Prime Minister Anwar Ibrahim used a June 9, 2026 special lecture at the University of Tokyo to warn that artificial intelligence is being shaped by a small cluster of countries and companies controlling the chips, cloud platforms, models, standards, and capital behind the technology. His argument was not that AI is moving too quickly, but that its center of gravity is moving too narrowly.
That distinction matters. The usual debate about artificial intelligence is framed as a race between innovation and regulation, as though society must choose between speed and safety. Anwar’s speech pointed to a more uncomfortable reality: the future of AI may be decided less by what the technology can do than by who owns the machinery that makes it possible.
Anwar’s warning landed in Tokyo, but its target was global. The infrastructure of modern AI is not evenly spread across the world; it is concentrated in a few layers of technical dependency. Advanced chips, massive data centers, foundation models, cloud distribution, and platform standards are not abstractions. They are chokepoints.
That makes AI different from many earlier waves of software. A country could build websites, mobile apps, or back-office systems with relatively modest capital and commodity hardware. Training and deploying frontier AI, by contrast, depends on scarce accelerators, industrial-scale electricity, high-end networking, elite engineering teams, and cloud contracts that can run into eye-watering sums.
This is why Anwar’s remarks should not be dismissed as another generic call for “ethical AI.” The ethical language is there, but the political economy is the story. If a handful of actors control the compute, the models, the payment rails, and the compliance frameworks, then everyone else is not merely adopting technology. They are renting access to someone else’s civilization stack.
For smaller and middle-power states, that raises the obvious sovereignty problem. Governments want AI for public services, education, health care, policing, cybersecurity, and economic modernization. But if the underlying systems are built, tuned, priced, and governed elsewhere, then national policy becomes downstream of foreign infrastructure.
Compute is not simply “more servers.” It is the combined availability of advanced chips, data center capacity, power, cooling, specialized software, and technical labor. In the AI era, compute determines who can train models, who can deploy them at scale, who can experiment cheaply, and who can set the defaults that others must accept.
That is why Anwar’s reference to chips, energy, cloud infrastructure, and capital cuts to the center of the issue. The world does not have an evenly distributed AI market in which every country chooses among neutral tools. It has a hierarchy. At the top are the firms that design the most sought-after accelerators, the companies that manufacture them, the hyperscalers that rent them, and the model builders that turn them into commercial products.
Below that are enterprises and governments that integrate AI into workflows. Further below are users who encounter AI through search engines, office suites, phones, social media, customer service systems, and public-sector portals. Each layer has less bargaining power than the one above it.
The uncomfortable part is that much of this hierarchy is invisible to ordinary users. A chatbot appears as a text box. A copilot appears as a button in an application. A recommendation system appears as convenience. But behind the interface is a supply chain of compute and policy decisions that most users, and many governments, cannot inspect.
For WindowsForum readers, this should feel familiar. Microsoft has spent years turning Windows, Microsoft 365, Azure, GitHub, and security tooling into a connected platform rather than a collection of products. That strategy has advantages for administrators: centralized identity, policy management, telemetry, endpoint security, and increasingly AI-assisted operations. But it also means that AI adoption often arrives through bundled enterprise ecosystems rather than discrete, easily separable choices.
The same dynamic applies beyond Microsoft. A company that builds on a hyperscaler may gain world-class infrastructure overnight, but it also inherits that provider’s pricing, regional availability, security model, acceptable-use policies, and strategic priorities. If the provider changes terms, sunsets a model, restricts a capability, or shifts pricing, customers can find themselves with limited leverage.
This is not a conspiracy; it is platform economics. The more valuable a platform becomes, the more customers build around it. The more customers build around it, the harder it becomes to leave. AI intensifies that lock-in because models are not interchangeable in the way virtual machines once appeared to be. Prompts, embeddings, fine-tuning, retrieval systems, evaluation pipelines, security reviews, and user workflows all create friction.
When Anwar warns that powerful actors may decide “what civilisation entails,” the phrase can sound grandiose. But in practical IT terms, it means something precise: platform owners set defaults, and defaults become behavior.
According to multiple reports, Anthropic resisted demands to remove restrictions that would have allowed its models to be used for mass domestic surveillance or fully autonomous weapons. The company’s stated position was that frontier AI systems are not reliable enough for such uses, and that mass surveillance poses grave democratic risks. Whether one views that stance as principled, strategic, self-protective, or all three, it exposes the governance problem.
