Automation and the AI buildout
AI is becoming cheaper to use and more physical to build. The durable advantage belongs to operators who connect models to trusted workflows, power, data, evaluation, and accountable judgment.
Model access is not the durable advantage.
Artificial intelligence is not one market and it is not one software feature. It is a linked industrial system that starts with electricity and ends in a useful application. The largest economic gains may occur at the application layer, but every application still depends on chips, data centers, networks, models, data, and human systems that make an output trustworthy.
The global winners will therefore not be determined by model access alone. They will be the countries and firms that can coordinate the full stack, and the operators that can turn a capable model into a repeatable result. Pakistan does not need to reproduce every layer to participate. It does need to know which layers it can influence, which it must buy, and where dependence creates cost or risk.
Every AI application pulls on five layers.
Jensen Huang’s five-layer frame, applied to the cited evidence. The diagram shows dependency, not market size or value captured.
Software diffusion and physical buildout are happening together.
AI adoption is real but incomplete. In results published in May 2026, the U.S. Census Bureau reported that 19.8% of U.S. businesses used AI in the survey period ending May 3. Information firms reported 39.7% use and finance and insurance firms reported 33.9%. Census broadened the question in November 2025, so this is evidence of uneven diffusion rather than a clean historical growth series.
Some task-level gains are also real. A peer-reviewed study of 5,172 customer-support agents found a 15% rise in issues resolved per hour after a generative AI assistant was introduced. Less-experienced workers gained more, while the strongest workers saw smaller speed gains and a slight quality decline. The result matters because it shows knowledge transfer inside a defined workflow. It does not prove economy-wide productivity or safe automation in every domain.
At the same time, the supply side has become physical and capital intensive. The IEA reports that global data-center electricity demand rose 17% in 2025, while electricity used by AI-focused data centers rose 50%. Capital expenditure by five large technology companies exceeded $400 billion during 2025. The IEA projects that total data-center electricity use could roughly double by 2030 in its central case, but also identifies grid connections, power equipment, chips, finance, permits, and public acceptance as constraints. That path is conditional, not installed capacity.
Labor effects remain uncertain. The ILO's 2025 task index places one in four workers in an occupation with some generative AI exposure. Exposure is not a prediction of job loss. Prices, demand, bargaining power, organizational design, new tasks, and the quality of worker transition will shape the result.
The important players are competing across five connected layers.
In March 2026, Jensen Huang described AI as a five-layer system: energy, chips, infrastructure, models, and applications. It is a useful map of dependency, provided it is not mistaken for a neutral market forecast.
Energy
Microsoft, Alphabet, Amazon, Meta, utilities, independent power producers, and data-center developers are trying to secure generation, connections, cooling, and more flexible loads. The Microsoft FY2026 third-quarter call said the company added one gigawatt of capacity during the quarter and expected roughly $190 billion of capital expenditure in calendar 2026, including higher component costs. That is a company plan, not a guaranteed completed buildout.
Pakistan's advantage at this layer is not cheap, abundant, dependable power. It is an expanding base of engineers, distributed solar adoption, and urgent local demand for efficiency and reliability. The gap is firm electricity, transmission, grid management, financing, and transparent cost allocation. As of August 19, 2026, we found no credible official evidence of a delivered Pakistan power project dedicated to large-scale AI compute. That absence should remain visible.
Chips
NVIDIA and AMD design major accelerators. TSMC's 2025 Form 20-F describes the advanced processes and packaging used for AI and high-performance computing. Microsoft and Google are also deploying custom accelerators, including Maia and TPU systems, to reduce cost and control more of their stack.
Pakistan has electrical engineers, computer scientists, a diaspora in global semiconductor firms, and a possible opening in design, verification, embedded systems, packaging skills, and training. It does not have a verified leading-edge commercial fabrication base. On April 7, 2026, the Ministry of IT's official news register recorded an upskilling milestone under the National Semiconductor Human Resource Development Programme. Training is a useful signal, not evidence of a working fab, supply chain, or export industry.
