This page could not be loaded. Check your connection, then reload the page.
AI Turns Sovereignty Into a Purchase Decision
AI turns the readiness question into an urgent purchase. AI is where sovereignty becomes an immediate operating decision. Organizations must decide which models they can use, where sensitive data is processed, where inference and training run, who controls the underlying infrastructure, and what happens if a model, cloud, or hardware supplier becomes unavailable.
51.9% of organizations call sovereign AI capability critical to their AI strategy. Almost as many, 49.7%, call independence from a single AI accelerator vendor critical. That is an aspiration more than an immediately available choice. NVIDIA remains the dominant supplier of AI data center accelerators and its CUDA software ecosystem makes moving to alternatives more difficult than simply replacing one chip with another. The finding does not mean buyers can substitute suppliers tomorrow. It means they recognize the exposure and want the ability to reduce it over time.
For AI, the operational question is straightforward: can the organization continue to run a critical AI service if a model provider, cloud provider, accelerator supplier, or critical software dependency changes terms or becomes unavailable? AI commitments are made faster than the platform, data governance, and infrastructure choices needed to support them. Without those foundations, organizations recreate the cloud sovereignty readiness gap in AI, only sooner.
Half call sovereign AI critical, and half say the same of chip independence All respondents
| Category | Sovereign-AI capability | Independence from a single AI accelerator vendor |
|---|---|---|
| Not important | 8% | 9% |
| Moderately important | 40% | 41% |
| Critical | 52% | 50% |
Source: n=1,940 respondents, normalized by region
Respondents that call a sovereign-AI capability critical: 51.9%
Respondents that call independence from a single AI accelerator vendor critical: 49.7%
Organizations that are bringing AI training or inference back to private or local infrastructure: 30%
Sovereign AI is critical for 58% of organizations in North America and 41% in EMEA.
Buyers choose models by workload and place sensitive AI where control matters
Buyers choose AI models by workload, not by type. Where they commit, open-weight models, meaning models with published parameters an organization can download and run itself, outrank proprietary frontier models reached through a vendor's interface. Preferred environments for sensitive AI follow the same logic: a private cloud platform leads, then a public hyperscaler limited to a sovereign region, a jurisdiction that keeps the workload in place, with an unrestricted public hyperscaler last. Most organizations choose the strongest frontier model where capability decides the outcome, and run open-weight models where control decides it, so a technology decision-maker needs a commercially supported platform that moves workloads across public cloud, sovereign region, private cloud, and on-premises.
Buyers choose models by workload and place sensitive AI where control matters All respondents · Open-weight models: 26%; proprietary frontier models: 18%
| Category | All respondents |
|---|---|
| Mix depending on workload | 29% |
| Open-weight/open-source models we control | 26% |
| Locally trained or in-region sovereign models | 24% |
| Proprietary frontier models | 18% |
| No preference | 4% |
Source: n=1,940 respondents, normalized by region
Buyers choose models by workload and place sensitive AI where control matters: preferred environment for sensitive AI workloads All respondents
| Category | All respondents |
|---|---|
| Private cloud platform | 20% |
| Public hyperscaler — in-country/sovereign region only | 17% |
| On-premises infrastructure | 16% |
| Local or regional sovereign cloud provider | 15% |
| Hybrid mix of multiple environments | 14% |
| Hyperscaler-managed private/dedicated infrastructure* | 11% |
| Public hyperscaler (any region) | 8% |
Source: n=1,940 respondents, normalized by region
As governments and government-backed national AI companies fund models trained locally and in-region, open-source AI models will appear from many more countries, widening the set an organization can run under its own control. And as other classes of models catch up to present-day frontier capabilities, despite the lag, leaders may find themselves much more able to embrace open or local models even for workloads requiring advanced, dependable capabilities, opening the door to more prioritization of sovereign, control, or open-source options.
AI workload movement is part of the near-term sovereignty agenda: 30% of organizations say they are actively repatriating or migrating AI workloads to private or local trusted infrastructure within the next 24 months, rising to 41% in Asia-Pacific against 26% elsewhere. Agentic AI, AI that acts and decides on its own instead of only answering a prompt, raises the stakes. As AI systems move from answering prompts to taking actions in regulated processes, organizations need clearer controls over where inference runs, what data and tools the system can access, how decisions are observed, and how the system can be stopped or moved if a dependency changes. A leader planning agentic AI for a regulated process should decide the required level of control over inference, data access, tools, and observability before choosing a model.
AI Training or Inference Repatriation, by Region All respondents
| Category | Repatriating AI training or inference to private/local infrastructure |
|---|---|
| North America | 28% |
| EMEA | 27% |
| APAC | 41% |
| LATAM | 23% |
| All respondents | 30% |
Source: n=289 / 470 / 979 / 202 / 1,940 respondents
Practitioners still pick the best model, for now
What companies say they prefer and what they actually use are often different. While buyers say they favor models they control, convenience and performance usually win out in practice. For control to matter, leaders must make it a firm requirement for sensitive workloads rather than just a general preference. The actual decisions vary substantially by the organization's needs; region may matter little, but industry may make a significant difference. In the case of these two European executives, the retailer optimizes for capability while the public agency accountable for citizen data focuses on control:
We don't factor sovereignty into AI. We use [a leading US AI assistant] and [another leading US AI assistant] because they're better than [a French AI vendor]. We want the best model. — CIO, enterprise retailer in EMEA
We'll run our own LLMs in our own environment, get them from a vendor but own the entire IPR, control it, and make sure it doesn't talk back to a mothership elsewhere. — CDIO, large public agency in EMEA
A technology decision-maker should know whether their own organization has already made that choice. Organizations that hold to a sovereign-AI policy under pressure are the ones that chose a controlled environment for their sensitive workloads before the choice became urgent.
What to Do About It
Decide the required level of control over inference, data access and tooling before you choose a model, not after.