batch shape and serving stack behind them have not been published. A reproducible benchmark report against global providers is on the roadmap further down, and until it exists these are the company's own numbers rather than an independent result.

The second is the uptime row. 99.99% is a reliability target, not a contractual commitment on the shared platform: the terms of service offer no uptime SLA there, and an availability guarantee would have to be agreed in a separate contract.

what the terms of service actually commit to

Why it matters: the sovereign AI gap

In 2026 the global inference market has attracted more than 40 billion dollars in venture capital, and every one of the best known providers is headquartered in the United States.

• Groq signed a 20 billion dollar infrastructure deal with Nvidia.
• Together AI raised 800 million dollars at an 8.3 billion dollar valuation.
• Fireworks AI reached 17.5 billion dollars, and Baseten 13 billion.
• Cerebras partnered with OpenAI.

The sovereign response has moved just as fast. France backed Mistral, the UAE launched G42 with Microsoft behind it, Saudi Arabia created HUMAIN with 200,000 GPUs, and India announced BharatGen with a 10,000 crore rupee commitment. Pakistan, with more than 600,000 registered software developers and three million freelancers, was not on that map. This launch is the claim that it now is.

The funding and valuation figures above are other companies' numbers, quoted as the announcement states them from public reporting in 2026. Nothing in them is a paranine figure.

One API call, two chip architectures

Every major inference provider today runs on a single chip vendor: Groq on its own LPU silicon, Cerebras on its wafer-scale engine, and Together AI, Fireworks AI, Lambda and Baseten on Nvidia GPUs. That is one hardware supply chain, and one export decision away from being a problem.

paranine runs a unified orchestration layer across two of them, Huawei Ascend NPUs and Nvidia GPUs, which the company describes as in production today for financial institutions in Pakistan's banking sector. A single API call is routed to the hardware that suits the model, the latency target and the throughput required. Customers see one endpoint, one set of credentials and one bill; the complexity stays on the platform side.

On the public catalogue the Pakistan NPU routes are still marked coming soon, so the routes anyone can call today are the ones the models page lists with their rates.

what actually changes when a model is served on Ascend instead of CUDA

How paranine compares, honestly

The comparison below is deliberately not a claim to be larger, better funded or more widely deployed than companies that have raised billions. It is a claim to be a different thing: one jurisdiction, two chip architectures, local currency, open models.

how to choose an inference provider when location is a hard requirement

Why the rebrand: from inference engine to AI factory

The name carries the same argument. Para evokes parallel processing, the architecture underneath modern AI, and nine stands for the nines of reliability the platform is built to hold. Stylised as P/9, it is meant to be legible to the banks, businesses and institutions deciding on this technology without stripping out the depth that makes it work.

The term AI factory is borrowed deliberately. Nvidia's Jensen Huang has used it to describe sovereign AI infrastructure as the bedrock of a modern economy, and the position here is simple: if every nation needs one, Pakistan's is being built by a team that has already shipped one.

Pakistan's demand for inference is real, and underserved

Every token processed through a foreign API today is data leaving Pakistan, revenue leaving Pakistan, and a dependency that a single policy change abroad could sever.

• More than 600,000 registered software developers, per the Pakistan Software Export Board.
• More than three million freelancers, the fourth-largest freelance workforce in the world.
• An estimated 5 to 7 trillion tokens of annual demand across enterprise and developer use.
• A banking sector already deploying AI, which is where the platform's first lighthouse customers are.
• No domestic alternative to the OpenAI, Anthropic or Google APIs that keeps the data in the country.

what the PDPB expects of personal data that leaves the country

What people are building on it

The platform serves frontier open-source families including Llama, Mistral, Qwen and DeepSeek variants, and model selection stays with the customer rather than being fixed by the provider. The catalogue lists every route that can be called today with its rate per million tokens in rupees.

• Conversational assistants for customer service and internal operations.
• Document processing, extraction and summarisation at scale.
• Code generation and review.
• AI features added to an existing product through the API.

the model catalogue and what each route costs in rupees

What comes next

The near-term ambition is to be the default AI infrastructure for any organisation in Pakistan that needs inference and will not send its data abroad.

• Scaling the Huawei Ascend NPU cluster against banking-sector demand.
• Expanding the catalogue with fine-tuning and custom model hosting.
• A developer portal with interactive documentation and a sandbox.
• Transparent, reproducible benchmark reports against global inference providers.
• Partnership discussions with government and education institutions.

how the PK Endpoint keeps a request inside Pakistan

About paranine

paranine (P/9) is Pakistan's sovereign AI factory. It serves frontier open-source models from infrastructure based entirely in Pakistan, billed in rupees from a prepaid balance, through an OpenAI-compatible API, with an orchestration layer that runs across both Huawei Ascend NPUs and Nvidia GPUs. It serves financial institutions, enterprises and developers who need local data processing, regulatory compliance and independence from foreign AI infrastructure.

Press enquiries go to info@paranine.com. Logos, the boilerplate and the fact sheet are in the press kit.

the press kit: logos, boilerplate and fact sheet