Compute Is Governance

For years, artificial intelligence policy was discussed primarily in terms of data, algorithms, ethics and regulation.

That is changing.

One of the most important resources shaping the future of AI is something much less visible to the public:

compute.

Compute refers broadly to the computational resources required to train, fine-tune, evaluate and run AI systems.

It includes advanced processors, servers, networking, cloud infrastructure, data centres and the energy required to operate them.

Without sufficient compute, many AI ambitions remain theoretical.

A national AI strategy can promise innovation.

A university can train talented researchers.

A company can develop sophisticated algorithms.

A government can create a regulatory framework.

But if the country lacks access to the infrastructure required to train, test, evaluate or deploy advanced systems, its practical AI capacity may remain limited.

This leads to a simple but increasingly important proposition:

Compute is not only infrastructure.

Compute is governance capacity.

Why Compute Matters

AI systems require different levels of computational resources.

A small machine-learning model can run on ordinary hardware.

More sophisticated systems require specialised accelerators.

Training frontier-scale models can require enormous clusters of advanced processors connected through high-speed networking.

Inference—the process of actually running a trained model—also requires substantial infrastructure when systems are used at scale.

Evaluation creates another demand.

If governments or independent institutions want to test advanced AI systems, they may need access to the systems themselves or to sufficient computing resources to reproduce experiments, conduct evaluations and test safeguards.

This creates an important governance relationship.

The ability to regulate AI increasingly depends partly on the ability to technically understand and evaluate it.

And technical understanding increasingly depends on infrastructure.

The Compute Gap

The distribution of AI compute is highly unequal.

The OECD notes that many national AI strategies do not adequately address domestic compute capacity, even though compute is increasingly necessary for implementing those strategies. It recommends that countries assess their existing capacity, future needs, security, sustainability and long-term sovereignty when developing national compute plans.

This is a significant observation.

A country may have an AI strategy without having an AI infrastructure strategy.

That creates a gap between policy ambition and implementation capacity.

The problem is particularly important for developing countries.

If domestic researchers cannot access sufficient compute, they may depend on foreign cloud providers.

If public institutions cannot evaluate models independently, they may depend on vendors.

If universities cannot experiment with advanced systems, local research capacity may stagnate.

If companies cannot afford computational infrastructure, innovation may remain concentrated in a small number of external platforms.

Compute therefore influences who can participate in AI development.

The Geography of Compute

AI infrastructure is not distributed evenly across the world.

Cloud providers operate geographically distributed data centres, but access to AI-capable infrastructure remains concentrated.

OECD data show that in 2025, 351 of more than 531 cloud availability zones operated by seven major cloud providers offered some level of AI-capable compute. The United States and China together hosted the largest number of AI-capable cloud availability zones in the OECD’s assessment.

This does not mean other countries have no AI capability.

It demonstrates something more fundamental:

AI capacity has a physical geography.

Servers have locations.

Data centres need land.

AI infrastructure requires electricity.

Chips move through global supply chains.

Cloud services operate under national laws.

These physical and legal realities matter for governance.

Compute and Technological Sovereignty

This is why compute has become part of technological sovereignty.

The concept does not necessarily mean that every country must manufacture its own chips or build its own frontier model.

That would be unrealistic for most countries.

Instead, technological sovereignty can mean having sufficient strategic capacity to make informed choices.

Can a country access compute when needed?

Can its researchers obtain meaningful access?

Can public institutions evaluate important AI systems?

Can sensitive workloads be processed under appropriate legal conditions?

Can the country maintain alternatives if an external provider becomes unavailable?

Can policymakers understand the infrastructure dependencies behind their national AI strategy?

These are sovereignty questions.

The EU’s Infrastructure Strategy

The European Union provides a useful example of this shift.

Europe’s AI policy increasingly combines regulation with infrastructure development.

The EU currently describes 19 AI Factories and 13 AI Factory Antennas as part of its emerging AI infrastructure ecosystem, with the network intended to connect computing capacity with universities, startups, industry and public authorities.

In July 2026, the European Commission launched a call to establish up to seven AI Gigafactories.

The initiative is backed by up to €10 billion in EU and national funding and is expected to mobilise at least €20 billion in private investment.

The proposed facilities are intended to combine large numbers of advanced AI processors with networking, cloud infrastructure and energy-efficient data centres.

The significance is broader than the facilities themselves.

Europe is treating access to compute as part of its technological strategy.

This is a different understanding of AI governance.

The question is no longer only:

How should AI companies be regulated?

It is also:

What infrastructure must Europe possess to remain capable of developing and governing AI?