Infrastructure
Microsoft, Amazon, Google, Oracle, CoreWeave, and other cloud and data-center operators are building dense computing systems, networks, storage, cooling, and orchestration. Alphabet's February 2026 earnings disclosure said it invested $91.4 billion in capital expenditure in 2025, with about 60% directed to servers and 40% to data centers and networking. These figures show the scale of action by one firm, not the total market.
Pakistan offers lower-cost technical labor, growing connectivity, and demand from public and private organizations. Its gaps are local high-performance compute, data-center power quality, cloud depth, high-speed network redundancy, procurement discipline, and published utilization data. The National AI Policy proposes a national compute grid, high-performance computing centers, and access for academic institutions. No current official source we reviewed proves that this promised national grid is already operating at the described scale.
Models
OpenAI, Google, Anthropic, Meta, Microsoft, and a growing set of open-model developers are competing on reasoning, multimodality, cost, safety, distribution, and control. Microsoft now emphasizes model choice, while Google couples proprietary models with its own chips and cloud. The important shift is that model access is broadening even while the most capable training runs remain concentrated.
Pakistan's practical advantage is language, culture, domain context, and people who can adapt existing models to local workflows. The gaps are high-quality Urdu and regional-language data, evaluation suites, research compute, safety testing, and institutions that can govern sensitive data. The National AI Policy calls for shared datasets and national or sectoral models. As of August 19, 2026, we found no authoritative public evidence of a Pakistan-trained frontier-scale model. That is not a failure. It is a boundary on what can honestly be claimed.
Applications
The application layer includes customer service, accounting, software delivery, health, agriculture, education, manufacturing, logistics, science, and government operations. This is where a model meets a user, a workflow, and a consequence. Census adoption data and the customer-support study show both movement and unevenness. The leading actors include established software firms, cloud platforms, domain specialists, and startups that can integrate AI into a process customers already need.
Pakistan's strongest near-term position is here. The country already sells digital services and has large domestic problems worth solving. The constraint is not access to a chat interface. It is the ability to design a bounded workflow, connect trusted data, evaluate outputs, secure the system, handle exceptions, and remain accountable. Pakistan's Economic Survey 2025–26 reports $3.388 billion in ICT export remittances in July to March FY2026 and $856.3 million from technology freelancers. It does not identify how much of that revenue came from AI.
Pakistan faces lower access costs and sharper capability pressure.
Pakistan faces falling access costs at the same time as routine digital work becomes easier to automate. That combination creates an opening and a warning. A firm can use global models to improve research, support, accounting, software, agriculture, health, or public administration without owning a data center. The same models can also compress the price of undifferentiated freelance and business-process work.
The rational response is to move from selling hours or generic output toward selling accepted outcomes. A defensible service combines a real customer problem, domain knowledge, proprietary context, measurable standards, security, exception handling, and a responsible person. Prompt technique alone is not a durable occupation or company strategy.
For government, the decision is not whether to endorse AI. It is where shared infrastructure and rules can lower costs without centralizing error or surveillance. Public procurement should specify data rights, evaluation, security, human override, and evidence of benefit before scale.
The credible opening is applied work, context, and accountable delivery.
Pakistan has a young technical workforce, an English-speaking services base, deep domain needs, lower operating costs in some roles, a globally distributed diaspora, and a large market in which local-language and low-bandwidth design matter. These are useful ingredients because applied AI needs people who understand both the work and the user.
The diaspora can add buyer access, standards, research relationships, apprenticeship design, and experienced management. Local operators can contribute domain context that a frontier provider will not gather on its own. Cost can help win an experiment, but trust, quality, and distribution must retain the customer.
The opportunity is asymmetric. Pakistan can create value in applications, evaluation, data preparation, system integration, domain services, and selected chip-design or infrastructure skills without claiming self-sufficiency in advanced fabrication or hyperscale cloud.
The hard gap lies between a fluent demo and a dependable system.
The most important gaps sit between demonstration and production: reliable electricity, dependable broadband, affordable compute, usable data, data rights, cybersecurity, evaluation, product management, domain supervision, and risk capital willing to fund patient capability.
