Seven AI hubs is the right design. It deserves ten times the money.

Disclosure: the author holds a personal, unpaid honorary advisory appointment connected to Pakistan's investment promotion work. ImpactLab has no institutional partnership with the Government of Pakistan, no Pakistani public body reviewed this issue before publication, and no one outside this publication had editorial input into it.

I have read a lot of national AI programme documents this year. On design, Pakistan's National AI Advancement Initiative is in the better half. On the specific question of geographic inclusion it is better than most European equivalents, which talk about regional balance and then site everything within an hour of the capital.

The allocation reported against it is Rs 524.64 million ^[2].

Seven hubs, two years. Roughly Rs 75 million each — about USD 265,000 per hub to establish it, operate it for twenty-four months, and incubate eighty companies. Before the seed grants, which on any realistic average would consume a majority of the total by themselves.

This is not a reason to be sceptical of the programme. It is the investment case for it. A well-designed instrument, under-funded by roughly an order of magnitude, is the easiest thing in development finance to fix — considerably easier than a badly designed instrument with money behind it, which is the more common situation and the one nobody can rescue.

The evidence

The design. Announced 27 July 2026 by the Ministry of IT and Telecommunication. Seven AI innovation hubs; 560 startups over two years; non-dilutive seed grants; national hackathons and awareness sessions; a large training-places target; stated priority for women and under-served regions; focus sectors in healthcare, agriculture, fintech, education, governance and smart cities. An operator is to be selected by competitive bid under a performance-based contract ^[1].

The allocation. Rs 524.64 million in the FY2026-27 Public Sector Development Programme, within Rs 16.29 billion for the IT and Telecommunication Division ^[2]. The Division-level figure is confirmed in the published PSDP reporting; the programme-level sub-allocation appears in trade reporting and we have not been able to verify it against the primary budget document. The arithmetic below is conditional on that number being right, and we will correct this issue if it is not. What is not in question is that the Division received a fraction of what it requested. The ministry did not choose this number.

The existing footprint it can build on. Ignite, the National Technology Fund, already operates National Incubation Centres in several cities, including most of the NAAI locations ^[3]. The Pakistan Startup Fund already provides equity-free grants on a rolling basis through a functioning diligence and disbursement process ^[4]. The National AI Fund, constituted in August 2025 from 30 percent of Ignite's research and development fund, separately provided for AI centres of excellence and regional hubs ^[5].

Status. Announced and partially funded. No operator contracted, no hub opened, no grant disbursed.

Why the existing footprint is the opportunity, not the objection

The obvious criticism of a seven-hub programme is that most of those cities already have a government incubator, and a second AI hub network was provided for by the same ministry a year earlier.

Read as duplication, that is a problem. Read as a delivery asset, it is the single most important thing about this programme.

A hub network built from nothing in seven cities on Rs 524.64 million is impossible. A hub network built inside existing incubation centres — sharing premises, administration, mentor networks, and the relationships with local universities and banks that take years to establish — is not obviously impossible at all. The marginal cost of adding an AI-specialised programme to a functioning incubator is a fraction of the cost of building a new one.

That reframing changes what the operator tender should be asking for. A tender that asks bidders to establish seven hubs will attract bids that cannot be delivered on the budget, and the honest ones will not bid. A tender that asks bidders to co-locate and specialise — running AI cohorts inside existing National Incubation Centre premises where they exist, and building new capacity only where they do not — is a procurement that a competent operator could actually win and deliver.

That is a two-hub build and a five-hub upgrade. It is a different programme on paper and the same programme in outcome, and it fits inside something much closer to the money available.

There is a related opportunity in the grants. The Pakistan Startup Fund already has the mechanism, the diligence process and the disbursement rails for equity-free grants. NAAI's grants do not need a parallel apparatus; they need a ring-fenced AI window inside an instrument that already works. Every rupee not spent standing up a second grant-making function is a rupee available to found a company.

What would unlock the full design

Three asks, in ascending order of cost.