Cloud and AI Infrastructure

The EU’s proposed Cloud and AI Development Act provides another example.

The European Commission proposed the legislation in June 2026 with the stated objective of strengthening Europe’s cloud and AI ecosystem, investment and infrastructure.

The proposal includes support for sustainable cloud and AI infrastructure and greater data-centre capacity.

This illustrates an important policy convergence.

AI policy is increasingly connected with:

cloud policy,

data-centre policy,

energy policy,

industrial policy,

research policy,

and digital sovereignty.

The boundaries between these policy areas are becoming less clear.

Compute and Semiconductors

Compute depends on hardware.

Advanced AI systems rely heavily on specialised processors and high-bandwidth memory, along with networking equipment and other components.

This makes semiconductor policy central to AI policy.

A country that lacks access to advanced processors may face constraints on frontier AI research.

A country with strong semiconductor manufacturing capabilities can occupy a strategically important position in the AI supply chain.

This is why export controls, supply-chain resilience and semiconductor investment have become increasingly connected to AI geopolitics.

The important governance lesson is that AI regulation cannot be completely separated from hardware policy.

The rules governing access to advanced computing can influence the development and diffusion of AI capabilities.

Compute and Security

There is also a security dimension.

Advanced compute can support beneficial research.

It can also enable capabilities that governments consider strategically sensitive.

This creates a difficult policy problem.

How should governments balance:

open scientific research,

commercial innovation,

international technology diffusion,

national security,

and misuse prevention?

The answer increasingly involves differentiated access.

Not every AI workload requires the same level of infrastructure.

Not every computing cluster creates the same security risk.

Not every model has the same capabilities.

This creates the possibility of risk-based compute governance.

But such systems also create difficult questions about measurement.

How much compute is significant?

Which types of chips matter?

Should training compute or inference compute be monitored?

How should distributed or cloud-based computing be treated?

How can regulators distinguish legitimate research from prohibited activity?

These questions remain technically and politically complex.

Compute Monitoring

If compute becomes strategically important, governments may increasingly want better visibility into its use.

Possible governance mechanisms could include reporting requirements, licensing, procurement controls, cloud-provider obligations and security standards.

But monitoring compute is not equivalent to controlling AI.

Compute is only one component of capability.

Algorithms, data, model architecture, talent and system design also matter.

A governance system that focuses too heavily on compute could therefore miss other pathways to capability.

Compute should be treated as one layer of a broader governance architecture.

The Energy Problem

Compute has another unavoidable requirement:

energy.

AI infrastructure operates inside the physical energy system.

The International Energy Agency reported in April 2026 that global electricity consumption from data centres increased by 17% in 2025, while electricity use by AI-focused data centres increased by around 50%.

The IEA also reports that capital expenditure by five major technology companies exceeded $400 billion in 2025 and was expected to rise substantially again in 2026.

These developments demonstrate that AI infrastructure is becoming an increasingly significant component of the physical economy.

AI policy therefore intersects with energy policy.

A country cannot simply announce large-scale AI ambitions without considering whether its electricity system, grid infrastructure and data-centre ecosystem can support them.

The Data-Centre Bottleneck

Data centres themselves are complex infrastructure projects.

They require:

electricity connections,

land,

cooling,

network connectivity,

construction capacity,

hardware,

security,

water or alternative cooling systems,

and long-term capital.

AI workloads can also create unusually high power densities.

The IEA’s 2026 analysis notes that AI-server power density has increased substantially and that rapid growth in data-centre demand is putting pressure on grids, supply chains and infrastructure planning.

This creates an important governance question:

Who decides where AI infrastructure should be built?

The answer may increasingly involve national governments, energy regulators, local authorities, utilities, technology companies and investors.

AI infrastructure therefore becomes a public-policy issue even when the underlying data centre is privately owned.

Environmental Governance

Compute also creates environmental questions.

The environmental impact of AI does not begin when a model generates a response.

It begins across the infrastructure lifecycle.

Chips require energy and materials.

Data centres require construction resources.

Servers require cooling.

Electricity generation has environmental consequences.

Hardware eventually becomes electronic waste.

This means responsible AI governance should eventually consider infrastructure impacts alongside model-level risks.

The OECD has specifically highlighted the need to measure AI’s environmental impacts across its lifecycle rather than treating AI sustainability as a single energy-consumption statistic.

This is an important methodological point.

There is no single number that captures the environmental footprint of AI.

The footprint depends on hardware, location, energy mix, utilisation, model efficiency, workload and lifecycle.

Efficiency Does Not Automatically Solve the Problem

AI systems are becoming more computationally efficient in many areas.

But efficiency can increase demand.