Who decides what a powerful model may be used for? The vendor that built it? The government that contracts for it? The military command that sees operational advantage? The public that bears the consequences? The courts, after the fact?
That is the unresolved issue Anwar was pointing toward. Democracies are built on accountability, but AI systems are often procured through contracts, embedded in classified settings, and governed by terms of service few citizens will ever read. The result is a democratic gap: public power exercised through private systems under opaque rules.
It is tempting to say governments should simply regulate the vendors. But the state is not always the neutral referee. In national security contexts, it may also be the customer demanding more capability and fewer constraints. That turns AI governance into a three-way contest among public accountability, commercial control, and state power.
The problem is that voluntary safety commitments do not resolve the legitimacy question. A private company can refuse a use case today and change its mind tomorrow. A board can be replaced. A revenue crisis can sharpen priorities. A government can threaten procurement consequences. A competitor can offer fewer restrictions and win the contract.
This is why Anwar’s emphasis on public institutions is important. AI ethics cannot survive as corporate noblesse oblige. If core democratic decisions are delegated to private AI providers, then the public is left hoping that founders, executives, and investors maintain the right instincts under pressure.
That is not governance. It is dependency with better press releases.
The challenge, however, is that public institutions are not automatically ready to do better. Many regulators lack technical capacity. Many procurement departments are outmatched by vendors. Many legislatures move slowly. Many agencies want AI benefits before they have AI oversight. The result is a vacuum that companies fill because someone has to make operational decisions.
But sovereignty does not have to mean autarky. It can mean bargaining power, audit rights, procurement discipline, domestic capacity, regional cooperation, open standards, and the ability to switch providers without breaking the state. It can mean knowing which systems are too sensitive to outsource blindly and which can safely be rented.
Malaysia’s own AI ambitions sit inside this tension. Like many countries, it wants the productivity gains and investment halo of AI without becoming a passive consumer of foreign infrastructure. That is a difficult line to walk. Governments can announce AI strategies quickly; they cannot conjure chip supply chains, data center ecosystems, and deep technical talent overnight.
Still, the middle ground matters. Countries can build national evaluation labs, require transparency for public-sector AI systems, support local-language datasets, invest in compute access for universities, and insist that critical services do not depend on a single foreign vendor. None of that guarantees independence. It does create leverage.
For IT administrators, the enterprise version of the same idea is avoiding architectural helplessness. If an organization’s AI strategy depends entirely on one vendor’s roadmap, one model family, one identity stack, and one cloud billing model, then it has made a strategic decision whether or not anyone called it that.
In AI, standards are still forming. That means early movers can influence the definitions. Large companies have the lawyers, engineers, lobbyists, and policy teams to participate in every consultation. Smaller states, civil society groups, independent researchers, and smaller vendors often do not.
The danger is not simply that standards will be too weak. It is that they will encode the assumptions of the largest incumbents. A compliance regime that only hyperscalers can afford may improve documentation while entrenching concentration. A safety framework built around frontier labs may ignore open-source models, local deployments, or public-sector needs. A benchmark that privileges English-language performance may underserve multilingual societies.
This is where global AI governance becomes more than summitry. If developing and middle-income countries are not present when standards are written, they will be present later as markets to be standardized. That is a very different kind of participation.
The Windows ecosystem offers a useful analogy. Microsoft’s dominance has often rested not merely on product quality but on formats, APIs, management tooling, certification regimes, and enterprise defaults. Once an ecosystem becomes the baseline, alternatives must prove not just that they work, but that they fit into the world the incumbent defined.
But convenience is also how infrastructure power becomes normal. Users rarely see the training data disputes, labor conditions, data center siting fights, energy demands, content moderation policies, model-routing decisions, and national security contracts behind the interface. The smoother the product, the easier it is to forget that it is part of a contested system.
This matters because AI adoption is increasingly happening by default. Operating systems, browsers, office suites, customer relationship platforms, developer tools, and security products are adding AI features as built-in capabilities. The decision is no longer “Should we buy an AI system?” It is “Which AI features do we disable, govern, or accept inside systems we already use?”
That is a much harder governance problem. Organizations are better at approving new tools than auditing new capabilities inside existing tools. A feature can appear in an update, a preview channel, a license tier, or an admin console toggle. By the time policy catches up, users may already have built habits around it.
