Artificial Intelligence and Computing

Chapter Overview

1. Executive Summary

This chapter begins with the question left by “Innovation and Entrepreneurship”: if Iraq can turn knowledge or a problem into a company, product and market, what infrastructure enables that company — and the state, university, hospital and factory — to use AI widely without every project becoming captive to an external provider, a single cloud or a model Iraq cannot audit or replace? The answer is neither buying more computers nor announcing a “national model” and then seeking a use for it. The required capability is an interconnected system: governance, data, computing, connectivity and energy, talent, tools and models, use cases, independent evaluation, and a clear right to stop or reverse course.

Iraq entered a new institutional phase in 2025–2026. Instructions No. (1) of 2025 on the Formations of the Prime Minister’s Office created the formal framework in which the National Centre for Artificial Intelligence operates. The Centre presents departments for AI applications, strategy management, policies, ethics and intellectual property. In 2026, a preliminary National Strategy for Artificial Intelligence and Smart Robotics 2026–2050 emerged, based on human capabilities, digital infrastructure, national industry, governance and legislation, with three stages beginning with sovereign and institutional foundations. This is a real institutional base, but not proof of complete capability.¹²

The strongest published Iraqi government-readiness figure comes from the National Centre’s assessment: only 19 of 64 institutions reported clear plans to improve their AI readiness, or around 29.7%, while the website shows strategy completion at 40% in 2026. Both figures should be read as working baselines, not international rankings: Iraq has begun measuring itself, but is still aligning plans, responsibility and readiness across institutions.³

Enabling infrastructure presents a better picture than the common impression that Iraq is “unconnected”. World Bank data based on ITU put internet use at 81.47% of the population in 2024, while ITU reports LTE/WiMAX coverage at 98.5% and at least 3G coverage at 99.2%. However, fifth-generation services had not entered commercial operation at this chapter’s evidence cut-off; the Communications and Media Commission announced a target launch in 2027. Access is therefore widespread, but AI readiness cannot be reduced to coverage: it needs stable access, data centres, computing, electricity, skills, usable data and institutions that know what they want from AI.⁴⁵⁶

In computing, the Ministry of Planning formally discussed infrastructure combining local physical servers and cloud servers in April 2026, alongside a sovereign national language model, “Iraq LLM”. In July, the National Centre announced the “Sumer Platform” to build the first Iraqi large language model. These are meaningful steps, but raise harder questions than an announcement: what is actual computing demand? What data and usage rights exist? What tests are required? What are the costs of training, operation and energy? And what merits local development rather than using an open, commercial or cloud model?⁷⁸

The chapter’s central decision rejects the “sovereign or cloud” binary. Iraq’s architecture should be hybrid and portable: local computing nodes for sensitive and critical functions, licensed Iraqi data centres for hosting and operation, multiple clouds for flexibility and variable capacity, and edge devices where necessary, with a unified standard for identity, permissions, logs, backup, encryption and exportability. In 2026, the Communications and Media Commission approved the Data Centre Services Regulation, bringing an organised local layer closer to implementation than before, but not yet establishing its capacity, quality or readiness for AI workloads.⁹

In regulation, on 23 September 2026 the Communications and Media Commission released the “Draft Regulation for Artificial Intelligence Services in the Republic of Iraq” for 30 days of public consultation. It stated aims of regulating use, protecting users’ rights and data, and promoting security, safety, transparency and innovation while respecting digital sovereignty. As a consultation draft at the data cut-off, it is not treated here as law in force. It is, however, evidence that governance has moved from general ethical discussion to a regulatory process requiring alignment with the Constitution, sectoral laws and the national AI strategy.¹⁰

There are early Iraqi applications that should be treated as learning cases. During the 2026 Arbaeen pilgrimage, the Communications and Media Commission reported that mobile companies used a machine-learning tool analysing data from more than 6,400 towers and thousands of cells to predict traffic and demand. The example illustrates an appropriate form of AI: a measurable operational problem, streaming data, prediction, human intervention and a network outcome. It does not establish that every sector is ready, or that every prediction should become an automatic decision.¹¹

