THE AI AGENDA
For countries and companies, AI sovereignty means keeping enough control over the AI they depend on without cutting themselves off from the global frontier
The dictionary definition of sovereignty is freedom from external control. Such freedom is an illusion when it comes to AI. Full ownership of the stack — from energy and chip production, through frontier models and the talent behind them, to proprietary data and software — is out of reach for any single country or company. Even the leading economies cannot make an advanced AI chip alone: The required tools, factories, and refined minerals are divided among a few countries and suppliers.
The realistic goal is controlled interdependence: having enough control over AI capabilities and dependencies to keep critical work running and governable, and a credible way to replace any of those dependencies.
Countries and companies share that goal. Countries are focused on maximizing economic competitiveness, security, freedom of action, and the resilience of critical services. Companies need continuity of operations, predictable costs, and control of the business. Some disruption will be deliberate — suppliers changing terms, states pursuing leverage, criminals seeking gain — and designed to defeat the backups.
In practice, AI sovereignty means choosing the terms of dependence: knowing what to rely on, how hard it would be to replace, and how to react if a critical dependency were compromised.
55%
expect at least moderate disruption if their company’s main AI provider becomes unavailable
65%
say their company has not put the ability to switch AI providers in place
Only0.1%
of customers fully switch cloud providers each year
Up to60
engineer-days needed to switch primary foundation models
46%
of CEOs say supplier and model concentration is formally registered as a board or enterprise risk
10%
of technology leaders have a signed statement defining what data and assets they retain if an AI supplier raises prices or cuts off access
74%
of CEOs expect a lasting advantage from redesigned processes and operating models; 64% from proprietary data and customer context
4%
say the same of privileged access to a leading AI model or supplier
41%
of CEOs rank geopolitics, trade policy, and industrial policy as a top threat — second only to macroeconomic conditions
32%
cite cybersecurity, data protection, and information integrity, including attacks on critical infrastructure, as a top threat
17%
of the world’s 100 largest economies own or host a hard-to-replace part of the AI supply chain
7
countries own a true chokepoint; another 10 own a narrower critical input or host hard-to-replace capacity
Source: Oliver Wyman Forum Global Voices Survey, August 2026 (footnote 1); UK Competition and Markets Authority (footnote 2); Oliver Wyman 2026 Parallel Surveys of CEOs and Tech Leaders on Agentic AI and IT (footnotes 2, 3, and 4); Oliver Wyman Forum x NYSE CEO Survey 2026 (footnote 5); Oliver Wyman Forum Sovereign AI Index, Q3 2026 snapshot (footnote 6)
REALITY CHECK 1
For a company or a country, relying on one model, one cloud, or one provider is often the cheapest and simplest starting point. But as an organization builds its workflows around that provider, switching becomes harder and more expensive. Protecting the ability to change providers is one of the costs of AI sovereignty, and that cost rises the longer a single provider stays embedded.
In cloud computing, the option to switch is rarely exercised; the UK competition authority recently found that only about one in a thousand cloud customers switched providers in a typical year. Companies are applying the lesson from cloud at the model layer, where swapping models still presents challenges. But dependency is forming around the model: in the agent memory, project context, orchestration software, and automation that accumulate as critical workflows are built, often across several suppliers. Once that constellation of providers is lined up, it is hard to reshuffle. Half of technology leaders estimate that switching just the primary foundation model on their main use case takes 21 to 60 engineer-days; the agent memory, context, and orchestration built around it add to that bill. Platforms that route work across several providers’ models hedge only the layer that was already easy to switch.
Dependency also can be shaped by commercial arrangements. Discounts tied to multiyear spending commitments, licensing terms, and bundled services can make an alternative uneconomic even when switching is technically possible. The decision to exit boils down to two questions: Can the organization make a move, and can it afford to move?
Regulators reached the same conclusion about cloud services, treating provider concentration as a resilience risk and requiring firms to maintain tested exit plans for critical providers — but only after nearly two decades of accumulated dependence. AI-enabled systems are being embedded into core operations on a much shorter timeline, so firms cannot assume supervision will arrive first.