For Windows administrators, this is not theoretical. The age of AI in the enterprise will be managed through identity, endpoint policy, data loss prevention, logging, retention, app control, and vendor contracts. The politics may be global, but the enforcement will often happen in a tenant admin center.
Even if a human remains “in the loop,” the machine’s recommendation can become the default. Staff may defer to it because it appears objective, saves time, or reduces personal responsibility. Managers may value it because it creates metrics. Vendors may defend it because it performed acceptably in aggregate. People harmed by it may struggle to understand how a decision was made.
Anwar’s argument about democracy and participation belongs here. Democratic accountability is not satisfied by telling citizens that an AI system is efficient. Public systems require contestability. People need to know when automated tools are used, what data they rely on, how errors are corrected, and who is responsible when the system fails.
This is harder than publishing a policy document. It requires procurement clauses, audit logs, appeal mechanisms, model evaluation, data governance, staff training, and political willingness to slow down deployments that are not ready. It also requires acknowledging that some AI uses may be inappropriate even if they are technically feasible.
The technology industry prefers the language of risk management because risk can be optimized. Democracy sometimes requires the language of limits.
The better response is pluralism. Not the vague pluralism of panel discussions, but practical pluralism: multiple providers, open interfaces, public-interest research, independent audits, regional compute initiatives, civil society participation, and procurement rules that prevent lock-in. The goal should not be to replace one center of power with another. It should be to prevent any center from becoming unchallengeable.
Open-source AI complicates this picture in useful ways. Open models can reduce dependence on closed vendors, enable local adaptation, and support research. But openness is not magic. Powerful open models still require compute, expertise, and governance. They can also be misused. The question is not whether open or closed AI is morally superior in all cases, but whether the ecosystem preserves meaningful choice.
For enterprises, the same principle applies. A healthy AI strategy may include commercial copilots, private models, open-weight systems, retrieval over internal data, and strict controls for sensitive workflows. The point is not ideological purity. The point is resilience.
The stack is where power lives. Chips determine capacity. Cloud determines access. Data determines relevance. Standards determine legitimacy. Contracts determine permissible use. Energy determines scale. Capital determines who can survive the next training run.
This is why the AI concentration debate will not be solved by asking chatbots to be nicer or more accurate. Model behavior is the visible surface of a deeper industrial structure. If that structure remains concentrated, then improvements in safety may still leave the world dependent on a narrow group of actors.
There is also a feedback loop. The largest AI firms attract more users, which generate more data and revenue, which fund more compute, which improves models, which attracts more users. Cloud partnerships reinforce the loop. Enterprise integrations deepen it. Developer ecosystems make it sticky.
Breaking that loop entirely may be impossible. Bending it toward accountability is the more realistic task.
The answer today is uneven. Shareholders get a vote through capital markets. Governments get a vote through procurement and regulation. Engineers get a vote through design choices. Cloud providers get a vote through infrastructure. Users get convenience, feedback buttons, and sometimes opt-outs. Citizens often get the consequences.
That imbalance is what makes AI power concentration a democratic issue rather than only a market issue. Markets can tolerate concentration for long periods if products are useful and margins are defensible. Democracies should be more suspicious, because concentrated informational power can shape the conditions under which democratic choices are made.
This does not mean every AI company is malicious or every government deployment is dystopian. It means incentives are not enough. Without enforceable accountability, even well-intentioned systems can drift toward surveillance, dependency, opacity, and exclusion.
That means the boring work matters. Contract language matters. Data classification matters. Exit plans matter. Audit rights matter. Energy planning matters. Public procurement rules matter. Model evaluation matters. So does the capacity to say no when a vendor’s roadmap outruns an institution’s ability to govern it.
For Windows-heavy enterprises, the next few years will test whether AI is managed as a strategic layer or merely enabled as a productivity feature. Microsoft, Google, Amazon, OpenAI, Anthropic, and others will continue packaging AI into tools that organizations already use. The pressure to adopt will be intense, because no CIO wants to explain why competitors are automating faster.
But speed is not the same as control. Organizations that move quickly without governance may find themselves unable to answer basic questions: which data entered which model, which outputs influenced which decisions, which logs exist, which vendor terms apply, and what happens if the provider changes the service.