In talent, the Ministry of Higher Education established ten AI colleges according to a June 2026 announcement and approved AI engineering clubs in five universities, with expansion expected. The University of Baghdad’s College of Artificial Intelligence opened tracks in engineering applications, biomedical applications and big data. The National Centre also runs training programmes and robotics and AI clubs. These expand “educational supply”, but there is still no published national baseline for ML/AI/MLOps practitioners, data engineers, evaluation and safety specialists, or the annual demand gap. This chapter therefore does not invent a promotional count of “AI experts”; it begins by measuring occupations and skills, then builds a capability ladder.¹²¹³¹⁴

The implementation conclusion is that Iraq needs neither a “Ministry of AI” nor a huge GPU centre without workloads. It needs a shared capability system: a national register of public AI use cases; risk classification; inventories of computing, data and skills; a shared computing platform purchased in demand-led stages; Iraqi language and evaluation infrastructure; a talent pathway from informed users to researchers, engineers and auditors; a use-case portfolio measured by outcomes before roll-out; procurement standards preventing vendor lock-in; and an evaluation laboratory testing accuracy, bias, security, privacy and impact. AI then becomes part of state and economic capability, rather than a display of tools.

2. The Central Question and the Chapter’s Scope

The central question is: what institutional, computing, human and regulatory capabilities does Iraq need between 2027 and 2045 to use AI widely and reliably in the state, economy, education and security, develop valuable local capabilities where viable, and prevent critical dependence or high-impact harm, without attempting to replicate the American, Chinese or European systems at their scale?

2.1 What the Chapter Resolves

• Define AI readiness as systemic capability, not simply the presence of a strategy or tools.

• Design national computing infrastructure: demand inventory, hybrid architecture, shared computing platform, operation, measurement and staged expansion.

• Place Iraq LLM / Sumer Platform within economic and technical logic: language assets, evaluation and adaptation before assuming a huge model must be trained from scratch.

• Build a talent ladder: users, practitioners, data and ML/MLOps engineers, researchers, sector experts and AI auditors.

• Select use cases in government, the economy, education and security by impact, measurability and risk.

• Govern risks, procurement and dependencies: system register, impact assessment, documentation, testing, human oversight, redress and supplier exit plans.

• Stages for 2027–2045, indicators, implementation and funding programmes, and rights and security safeguards.

2.2 Boundaries That Prevent Duplication

The chapter does not repeat R&D from V3-D06-C02 or incubators and venture capital from V3-D06-C03. It does not build digital identity, payments or shared government services; those belong to V3-D06-C05. It does not settle detailed data sovereignty, cybersecurity or critical infrastructure; those belong to V3-D06-C06, with only the minimum needed for AI governance used here. Nor does it design smart factories or industrial robotics in detail; that is V3-D06-C07.

3. Operational Definitions and Measurement Rules

4. The Iraqi Baseline: From AI Initiatives to National Capability

4.1 An Institutional Structure Taking Shape

The National Centre for Artificial Intelligence within the Prime Minister’s Office system is a central coordination point. Its official page divides its work among AI applications, applied research and clubs; strategy, policy, ethics and intellectual-property management; and administrative and legal support. This is initially appropriate for coordination and standards, but success must be measured by the unified government readiness and evaluation cycle it establishes, not by activity counts.¹⁵

The preliminary 2026–2050 strategy proposes long stages extending to 2050, whereas Iraq Vision’s horizon is 2045. There is no conflict: the Vision uses the current government process as an existing asset, but builds targets to its own 2045 horizon and requires periodic review. Technology does not permit freezing a document for 24 years; principles should remain long-term, while programmes, computing and models are reviewed every 12–24 months.

4.2 Baseline Indicators That Can Be Established

This dashboard prevents two opposing mistakes: claiming Iraq starts from zero, and claiming that a centre, strategy and internet coverage mean “the infrastructure is ready”. Readiness is the ability to turn these assets into reliable, productive systems, a gap to measure in 2027 rather than assume away.