If each AI operation becomes cheaper, organisations may use many more operations.

This creates a possible rebound effect.

Therefore, efficiency improvements do not necessarily imply that total AI resource consumption will fall.

Governance needs to measure both efficiency and aggregate demand.

Compute and AI Safety

Compute also matters for safety research.

Advanced AI evaluation can require significant technical infrastructure.

Independent researchers may need access to models, controlled environments and computational resources.

Public AI safety institutes may need their own technical capabilities.

Universities may need sufficient compute to reproduce research.

Without access, independent verification becomes difficult.

This creates a potential asymmetry.

The organisations developing advanced AI may possess vastly greater computational resources than the institutions responsible for evaluating them.

That is a governance problem.

Who Audits the Powerful?

This connects directly to Part 15.

An effective AI assurance ecosystem requires evaluators who can actually test the systems they are assessing.

If frontier developers possess much more compute than public institutions or independent evaluators, the information and capability gap can become substantial.

Public technical capacity therefore matters.

Governments do not necessarily need to own frontier-scale infrastructure.

But they may need enough technical capacity to:

access relevant systems,

run evaluations,

analyse results,

conduct red-team exercises,

understand technical documentation,

and commission independent research.

Otherwise, regulatory oversight can become dependent on vendor-provided evidence.

The Procurement Problem

Compute also matters to governments as buyers of AI.

Public-sector institutions increasingly procure cloud services, AI models and digital systems.

If procurement contracts do not address infrastructure dependency, governments may become locked into particular providers.

This can create long-term strategic risks.

A procurement authority should therefore consider not only:

Does the AI system work?

but also:

Where does the system run?

Who controls the infrastructure?

What happens if the provider changes its terms?

Can data be migrated?

Can logs be retained?

Can the system be independently evaluated?

What happens if access is interrupted?

These are infrastructure-governance questions.

Compute and Universities

Universities are particularly important.

Academic AI research depends increasingly on access to computational resources.

But university compute can be expensive.

If advanced research is concentrated in a few wealthy institutions, the global knowledge ecosystem becomes more unequal.

This is especially relevant to the Global South.

A researcher may have excellent theoretical expertise but still be unable to conduct meaningful experiments because compute is inaccessible.

That can produce a structural research gap.

The result is not simply fewer AI startups.

It is fewer independent voices in the scientific and policy debate.

A Global Research Infrastructure Problem

This means compute access should also be understood as a research-equity issue.

International organisations, governments and universities could explore shared infrastructure models.

Regional compute centres.

Research cloud credits.

Open evaluation environments.

Shared GPU clusters.

International research programmes.

Public-interest AI infrastructure.

Such mechanisms could expand access without requiring every country to build its own frontier-scale data centre.

For developing countries, this may be more realistic than technological self-sufficiency.

The Bangladesh Question

For Bangladesh, compute policy deserves much more serious research.

The immediate question is not:

Can Bangladesh build a frontier AI data centre comparable to the world’s largest facilities?

A more useful question is:

What level of compute does Bangladesh actually need?

Different sectors have different requirements.

Universities may need research clusters.

Startups may need affordable cloud access.

Government agencies may need secure inference environments.

AI safety researchers may need evaluation infrastructure.

Public-sector developers may need controlled testing environments.

Schools and smaller organisations may primarily need access to efficient inference services.

This suggests that a national compute strategy should begin with demand assessment.

The OECD makes a similar point: countries should first understand existing capacity and future ecosystem needs before designing national compute strategies.

Bangladesh’s Strategic Options

A realistic national strategy could involve several layers.

The first is access.

Researchers and startups need affordable access to high-performance computing.

The second is security.

Sensitive public-sector workloads may require stronger infrastructure and governance than ordinary commercial applications.

The third is research.

Universities need predictable access to compute for experimentation and evaluation.

The fourth is procurement.

Government contracts should avoid unnecessary dependence on a single provider.

The fifth is regional cooperation.

Bangladesh may benefit from shared infrastructure arrangements rather than attempting complete domestic self-sufficiency.

The sixth is energy planning.

If domestic AI infrastructure expands, it should be aligned with electricity and data-centre planning.

The seventh is human capacity.

Compute without researchers, engineers and governance expertise does not create an AI ecosystem.

The Infrastructure-Capability Relationship

This last point is critical.

Compute alone does not create AI capability.

The relationship is closer to:

Compute + Talent + Data + Research + Capital + Institutions = AI Capacity.

Weakness in any one layer can constrain the others.

A country can build a data centre without producing strong research.

It can train researchers without giving them infrastructure.

It can purchase cloud access without developing evaluation capacity.

It can write AI policy without understanding the underlying technology.