The useful response is not panic. It is discipline.
That distinction matters. The usual debate about artificial intelligence is framed as a race between innovation and regulation, as though society must choose between speed and safety. Anwar’s speech pointed to a more uncomfortable reality: the future of AI may be decided less by what the technology can do than by who owns the machinery that makes it possible.
The AI Race Is Becoming a Geography Lesson in Power
Anwar’s warning landed in Tokyo, but its target was global. The infrastructure of modern AI is not evenly spread across the world; it is concentrated in a few layers of technical dependency. Advanced chips, massive data centers, foundation models, cloud distribution, and platform standards are not abstractions. They are chokepoints.That makes AI different from many earlier waves of software. A country could build websites, mobile apps, or back-office systems with relatively modest capital and commodity hardware. Training and deploying frontier AI, by contrast, depends on scarce accelerators, industrial-scale electricity, high-end networking, elite engineering teams, and cloud contracts that can run into eye-watering sums.
This is why Anwar’s remarks should not be dismissed as another generic call for “ethical AI.” The ethical language is there, but the political economy is the story. If a handful of actors control the compute, the models, the payment rails, and the compliance frameworks, then everyone else is not merely adopting technology. They are renting access to someone else’s civilization stack.
For smaller and middle-power states, that raises the obvious sovereignty problem. Governments want AI for public services, education, health care, policing, cybersecurity, and economic modernization. But if the underlying systems are built, tuned, priced, and governed elsewhere, then national policy becomes downstream of foreign infrastructure.
The New Dependency Is Not Oil, but Compute
The old geopolitical metaphor for strategic dependence was energy. States worried about oil lanes, gas pipelines, refineries, and reserves. AI replaces none of that, but it adds a new dependency: compute.Compute is not simply “more servers.” It is the combined availability of advanced chips, data center capacity, power, cooling, specialized software, and technical labor. In the AI era, compute determines who can train models, who can deploy them at scale, who can experiment cheaply, and who can set the defaults that others must accept.
That is why Anwar’s reference to chips, energy, cloud infrastructure, and capital cuts to the center of the issue. The world does not have an evenly distributed AI market in which every country chooses among neutral tools. It has a hierarchy. At the top are the firms that design the most sought-after accelerators, the companies that manufacture them, the hyperscalers that rent them, and the model builders that turn them into commercial products.
Below that are enterprises and governments that integrate AI into workflows. Further below are users who encounter AI through search engines, office suites, phones, social media, customer service systems, and public-sector portals. Each layer has less bargaining power than the one above it.
The uncomfortable part is that much of this hierarchy is invisible to ordinary users. A chatbot appears as a text box. A copilot appears as a button in an application. A recommendation system appears as convenience. But behind the interface is a supply chain of compute and policy decisions that most users, and many governments, cannot inspect.
Cloud Platforms Have Become the New Public Square’s Landlords
The concentration problem is not limited to model developers. The cloud matters because cloud providers increasingly determine what can be built, where it can run, how it is billed, and what compliance posture is required. Amazon Web Services, Microsoft Azure, and Google Cloud are not merely hosting vendors in this story. They are the distribution layer for modern AI.For WindowsForum readers, this should feel familiar. Microsoft has spent years turning Windows, Microsoft 365, Azure, GitHub, and security tooling into a connected platform rather than a collection of products. That strategy has advantages for administrators: centralized identity, policy management, telemetry, endpoint security, and increasingly AI-assisted operations. But it also means that AI adoption often arrives through bundled enterprise ecosystems rather than discrete, easily separable choices.
The same dynamic applies beyond Microsoft. A company that builds on a hyperscaler may gain world-class infrastructure overnight, but it also inherits that provider’s pricing, regional availability, security model, acceptable-use policies, and strategic priorities. If the provider changes terms, sunsets a model, restricts a capability, or shifts pricing, customers can find themselves with limited leverage.
This is not a conspiracy; it is platform economics. The more valuable a platform becomes, the more customers build around it. The more customers build around it, the harder it becomes to leave. AI intensifies that lock-in because models are not interchangeable in the way virtual machines once appeared to be. Prompts, embeddings, fine-tuning, retrieval systems, evaluation pipelines, security reviews, and user workflows all create friction.
When Anwar warns that powerful actors may decide “what civilisation entails,” the phrase can sound grandiose. But in practical IT terms, it means something precise: platform owners set defaults, and defaults become behavior.