5. State and Institutional Readiness

The first challenge is not “persuading every ministry to use AI”. The reverse is true: uncontrolled proliferation must be prevented until an institution has a use-case owner, data, a legal basis, an outcome measure and risk controls. The 19/64 result means most covered institutions do not present a clear readiness plan; the first objective is therefore a common planning method, not buying tools for everyone.

5.1 A Unified Institutional Readiness Card

5.2 A National Systems Register, Not a Shopping List

The state begins with a register of AI systems used or planned by covered entities. It does not publish security secrets or sensitive data, but the feasible minimum: purpose, owning entity, system type, risk level, impact-assessment status, whether it affects rights, benefits or safety, and whether oversight and redress exist. Sensitive security systems are registered internally under appropriate oversight, not exempted from inventory because they are secret.

6. Enabling Digital Infrastructure

Wide mobile internet coverage is an important base, but is not computing infrastructure. AI depends on four interconnected networks: connectivity carrying data and requests; data centres hosting storage and computing; electricity and cooling sustaining workloads; and specialised/cloud computing executing models. LTE coverage may be 98.5%, yet a national AI system may fail if latency, outages, power, inference costs or storage capacity are unsuitable.

The announced 2027 5G launch is useful for cases needing lower latency and greater capacity, but is not a universal prerequisite for AI projects. Government document processing or a knowledge assistant may work efficiently on existing fixed/cloud infrastructure; edge analytics, robotics, vehicles and some industrial applications may benefit more from 5G/edge. The Vision’s rule is to choose technology from the use case, not the use case from the technology.

7. The National Computing Gap

The verifiable public sources at the research cut-off contain no unified national inventory of operational GPUs/accelerators across the state, universities and companies, their capacities, ages, utilisation, locations, costs or expected workloads. This is not a call to publish sensitive details; it is a decision gap. Without an inventory, it is impossible to know whether the problem is insufficient hardware, fragmentation, low use, weak networks, power or software, or difficulty for researchers and companies in accessing it.

7.1 Before Purchasing: Forecast National Workloads

7.2 Success Criterion: Usable GPU-hours, Not GPU Counts

Unit counts are easy to announce but reveal little. A successful platform measures actually available computing hours, utilisation, waiting time, task cost, energy, job failures, support quality and the share of workloads portable to another provider. A unit left unused because of network, software or skills limitations is not full national capability.

8. Target Computing Architecture: Hybrid, Shared and Portable

The Vision proposes a “shared national AI computing system”, not a single centre monopolising every workload. It comprises a national computing marketplace/portal distributing work across local government, university and private capacities, licensed Iraqi data centres and multiple clouds, with sovereign or critical nodes where needed. The state owns policy, identity, standards, logs and exit contracts; it need not own every device.

This design applies the meaning of sovereignty established in earlier chapters: independence is not isolation. Foreign chips and services and global clouds may be used when economically and technically superior. The relationship becomes a sovereign risk when Iraq cannot move its workloads or data, understand service-termination conditions, access its logs or sustain a critical function when contracts, policies or markets change.

9. Iraq LLM and the Sumer Platform: Language Capability, Not a Prestige Project

The Ministry of Planning discussed “Iraq LLM” as a sovereign national language model, followed by the National Centre’s announcement of the Sumer Platform to build an Iraqi language model. The Vision takes this pathway seriously but rejects the assumption that “training a foundation model from scratch” is the only definition of sovereignty. Costs and methods change rapidly: according to AI Index 2025, inference prices for performance equivalent to GPT-3.5 fell from around $20 per million tokens in November 2022 to $0.07 in October 2024, alongside rapid hardware and efficiency improvements. Decisions for 2045 must remain updatable, not lock the country into a 2026 architecture.¹⁶

9.1 The Iraqi Language Capability Ladder

10. Data, Evaluation and Testing

This chapter does not construct the full data-sovereignty policy — that belongs to Chapter Six — but needs an operational rule: no consequential public model enters service because its demonstration is “impressive”. There must be test data separate from development data, function-appropriate metrics, tests on Arabic and Iraqi context, deliberate failure cases, and documentation of the version, model, source and cost.