Workforce risk is equally important. Entry-level tasks often serve as apprenticeship. If firms automate those tasks without redesigning how people learn, the experienced workforce of the future gets thinner. Operators should measure quality, rework, exceptions, auditability, and cost per accepted result, not only minutes saved.
Platform dependence remains unavoidable in the near term. Imported cloud, chips, and models expose firms to foreign-exchange costs, policy changes, service interruption, and shifting terms. That dependence can be managed through portable data, multi-model evaluation, clear exit plans, local operational competence, and careful selection of what must remain human.
Policy direction is clearer; delivery evidence is still limited.
- July 2025: The federal cabinet approved the National AI Policy 2025. Its training, compute, data, and institutional targets are commitments that require public delivery evidence.
- April 7, 2026: The Ministry of IT recorded a semiconductor upskilling milestone under its national human-resource program. This is evidence of training activity, not commercial fabrication capacity.
- May 2026: The Pakistan Economic Survey reported continued growth in ICT export remittances through March 2026, but did not provide an AI-specific revenue series.
- August 19, 2026 evidence boundary: We found no authoritative evidence that a leading-edge fab or the promised national AI compute grid was operational.
These signals matter because policy direction is becoming clearer. They do not yet establish execution, access, utilization, safety, or economic return.
PK Ventures is testing bounded workflows, not making a national technology claim.
PK Ventures is treating AI as an operating-capability question. The immediate work is to identify bounded professional workflows where source quality, acceptance criteria, human review, and customer value can be measured. The companion analysis on professional-services disruption examines that narrower operating problem.
This is not a claim that PK Ventures is building chips, models, or national infrastructure. It is a disciplined choice to work where evidence can be produced locally: a real workflow, a responsible operator, a measured outcome, and a clear stop condition when the system is not reliable enough.
Do not stop at agreement
Challenge the argument, add what we missed, or show us where it should become work.
Sources and scope notes
- [1]NVIDIA: AI Is a 5-Layer Cake
Jensen Huang's March 10, 2026 explanation of the energy, chips, infrastructure, models, and applications stack. It is a company leader's framework, not an independent market forecast.
- [2]U.S. Census Bureau: AI use by businesses
Reports nationally representative Business Trends and Outlook Survey results and the November 2025 wording change.
- [3]Generative AI at Work: final open-access paper
Author-hosted final version of the peer-reviewed Quarterly Journal of Economics study. It evaluates a staggered deployment to 5,172 customer-support agents and limits the finding to that setting.
- [4]ILO: Generative AI and Jobs, 2025 update
Maps task exposure across occupations and distinguishes exposure from job loss.
- [5]IEA: Key Questions on Energy and AI
Separates observed 2025 data-center demand and capital spending from uncertain projections through 2030.
- [6]Microsoft: Fiscal Year 2026 third-quarter earnings call
Describes Microsoft's current data-center capacity, custom silicon, model platform, and 2026 capital-spending plan. Company statements document actions, not independently verified performance.
- [7]Alphabet: Q4 2025 earnings transcript
Published February 4, 2026 and describes Alphabet's 2025 infrastructure investment, TPU and NVIDIA deployment, model activity, and application distribution. Company statements are treated as disclosures of action.
- [8]TSMC: 2025 Form 20-F
Filed April 16, 2026 and documents advanced-process and packaging capability used for AI and high-performance computing.
- [9]World Bank: Digital Progress and Trends Report 2025
Assesses connectivity, compute, context, and competency foundations for developing economies.
- [10]Pakistan Economic Survey 2025–26: Information Technology
Reports Pakistan's ICT export remittances and connectivity base through March 2026. The article follows the report's table and cover value of $3.388 billion, not a narrative unit typo in the report body.
- [11]Pakistan Ministry of IT: National AI Policy
Sets policy pillars and targets; these are commitments rather than delivered capacity.
- [12]Pakistan Ministry of IT: 2026 latest-news register
Provides dated official records of current programs, including the April 7, 2026 semiconductor upskilling milestone.
Related insights
Continue the thread.
Each note should move into the next useful question, not end as a standalone argument.