The Military-AI Dispute Shows Where the Soft Language Ends
The most revealing AI governance fights are not happening in consumer chatbots. They are emerging where commercial AI meets the state: defense, intelligence, policing, border systems, and cyber operations. The reported clash between Anthropic and the Pentagon over Claude safeguards is a case study in how quickly abstract principles become institutional conflict.According to multiple reports, Anthropic resisted demands to remove restrictions that would have allowed its models to be used for mass domestic surveillance or fully autonomous weapons. The company’s stated position was that frontier AI systems are not reliable enough for such uses, and that mass surveillance poses grave democratic risks. Whether one views that stance as principled, strategic, self-protective, or all three, it exposes the governance problem.
Who decides what a powerful model may be used for? The vendor that built it? The government that contracts for it? The military command that sees operational advantage? The public that bears the consequences? The courts, after the fact?
That is the unresolved issue Anwar was pointing toward. Democracies are built on accountability, but AI systems are often procured through contracts, embedded in classified settings, and governed by terms of service few citizens will ever read. The result is a democratic gap: public power exercised through private systems under opaque rules.
It is tempting to say governments should simply regulate the vendors. But the state is not always the neutral referee. In national security contexts, it may also be the customer demanding more capability and fewer constraints. That turns AI governance into a three-way contest among public accountability, commercial control, and state power.
Silicon Valley’s Moral Vocabulary Cannot Carry the Whole Burden
The AI industry has become fluent in the language of safety. Companies publish model cards, system cards, red-team summaries, usage policies, safety frameworks, responsible deployment principles, and alignment research. Some of this work is serious. Some of it is defensive branding. Much of it is both.The problem is that voluntary safety commitments do not resolve the legitimacy question. A private company can refuse a use case today and change its mind tomorrow. A board can be replaced. A revenue crisis can sharpen priorities. A government can threaten procurement consequences. A competitor can offer fewer restrictions and win the contract.
This is why Anwar’s emphasis on public institutions is important. AI ethics cannot survive as corporate noblesse oblige. If core democratic decisions are delegated to private AI providers, then the public is left hoping that founders, executives, and investors maintain the right instincts under pressure.
That is not governance. It is dependency with better press releases.
The challenge, however, is that public institutions are not automatically ready to do better. Many regulators lack technical capacity. Many procurement departments are outmatched by vendors. Many legislatures move slowly. Many agencies want AI benefits before they have AI oversight. The result is a vacuum that companies fill because someone has to make operational decisions.
Digital Sovereignty Is Easy to Say and Expensive to Build
“Digital sovereignty” has become a fashionable phrase because it promises control without always specifying cost. For countries outside the AI superpower tier, sovereignty cannot mean building everything domestically from chips to models to cloud regions. That would be unrealistic for most states and wasteful even for some wealthy ones.But sovereignty does not have to mean autarky. It can mean bargaining power, audit rights, procurement discipline, domestic capacity, regional cooperation, open standards, and the ability to switch providers without breaking the state. It can mean knowing which systems are too sensitive to outsource blindly and which can safely be rented.
Malaysia’s own AI ambitions sit inside this tension. Like many countries, it wants the productivity gains and investment halo of AI without becoming a passive consumer of foreign infrastructure. That is a difficult line to walk. Governments can announce AI strategies quickly; they cannot conjure chip supply chains, data center ecosystems, and deep technical talent overnight.
Still, the middle ground matters. Countries can build national evaluation labs, require transparency for public-sector AI systems, support local-language datasets, invest in compute access for universities, and insist that critical services do not depend on a single foreign vendor. None of that guarantees independence. It does create leverage.
For IT administrators, the enterprise version of the same idea is avoiding architectural helplessness. If an organization’s AI strategy depends entirely on one vendor’s roadmap, one model family, one identity stack, and one cloud billing model, then it has made a strategic decision whether or not anyone called it that.
The Standards Fight Is the Quiet Fight
Anwar mentioned standards, and that may be the least flashy but most consequential part of the debate. Standards determine what counts as safe, interoperable, compliant, auditable, and acceptable. They shape procurement checklists, insurance requirements, vendor claims, and regulatory enforcement.In AI, standards are still forming. That means early movers can influence the definitions. Large companies have the lawyers, engineers, lobbyists, and policy teams to participate in every consultation. Smaller states, civil society groups, independent researchers, and smaller vendors often do not.