21 to 60 engineer-days needed to switch the primary foundation model
Preparation lags awareness of the risk. Forty-six percent of CEOs say supplier and model concentration is formally registered as a board or enterprise risk, and 55% of companies expect at least moderate disruption if their main AI provider becomes unavailable. Yet only 35% have set up the ability to switch providers, and just 10% of technology leaders have a written statement, signed by their executive sponsor, of which data and assets stay under their control if a key supplier raised prices or cut off access. These are self-assessments; in a real shock, fallback plans could fare worse than expected. For a critical dependency, a fallback plan that has not been funded and tested is not a reliable option.
Measures in place to reduce dependence on any single AI provider reported by respondents with visibility into company AI spending
Note: Respondents could select multiple measures.
Source: Oliver Wyman Forum Global Voices Survey (August 2026)
Survey question: “Which of the following has your company put in place to reduce its dependence on any single AI provider?”
Base: N=5,789 (respondents with visibility into company AI spending)
REALITY CHECK 2
Not every AI service or workflow needs the same protections against disruption. A consumer-facing app can tolerate delay or substitution, but interruptions of AI in health records, energy systems, and payments can cause wider harm. One standard for every service could waste money where the stakes are low while leaving critical services vulnerable.
The cost of sovereignty, whether it means paying more or accepting less capable AI, should rise with the potential harm and fall for more replaceable dependencies. Asked which services should have the most resilient AI access, people in 12 of 15 markets put healthcare first, ahead of energy, national security, and payments. Public pressure will favor the services people see, but the sovereignty premium should also follow the chain of dependence down to the energy and critical infrastructure those services run on.
Services that respondents say they want their country to maintain if there’s a prolonged loss of access to foreign AI providers, ranked
Note: Respondents could select up to three services.
Source: Oliver Wyman Forum Global Voices Survey (August 2026)
Survey question: “Which AI-enabled services should your country be able to keep operating during a prolonged loss of access to foreign AI providers?”
Base: N=12,837 (all respondents)
Prioritization is only the first step. Any extra cost paid for greater control and resilience should buy a specific protection against a defined failure, regardless of the cause. Local hosting addresses jurisdiction and access risk, while redundant capacity, portable data, contractual rights, and trusted suppliers address different failures. None is interchangeable, and even several together can leave a shared dependency hidden.
The sovereignty premium should target the environment around the model: the computing power, energy, and connectivity an alternative needs to run, and the data architecture, identity, evaluation, tools, and workflows that put the AI to work. Whether the model, software, and integrations come from one supplier or several, an exit starts to resemble a rebuild rather than a switch.
Where substitution is slow or costly, firms should have plans for implementing a smaller model, narrower functionality, or more human handling to keep essential functions running until a full solution is in place.
Note: Purple shading marks dependencies that are slow or costly to replace.
Source: Oliver Wyman Forum analysis
REALITY CHECK 3
Leaders do not expect advantage to come from the AI model itself: Only 4% of CEOs say privileged access to a leading model or supplier will be a lasting advantage, while 74% point to redesigned processes and operating models and 64% to proprietary data and customer context. AI sovereignty follows the same logic. Critical work needs enough capability under the company’s control to keep running, and that is often possible with less than frontier performance. For most workflows, reliable access, organizational context, security, and integration matter more. Frontier capability is worth its added dependency only where the performance gap changes crucial outcomes.
Open-weight models have trailed the frontier by a matter of months, and the computing power needed for a given level of performance has halved roughly every eight months. At that pace, today’s frontier capability becomes cheap and controllable within a year. And unless the demands of routine work rise as fast as capability does, tomorrow’s frontier models will be overkill for most of that work.
Among respondents with access to custom or in-house AI
Source: Oliver Wyman Forum Global Voices Survey (August 2026)
Survey questions: “If you could only keep one for your work, which would you choose?”; “How often do you manually move information or work between a public AI tool and company AI tool because each is better at a different part of the task?”
Base: N=2,278 (respondents with access to custom or in-house AI); N=1,268 (respondents who would keep company AI)
Inside companies, controlled and frontier AI are already complements rather than substitutes. Among respondents with access to custom or in-house AI, 56% would choose the company tool if they could keep only one, yet 65% of that group still move tasks between company and public AI at least weekly. Company tools, which trade capability for control, win on internal context and data protection. Public tools, which offer the frontier directly, win on speed, quality, and breadth.