10.1 A National AI Assurance Lab

An evaluation laboratory/network should be functionally independent of the development team and draw on universities, regulators and sector experts. It does not centrally approve every product; it sets protocols, tests high-risk and cross-agency systems, builds benchmarks and accredits subsidiary laboratories. The NIST model proposes a Govern–Map–Measure–Manage cycle and stresses continuous testing and monitoring throughout the lifecycle, not only at purchase.¹⁷

11. Digital Talent: From an “AI Course” to a Career Ladder

Expanding colleges and clubs gives Iraq a talent channel, but the most dangerous mistake is equating completion of an AI course with engineering capability to operate a production system. National capability needs complementary layers. Singapore, for example, distinguished Creators, Practitioners and Users and targeted expansion to 15 thousand practitioners, then updated its strategy in 2026 towards broad capabilities and organised computing access. The lesson for Iraq is not the number, but classifying skills by function.¹⁸¹⁹

The 2027 priority is not a total talent figure, but an “AI Occupations Observatory” linking education to demand: actual workers by role, vacancies, salaries, skills, time to fill positions, attrition and emigration, and pathway graduates. Only then should 2030 and 2035 numbers be determined by expected demand, not promotional comparisons with another country.

12. AI in Government: From Chatbots to State Productivity

The best initial government use cases combine measurable impact and manageable risk. Summarising correspondence, classifying applications, internal legislative research with sources, extracting document data, forecasting service demand, detecting duplication or financial anomalies for human review, and routing reports are better starting points than systems automatically determining citizens’ entitlements, guilt or employment.

12.1 Government Use-Case Gateway

Every project begins with a limited trial and demonstrates impact against a baseline. If it does not reduce time, cost or error, or increase satisfaction/access within risk limits, it should stop even if technically advanced. Government does not need “AI” when simpler procedures, a database or a search engine solve the problem at lower cost and risk.

13. AI in the Economy: A Productivity Engine, Not a Separate Sector

The greatest economic value comes not only from AI companies, but from introducing AI into sectors addressed in Part Four: industry, agriculture, energy, transport, finance, telecommunications and services. An industrial company using computer vision for maintenance and quality may create more value than a company developing a chatbot without a market. AI support should therefore be linked to a sectoral indicator: less waste, higher productivity, less downtime, better quality, greater sales or lower cost.

The telecommunications sector’s Arbaeen experience is a useful example because the measure precedes the technology: predict pressure and improve resource allocation. The same logic should govern other sectors. Public funding supports not “AI use” but a verifiable productivity outcome, preferably with private beneficiaries bearing a growing share of costs once value is demonstrated.

14. AI in Education and Research

Education faces opportunity and risk simultaneously. Generative AI can assist explanation, feedback, translation, exercise development, and researchers’ coding and search. But it may weaken assessment if submitting machine output replaces learning, leak student data or create fabricated information. UNESCO’s guidance on generative AI in education calls for a human-centred approach, privacy protection, pedagogical validation and age- and context-appropriate design.²⁰

14.1 The Education Rule: AI Literacy Belongs in the Curriculum, Not a Cheating/Banning Binary

• Update assessment towards explanation, defending work, projects and tests measuring understanding, rather than merely detecting machine-generated text.

• Specify when AI may be used and how it must be disclosed; distinguish assistance from plagiarism.

• Do not enter personal data, grades, or health or psychological records into an unapproved public service.

• Develop institutional assistants linked to documented content and curricula, with logs and clear teacher roles.

• Use AI to support teachers and researchers, not replace their professional responsibility for assessment or publication.

• Introduce data, modelling, evaluation, security and ethics skills across non-computing disciplines too.

15. AI in Security: Analytical Capability under the Law

Part Three established that security is not equivalent to collecting more data and that intelligence does not automatically become evidence of guilt. This applies particularly strongly to AI. Lawful uses may include detecting cyber anomalies, triaging reports, logistics analysis, image/map support, network-load or maintenance forecasting, and linking patterns for an analyst. Risk becomes high when a system suggests detention, force or surveillance of a person, or creates an unchallengeable “risk score”.