The danger is not simply that standards will be too weak. It is that they will encode the assumptions of the largest incumbents. A compliance regime that only hyperscalers can afford may improve documentation while entrenching concentration. A safety framework built around frontier labs may ignore open-source models, local deployments, or public-sector needs. A benchmark that privileges English-language performance may underserve multilingual societies.
This is where global AI governance becomes more than summitry. If developing and middle-income countries are not present when standards are written, they will be present later as markets to be standardized. That is a very different kind of participation.
The Windows ecosystem offers a useful analogy. Microsoft’s dominance has often rested not merely on product quality but on formats, APIs, management tooling, certification regimes, and enterprise defaults. Once an ecosystem becomes the baseline, alternatives must prove not just that they work, but that they fit into the world the incumbent defined.
The User Experience Hides the Political Economy
AI arrives to most people as convenience. A document summarizes itself. A search result becomes conversational. A coding assistant suggests a fix. A meeting transcript produces action items. A helpdesk bot answers at midnight. These are real benefits, and dismissing them would be foolish.But convenience is also how infrastructure power becomes normal. Users rarely see the training data disputes, labor conditions, data center siting fights, energy demands, content moderation policies, model-routing decisions, and national security contracts behind the interface. The smoother the product, the easier it is to forget that it is part of a contested system.
This matters because AI adoption is increasingly happening by default. Operating systems, browsers, office suites, customer relationship platforms, developer tools, and security products are adding AI features as built-in capabilities. The decision is no longer “Should we buy an AI system?” It is “Which AI features do we disable, govern, or accept inside systems we already use?”
That is a much harder governance problem. Organizations are better at approving new tools than auditing new capabilities inside existing tools. A feature can appear in an update, a preview channel, a license tier, or an admin console toggle. By the time policy catches up, users may already have built habits around it.
For Windows administrators, this is not theoretical. The age of AI in the enterprise will be managed through identity, endpoint policy, data loss prevention, logging, retention, app control, and vendor contracts. The politics may be global, but the enforcement will often happen in a tenant admin center.
Accountability Cannot Be Retrofitted After Deployment
The most dangerous phrase in technology policy is “we’ll fix it later.” It is especially dangerous with AI because deployed systems create institutional habits. Once a government agency uses AI to triage benefits, flag risks, summarize intelligence, or prioritize enforcement, the system becomes part of bureaucratic reality.Even if a human remains “in the loop,” the machine’s recommendation can become the default. Staff may defer to it because it appears objective, saves time, or reduces personal responsibility. Managers may value it because it creates metrics. Vendors may defend it because it performed acceptably in aggregate. People harmed by it may struggle to understand how a decision was made.
Anwar’s argument about democracy and participation belongs here. Democratic accountability is not satisfied by telling citizens that an AI system is efficient. Public systems require contestability. People need to know when automated tools are used, what data they rely on, how errors are corrected, and who is responsible when the system fails.
This is harder than publishing a policy document. It requires procurement clauses, audit logs, appeal mechanisms, model evaluation, data governance, staff training, and political willingness to slow down deployments that are not ready. It also requires acknowledging that some AI uses may be inappropriate even if they are technically feasible.
The technology industry prefers the language of risk management because risk can be optimized. Democracy sometimes requires the language of limits.
AI Nationalism Will Not Save the Public Interest
One possible response to concentration is AI nationalism: every country races to build its own models, clouds, datasets, and champions. That may sound attractive, but it can reproduce the same problem at a national scale. A domestic monopoly is still a monopoly. A state-controlled model can be less accountable than a foreign one if institutions are weak.The better response is pluralism. Not the vague pluralism of panel discussions, but practical pluralism: multiple providers, open interfaces, public-interest research, independent audits, regional compute initiatives, civil society participation, and procurement rules that prevent lock-in. The goal should not be to replace one center of power with another. It should be to prevent any center from becoming unchallengeable.
Open-source AI complicates this picture in useful ways. Open models can reduce dependence on closed vendors, enable local adaptation, and support research. But openness is not magic. Powerful open models still require compute, expertise, and governance. They can also be misused. The question is not whether open or closed AI is morally superior in all cases, but whether the ecosystem preserves meaningful choice.