Organizations can maximize their AI sovereignty by owning the control layer while keeping the frontier replaceable. The control layer is the part of the model’s environment that accumulates through use: the data, context, and evaluation that critical work depends on, as well as the memory, permissions, tool connections, and workflow state that AI agents build up inside orchestration software. By keeping those portable, a company can preserve the ability to plug in better models as they emerge. Frontier access then becomes more of a managed option than a dependency.
Control also can mean proximity. Where critical work must continue without a reliable connection to centralized infrastructure — say, in a vehicle, facility, clinic, or deployed unit — smaller models running at the edge preserve essential functions.
Control is also cultural. A model shapes what users see. One trained largely outside a country’s language, institutions, and legal tradition carries blind spots and biases. Companies inherit them. Anchoring models in users’ context is a sovereignty premium worth paying.
Only 4% of CEOs say privileged access to a leading model or AI supplier will be a lasting advantage
REALITY CHECK 4
Regulation and procurement redirect demand, investment, and corporate behavior at a scale that public budgets rarely reach. Procurement rules create anchor customers. Where national security is involved, government is often the dominant buyer and can define the market outright. With hundreds of billions of dollars flowing into AI globally each year, the most effective sovereign AI investment may be redirecting private capital through the levers of government rather than spending public funds.
That leverage runs in both directions. A well-calibrated government requirement gives more suppliers enough demand to reach scale, and each additional supplier increases optionality for the private sector. But a requirement set at a level only the largest providers can meet entrenches the supplier concentration it was meant to reduce. Localization rules split the same way: They attract real capacity where the economics work, but where they don’t, such rules simply add cost and a local address to the same foreign stack. Government policies should be judged by whether they increase the number of providers a country or company can actually switch to.
Some constraints still call for direct investment or partnership. Regulation alone rarely produces energy, computing power, or specialist talent at the speed required. Public investment can accelerate all three. But government investment in national AI capability faces political and fiscal constraints. Across 15 markets, three-quarters of residents support major public spending to build AI capability, yet only 19% support it if it requires higher taxes or spending cuts elsewhere. This constraint is tightest in Europe and loosest in Asia and the Gulf. AI may also squeeze the budgets that would fund it: If it displaces taxable work and adds public spending on retraining and income support, governments will have less to spend. Few governments will find the fiscal room or political leeway to fund national AI capability at the scale their ambitions imply. That leaves most of the work to the instruments that do not require public money.
Conditional and unconditional support for AI investment, by market
Note: Conditional = support only if investment does not raise taxes or reduce other spending; unconditional = support even if taxes/cuts are required. Conditional support share is calculated from unrounded respondent counts as conditional support divided by total support. Markets are ranked by this normalized measure.
Source: Oliver Wyman Forum Global Voices Survey (August 2026)
Survey question: “Would you support your government making a major public investment to build national AI capabilities?”
Base: N=12,837 (all respondents); market Ns=503–936
REALITY CHECK 5
No country holds every capability that matters for sovereign AI. Few have the scale on their own to influence the terms of global suppliers. This makes alliances and partnerships crucial. Our Sovereign AI Index makes the constraint concrete: Hard-to-replace positions in the AI supply chain sit in just 17 of the world’s 100 largest economies, and only seven hold a true chokepoint. For many countries and companies, sovereignty means credible access to more than one technology ecosystem and the freedom to combine elements of each. Partnerships can add capacity, bargaining power, and routes around a disrupted relationship.
Critical capabilities for advanced AI chips are concentrated in different economies, creating mutual dependencies
Note: Percentages are estimated shares only; the fabrication range reflects differing definitions of the leading edge.
Source: Oliver Wyman Forum analysis
A partnership strengthens sovereignty only when it widens the choices available. It should add a supplier, an architecture, a route to capacity, or expertise that the partners lacked on their own. Many arrangements fail that test on technical grounds. Unless systems interoperate so that data, identity, and tools move between them, a second provider adds cost without adding choice. Companies build choice through interoperable architectures, joint procurement, and shared evaluation; countries do so through reciprocal access, shared capacity, and research and procurement commitments. Where scarcity is structural, as with critical minerals, semiconductors, energy, and specialist talent, the arrangement has to be durable because replacement takes years.