Effective AI strategy requires alignment among these components.

The Geopolitics of Compute

Compute is also becoming a geopolitical resource.

Countries want access to advanced chips.

Technology companies want access to global markets.

Governments want resilient supply chains.

National-security institutions want to prevent sensitive capabilities from reaching strategic competitors.

Developing countries want affordable access.

The resulting system is inherently political.

This does not mean every compute policy is motivated by geopolitical competition.

But the strategic value of computing infrastructure increasingly affects international economic policy.

Compute Diplomacy

A new form of international cooperation may therefore emerge around compute.

Countries could cooperate on:

research infrastructure,

regional data centres,

AI safety evaluation,

energy-efficient computing,

standards,

capacity building,

and access for developing-country researchers.

This would be particularly important for countries that cannot economically justify large-scale domestic infrastructure.

Compute cooperation could become part of development policy.

AI Infrastructure as Development Infrastructure

This is an important shift in perspective.

Historically, development policy focused on roads, electricity, telecommunications and internet connectivity.

AI adds another layer.

Computational infrastructure may increasingly become part of knowledge infrastructure.

The countries that can access advanced computing can participate more deeply in AI research and innovation.

Those without access may increasingly depend on external platforms.

This creates a new dimension of the digital divide.

The Digital Divide Becomes the Compute Divide

The first digital divide concerned access to computers.

The second concerned access to the internet.

The next may increasingly concern access to advanced AI infrastructure.

But the compute divide is not simply about whether someone has a device.

A country may have widespread internet access while having almost no domestic AI compute capacity.

Its citizens can use AI services.

Its researchers may still struggle to build or evaluate AI systems independently.

This distinction between AI consumption and AI capability is strategically important.

Consumption is not sovereignty.

Access Is Not Capability

A country can access a foreign AI model without possessing the ability to understand its architecture, evaluate its behaviour or reproduce its results.

It can deploy cloud-based AI without controlling the infrastructure.

It can use AI products without developing domestic technical expertise.

Therefore:

AI adoption ≠ AI capability.

AI access ≠ AI sovereignty.

AI usage ≠ AI governance capacity.

This is one of the most important distinctions for developing-country AI policy.

What Should Governments Measure?

A mature national AI strategy should therefore track more than the number of AI users or AI startups.

It should examine:

available AI compute,

research access,

cloud availability,

data-centre capacity,

advanced chip access,

energy reliability,

AI research capacity,

evaluation infrastructure,

public-sector technical expertise,

and access to international research networks.

These indicators provide a much better picture of national AI readiness.

A Potential Atlas Research Framework

This creates a strong opportunity for Atlas AI Institute.

Atlas could develop a:

“National AI Compute Readiness Framework.”

The framework could examine countries across several dimensions:

Compute availability.

Affordability.

Research access.

Cloud dependency.

Energy resilience.

Data-centre capacity.

Evaluation infrastructure.

Talent.

Public-sector access.

Strategic autonomy.

Regional cooperation.

The framework should avoid simplistic rankings.

Its purpose would be diagnostic rather than competitive.

It could help identify where countries have infrastructure, where bottlenecks exist and which governance interventions are realistic.

For Bangladesh, such a study could become part of a broader national AI-readiness research programme.

The Governance Implication

Compute changes how we think about regulation.

Suppose a government creates sophisticated AI regulations.

But the government lacks technical staff.

It cannot independently test models.

Universities lack compute.

Public agencies depend entirely on foreign cloud infrastructure.

In that environment, formal regulation may exist without meaningful implementation capacity.

This is the central lesson:

Regulatory sovereignty requires technical capacity.

And technical capacity increasingly requires infrastructure.

The Future of Compute Governance

The next stage of AI governance may therefore include policies concerning:

compute access,

chip supply,

cloud infrastructure,

data centres,

energy,

research infrastructure,

public-sector compute,

and international infrastructure cooperation.

These issues may not traditionally have been called AI governance.

Increasingly, they should be considered part of the same ecosystem.

The Strategic Direction

AI governance began with principles.

It moved toward regulation.

Then implementation.

Then assurance.

Then frontier-model oversight.

Now infrastructure is becoming part of the governance conversation.

This is not a replacement for regulation.

It is an expansion of the governance problem.

The countries that can build and access AI infrastructure will have greater ability to develop AI.

The countries that can evaluate that infrastructure and its outputs will have greater ability to govern AI.

The countries that lack both may become dependent on decisions made elsewhere.

That is why compute matters.

The most important AI governance question may increasingly be not only:

What rules should govern AI?

but also:

Who has the technical capacity to make those rules meaningful?