For enterprises, the same principle applies. A healthy AI strategy may include commercial copilots, private models, open-weight systems, retrieval over internal data, and strict controls for sensitive workflows. The point is not ideological purity. The point is resilience.
The AI Debate Has Finally Caught Up With the Supply Chain
For years, AI policy focused heavily on model behavior: bias, hallucination, misinformation, copyright, and safety. Those issues remain important, but Anwar’s speech reflects a broader shift. The debate is moving down the stack.The stack is where power lives. Chips determine capacity. Cloud determines access. Data determines relevance. Standards determine legitimacy. Contracts determine permissible use. Energy determines scale. Capital determines who can survive the next training run.
This is why the AI concentration debate will not be solved by asking chatbots to be nicer or more accurate. Model behavior is the visible surface of a deeper industrial structure. If that structure remains concentrated, then improvements in safety may still leave the world dependent on a narrow group of actors.
There is also a feedback loop. The largest AI firms attract more users, which generate more data and revenue, which fund more compute, which improves models, which attracts more users. Cloud partnerships reinforce the loop. Enterprise integrations deepen it. Developer ecosystems make it sticky.
Breaking that loop entirely may be impossible. Bending it toward accountability is the more realistic task.
The Tokyo Warning Was Really About Who Gets a Vote
Anwar framed AI as a human question, and that is easy to mock because political leaders often reach for lofty language around technology. But in this case the lofty language is doing real work. “Humanity” is a way of asking who gets to participate in decisions that increasingly shape labor, security, education, administration, and war.The answer today is uneven. Shareholders get a vote through capital markets. Governments get a vote through procurement and regulation. Engineers get a vote through design choices. Cloud providers get a vote through infrastructure. Users get convenience, feedback buttons, and sometimes opt-outs. Citizens often get the consequences.
That imbalance is what makes AI power concentration a democratic issue rather than only a market issue. Markets can tolerate concentration for long periods if products are useful and margins are defensible. Democracies should be more suspicious, because concentrated informational power can shape the conditions under which democratic choices are made.
This does not mean every AI company is malicious or every government deployment is dystopian. It means incentives are not enough. Without enforceable accountability, even well-intentioned systems can drift toward surveillance, dependency, opacity, and exclusion.
The Practical Lesson Is to Build Leverage Before the Lock-In Hardens
The most concrete implication of Anwar’s warning is that users, companies, and states should stop treating AI adoption as a simple upgrade cycle. AI is infrastructure. Infrastructure choices compound.That means the boring work matters. Contract language matters. Data classification matters. Exit plans matter. Audit rights matter. Energy planning matters. Public procurement rules matter. Model evaluation matters. So does the capacity to say no when a vendor’s roadmap outruns an institution’s ability to govern it.
For Windows-heavy enterprises, the next few years will test whether AI is managed as a strategic layer or merely enabled as a productivity feature. Microsoft, Google, Amazon, OpenAI, Anthropic, and others will continue packaging AI into tools that organizations already use. The pressure to adopt will be intense, because no CIO wants to explain why competitors are automating faster.
But speed is not the same as control. Organizations that move quickly without governance may find themselves unable to answer basic questions: which data entered which model, which outputs influenced which decisions, which logs exist, which vendor terms apply, and what happens if the provider changes the service.
The Compute Oligarchy Leaves Administrators With Homework
The story Anwar told in diplomatic language becomes very operational when it reaches the IT department. The global concentration of AI power cannot be fixed by a sysadmin, but its risks will be experienced through systems that administrators are expected to secure, document, and explain.The useful response is not panic. It is discipline.
- Organizations should inventory AI features already present in their operating systems, productivity suites, developer tools, security platforms, and cloud services.
- Administrators should treat AI outputs that touch business decisions, security operations, or regulated data as governed workflows rather than casual productivity aids.
- Procurement teams should ask vendors how data is retained, where inference runs, which subcontractors are involved, and whether customers can opt out of model training or secondary use.
- Public-sector buyers should require appeal paths, auditability, and human accountability before deploying AI in services that affect rights, benefits, policing, or access to government.
- Enterprises should avoid designing AI architectures that cannot be migrated, inspected, or partially replaced without breaking core operations.
- Policymakers should focus not only on model behavior but also on compute access, cloud concentration, standards-setting, and public-interest technical capacity.