Restate the targets against the money, or state where the rest comes from. This is the free one, and it protects the programme. Targets published above what the appropriation supports do not damage the ministry that set them so much as they damage the next programme, because the next appropriation is negotiated against the last delivery record. A ministry that announced two fully funded hubs with a published path to seven would be in a stronger position at the FY2027-28 budget than one that announced seven and delivered two. Pakistan has a queue of next programmes and should protect their credibility.

Publish an independent evaluation of the incubation centres. The government's own figures on the existing network are strong — more than 2,250 startups since 2017 and Rs 36.3 billion in revenue on the most recent account — but the ministry's project page reports materially different numbers, and no independent audit exists that I have been able to locate ^[6]. I suspect the real story is better than what can currently be proven, and that is the frustrating part. A verified track record is collateral: it is what a development finance institution, a philanthropic funder, or a corporate partner underwrites a co-investment against. An unverified one is a discount applied to every subsequent ask. Commissioning the evaluation is the cheapest way to raise the price Pakistan can charge for its own record.

Pair hub siting with a connectivity guarantee. Internet service in Mirpur, in Azad Jammu and Kashmir, was restored in late July after a shutdown reported at between forty-five and forty-eight days ^[7]. Siting hubs in under-served regions is the right instinct and I would not want it reversed. But an AI hub is a connectivity-dependent asset, and the programme document should carry a commitment — from the telecommunications authority rather than from the IT ministry — that designated hub sites are exempt from administrative service suspensions. That is a coordination ask, not a budget ask, and it would make the inclusion commitment real rather than symbolic.

The wider point about sequencing

There is a version of this argument that says Pakistan should stop announcing and start delivering. I do not think that is quite right, and it is worth saying why.

Announcements are how a state with limited fiscal room signals direction to actors it cannot pay. The compulsory university AI course, mandated in February, will do more for Pakistani AI capability over ten years than any hub network, and it cost the treasury almost nothing because it directed institutions that were already funded. The Digital Presence Proceeds Tax changed a multinational's operating footprint, as Issue 14 set out, and it cost nothing because it was a rule rather than a subsidy.

Pakistan is unusually good at this category of instrument — the rule, the mandate, the tax design — and it is the right category for its fiscal position. Where it struggles is the category that requires sustained appropriation across budget cycles, which is precisely what a seven-hub network is.

The strategic conclusion is not to announce less. It is to route as much of the AI programme as possible through instruments that do not require appropriation, and to reserve the appropriated money for the two or three things that genuinely cannot be done any other way. Under that test, the under-served regions are exactly where the Rs 524.64 million should go, and the cities that already have incubators should be served by a mandate rather than a build.

The objection worth taking seriously

The strongest objection is that this is too generous. Several overlapping AI institutions, a hub network provided for twice, and a delivery ecosystem that cannot currently reconcile its own published figures on the incubation centres — on this reading the problem is not funding but institutional discipline, and more money would be absorbed rather than deployed.

That is the serious version and it deserves a serious answer rather than a rebuttal. The reporting discrepancy is real and documented, and anyone advocating a larger appropriation has to reckon with it.

The answer is that the two asks are sequenced. The independent evaluation and the published division of labour come before the larger appropriation, not after it, and they are the conditions on which the larger appropriation becomes defensible. That is the standard structure of a results-based financing conversation, and it is available here. What would not be defensible is a tenfold increase into an ecosystem that cannot yet reconcile its own numbers.

What ImpactLab is doing

The Norway–Pakistan publication track is the answer in publication mode: analysis published annually, with a named diaspora and academic contributor list. As stated in Issue 11, there is no institutional partnership counterpart and we are not claiming one.

The Middle East ATLAS carries the Arabic-language and Urdu-language vector for AI literacy work under Pillar I, with a scoping note publishing in the fourth quarter of 2026.

Both sit alongside AI for Norway as the small-state and large-state halves of one question. The comparison keeps producing the same finding from opposite directions. Norway has money and no maintained inventory of the AI its own state runs. Pakistan has a mandate reaching every university in the country and roughly USD 265,000 per hub. Neither problem is a shortage of intent, and each country has the thing the other lacks.