UNESCO’s global recommendation stresses human rights, dignity, transparency, fairness and human oversight. It states that ethical and legal responsibility must remain attributable to individuals or legal entities and that AI must not replace ultimate human responsibility. The European Union provides a comparative example of risk-based regulation, including strict restrictions on some biometric and law-enforcement uses. The Vision does not import EU law, but transfers the principle: higher risk requires higher barriers, and some uses may warrant prohibition rather than “improvement”.²¹²²

16. Risk Governance and Human Responsibility

Governance is not an “ethics committee” reviewing after implementation. It is lifecycle design. NIST divides risk management into Govern, Map, Measure and Manage; OECD emphasises human rights, transparency, comprehensibility and contestability of outputs, robustness, security and accountability. Iraq’s Vision can unify these principles in one government system, then tailor it by sector.¹⁷²³²⁴

16.1 Four Proposed National Risk Categories

16.2 Minimum AI Impact Assessment

• Purpose and alternatives: why AI rather than a simpler solution?

• Affected parties, rights and potential unintended consequences.

• Data sources, quality, representation, legal basis, retention and access.

• Model/provider, version, known limitations and processing location.

• Pre-operation measurement, test sets, thresholds and critical failure cases.

• Human intervention level: who can override/stop the system, and when?

• Explanation/disclosure and redress according to context.

• Security, red-teaming, supply chain and updating.

• Monitoring, incident response, rollback and decommissioning plan.

• Periodic review after changes to the model, data or purpose.

17. Public Procurement and Supplier Dependence

The greatest practical risk may be not a “chip monopoly” but a bad purchasing contract: a provider owning prompts, preventing embedding export, changing the model without notice, withholding logs, imposing high egress costs, retaining data for training or closing an API. Procurement therefore becomes an instrument of sovereignty.

Buy/build rule: The state should not build a model simply because it can, or buy a service because it is cheaper today. It compares five years of cost, risk, dependence, speed and skills. For a general low-sensitivity workload, a commercial API may be best; for a stable critical service, a locally hosted open model may be best; for a distinctive language or sectoral advantage, local adaptation may be the point of value.

18. Energy, Cooling and Sustainability

Computing is not only an ICT matter. The IEA’s 2025 report projects, in its base case, that global data-centre electricity consumption will roughly double from around 460 TWh in 2024 to around 945 TWh in 2030, with rapid growth driven by accelerated servers and AI. This is a global figure, not an Iraqi forecast, but establishes a planning principle: computing, electrical capacity and cooling have different lead times and must be planned together.²⁵

Iraq should not allocate unmeasured subsidised electricity to AI centres merely for “strategy”. Every large computing project needs a grid and power location, lifecycle cost, PUE, cooling and water, resilience, backup power, ambient temperature assessment, scalability and outage alternatives. Where electricity is constrained, purchasing flexible external computing for some workloads may offer greater economic sovereignty than building a local asset that cannot operate.

19. Comparative Lessons: Transfer the Mechanism, Not the Scale

19.1 India: Computing as a Shared Service, Not Merely Government Inventory

IndiaAI Mission was approved in 2024 with seven pillars covering computing, data and models, applications, skills, company funding, and safe and trusted AI. It began with a target of more than 10,000 GPUs, with announced capacity later expanding. Government used empanelment of cloud and data-centre providers to make computing available to universities, companies and public entities. The lesson for Iraq is demand aggregation and computing access under clear pricing and governance, not copying India’s GPU count or budget.²⁶

19.2 European Union: An “AI Factory” Combining Computing, Data and Talent

European AI Factories connect supercomputing centres to universities, small companies, industry and funding; by 2026, 19 factories and 13 regional antennas were being deployed. GPUs alone do not create an ecosystem; successful access points combine computing, support, data, talent and applications. Iraq needs a smaller, distributed version connecting universities, companies and ministries rather than an isolated technology centre.²⁷