Supplier diversity is not dependency diversity. Two separate providers may still rely on the same chipmaker, power grid, undersea cable, or export-control regime. A partnership that leaves everyone dependent on the same underlying provider or stack doesn’t strengthen anyone’s sovereignty.
Strengths in models, research, compute, supply chain chokepoints, private capital, and government readiness are distributed across economies, and even the economy that leads on most counts, including chokepoints, cannot make an advanced AI chip on its own. That is what makes interdependence unavoidable.
Note: Profiles show selected dimension ranks only and are not an overall ranking. Ranks show position, not distance; small differences in rank should be interpreted cautiously.
Source: Oliver Wyman Forum Sovereign AI Index. Data snapshot Q3 2026
Significant AI models developed by organizations associated with each economy
88% of large-scale AI models captured by the index are attributable to the United States and China
Published AI research authors affiliated with institutions in each economy
69% of the indexed AI research workforce sits in these ten economies
Operating AI hardware the index could verify as physically located in each economy, regardless of ownership
91% of the verified AI compute captured by the index is hosted in the United States
Ownership or domestic hosting of technologies and capacity judged hard to replace in the AI supply chain
17 economies own or host an in-scope chokepoint position; seven own a true chokepoint
Private investment tagged to AI and assigned to each economy in the 2025 source data
78% of the recorded 2025 AI-related private capital went to the United States
Institutional capacity to adopt, govern, and support AI, as measured by the Oxford Insights Government AI Readiness Index
Note: China has been intentionally excluded from ranking due to insufficient coverage in the source material
Note: Government AI funding is omitted from this publication view because the current evidence does not cover several material markets, including China and the United Arab Emirates, and therefore does not support a credible comparative ranking. No replacement composite ranking has been calculated. Verified onshore AI compute reflects documented operating hardware and may undercount capacity in markets with limited public disclosure. AI research workforce counts published authors and does not represent the full engineering or practitioner workforce. Recorded private AI capital reflects one year of deal data and may be influenced by a small number of large transactions. Callout shares describe evidence captured by the index, not complete global totals. The index does not directly score energy, availability, data control, language or cultural context, orchestration, applications, enterprise adoption, or operating security. Small differences in rank should be interpreted cautiously.
Source: Oliver Wyman Forum Sovereign AI Index. Data snapshot Q3 2026. The index covers the world’s 100 largest economies
SETTING THE TERMS
The Sovereign AI Index settles the question of whether anyone can go it alone. The United States hosts 91% of the verified AI computing power in the index and leads five of six dimensions, yet it cannot make an advanced chip without Dutch lithography, Japanese materials, Korean memory, Chinese critical minerals, and Taiwanese fabrication. The leader depends on others as surely as others depend on the leader. And the economies with the most leverage relative to their size did not buy it; each became hard to replace in one category.
That is the lesson the five reality checks share. Every decision on sovereignty comes down to one test: Does it add a credible alternative to keep critical work running? The test cuts both ways. The assets CEOs expect to deliver lasting advantage — proprietary data and processes, not privileged access to a model — are also what make an organization hard to replace, and therefore worth partnering with.
No country or company is a sovereign island, and none needs to be. The question is how to set the terms: what to own, what to secure through partners, what to keep replaceable, and what cannot be allowed to fail.
Every decision on sovereignty comes down to one test: Does it add a credible alternative to keep critical work running?
Abhishek Sharma is a Partner in Oliver Wyman's IMEA Government and Public Institutions practice
Darren Goldie is a Partner in Oliver Wyman's Government and Public Institutions practice
Jad Haddad is a Partner and Global Head of Quotient, AI by Oliver Wyman, overseeing AI capabilities and client offerings
Jose Lopez is a Senior Fellow at the Oliver Wyman Forum and Principal in Oliver Wyman's Government and Public Institutions and Digital practices
Mario Rizk is the IMEA partnerships lead for Quotient, AI by Oliver Wyman, and a Partner in the Digital and AI practice
Matthieu De Clercq is a Partner and lead of the IMEA Government and Public Institutions practice
Xavier Boileau is a Partner in Oliver Wyman's Digital and Financial Services EMEA practices
