What is verified. Pakistan’s National Telecommunication Corporation signed a three-year agreement with Data Vault Pakistan on 16–17 July 2026 to supply GPU-as-a-Service and Models-as-a-Service from government-managed data centres. The hardware named is NVIDIA H200, B200 and A100. The hosted models are Llama 3.1, GPT-OSS 20B, Qwen3 32B and Kimi K2 — American and Chinese open weights running side by side on the same government racks. The contract value was not disclosed, nor was a delivery schedule or go-live date.
What the deal is, precisely. This is a services procurement, not a facility. Sovereignty here is jurisdictional rather than industrial: the accelerators are American and export-approved, the models are foreign open weights, and what remains under national control is the data and the law governing it. The chosen models are also modest in size — this is a serving stack, not a training programme.
Unclear. The contract value. Whether any GPU has been delivered. When services go live. Whether the agreement was competitively tendered — no procurement notice or tender document has been published, and the record rests on company statements and local press reporting.
Why it’s on the tracker. Data Vault said in December 2025 it had secured more than 3,000 GPUs and was Pakistan’s only AI-enabled data centre. It is now also the supplier to the federal government and the national physics research centre, with a Punjab partnership in prospect. National compute sovereignty is being built on a single private vendor, days after Pakistan joined the Chinese-led World AI Cooperation Organisation — a hedge held in each layer of the stack rather than chosen between.