19.3 Singapore: Differentiating Talent and Updating Strategy as Markets Change

NAIS 2.0 distinguished industry, government and research activity; people and communities; and infrastructure and environment. It targeted practitioner expansion, then established an AI council in 2026 and updated priorities towards broad capabilities, resource efficiency and green computing access. The lesson is periodic governance and functional differentiation, not importing the institutions of a small high-income state unchanged.¹⁸¹⁹

19.4 International Standards: Proportionate Risks and a Lifecycle Approach

UNESCO, NIST and OECD converge on principles Iraq can turn into procedures: rights, dignity and human oversight; clear use context; measurement and monitoring throughout the lifecycle; appropriate transparency; robustness and security; and attributable responsibility. Good regulation does not impose identical burdens on a drafting assistant and a system affecting rights or safety; it increases the burden with risk.

20. Iraq in 2045: What Does “National AI Capability” Mean?

In 2045, success will not mean Iraq “owns a model”. It will mean a ministry, university or company can move from a problem to a reliable AI system through a known pathway; access computing within predictable time and cost; find data, language assets and tests; employ a qualified team or provider under a contract with an exit route; subject high-impact systems to independent evaluation; and allow citizens to know when a system affected their rights and challenge it where law requires.

21. Transformation Stages, 2027–2045

22. Indicator and Target Dashboard

The following targets are proposed policy commitments, not forecasts. Where no published baseline exists, the 2027–2028 target is to establish measurement; no 2026 figure is invented to beautify the table.

The Oxford Government AI Readiness Index ranking is not a numerical target, despite its comparative value. Its 2025 methodology changed substantially from 2024, broadening the question from government use to government as buyer, enabler and regulator, with pillars covering computing, infrastructure, data, governance, adoption, skills and safety. The Vision uses these dimensions as an external checklist, not a promotional ranking ladder.²⁸

23. Implementation Programme Package

24. Implementation and Funding Matrix

The chapter sets neither a total national computing cost nor a GPU count until the 2027–2028 inventory is complete. Costs change rapidly and hardware prices and efficiency improve quickly, according to AI Index; fixing a twenty-year procurement figure would be poor engineering, not ambition. Funding is staged: initial capacity tied to proven workloads, then expansion when utilisation, demand and energy thresholds are met.¹⁶

25. Risks and Safeguards

26. What Do We Not Do?

• We do not make training the largest language model a standalone national objective; we build language assets, data, tests and use cases first.

• We do not announce GPU counts as strategic targets before workload inventory or buy capacity without knowing how it will be used or operated.

• We do not store all data within Iraq as the sole definition of sovereignty or send everything to the cloud as if it were always cheaper.

• We do not demand an AI project from every ministry; we demand a problem and baseline, using AI only if it outperforms the alternative.

• We do not turn every prediction or score into a decision or make humans rubber stamps for automated outputs.

• We do not use AI to create unrestricted political surveillance or classify citizens by loyalty or opinion.

• We do not equate an AI college or training course with a production-ready practitioner; competence is demonstrated through projects, skills and work.

• We do not copy the EU AI Act, NIST or another country’s strategy as Iraqi law; we transfer mechanisms and align them with the Constitution, law and reality.

• We do not automatically grant external providers data ownership or the right to train models on data in exchange for a “free service”.

• We do not bind the state to one model, provider or closed format where lock-in is an unjustified risk.

• We do not consider “AI” a cure for poor data or bad government procedures; it may accelerate error rather than correct it.

• We do not hide power, cloud, licensing, network and support costs when assessing AI viability.

27. Conclusion and Bridge to Public Digital Infrastructure

Within the Vision’s horizon, AI will become a normal layer of the state and economy, like databases and communications, but should not become an independent end. National strength appears when Iraq can choose suitable technology, access computing, build and employ skills, operate models under clear responsibility, measure impact, change or stop them, and protect rights and continuity. Technology whose benefits cannot be measured or harms audited is not “capability”, however advanced its name.