Artificial intelligence is entering a phase where the question is no longer simply what an AI system can do. Increasingly, organizations want to know where it runs, who controls the infrastructure behind it, where the data goes, and whether the system can keep working when an outside provider changes its terms or becomes unavailable.
That shift is giving sovereign AI a much larger role in technology planning. What began largely as a discussion around data residency and regulation is becoming a much broader question of physical and operational control.
Gartner’s September 2026 research, Emerging Tech: Top-Funded Startups in Sovereign AI and Trust Infrastructure, examines close to 30 startups that received venture capital, institutional, asset finance, or sovereign-backed funding between July 2021 and June 2026. Together, the companies represent $13.5 billion across 25 distinct deals.
The research divides the market into four areas: compute and physical infrastructure, deterministic governance and trusted execution, multicloud and decentralized orchestration, and sovereign models and regional languages. Each addresses a different weakness in the way AI is currently built and deployed.
The AI Race Is Moving from Software Screens to Power Grids
The physical side of AI is becoming harder to ignore. Advanced models require enormous computing capacity, and that computing capacity requires dependable electricity. For countries and highly regulated organizations, relying entirely on infrastructure outside their jurisdiction can create problems that go well beyond technology procurement.
Gartner’s research describes sovereign AI as a physical asset rather than simply a software service, the implication is significant. If organizations want operational control, they need to think about the location of GPU clusters, data centers, energy supplies, and the networks connecting them.
The compute and physical infrastructure category is the largest part of the investment tracked in the research, accounting for more than half of the total capital. Gartner identifies HUMAIN, Media Stream AI, Nuvena.ai, Rapidus, Sharon AI, and Stargate Norway among the companies in this segment.
Their approaches vary. Some are developing high-density GPU facilities, while others are working on semiconductor manufacturing or domestic data center infrastructure. Energy is becoming a particularly important part of the equation. AI facilities can be located alongside hydropower, geothermal resources, or other dependable sources of electricity. The research also points to behind-the-meter nuclear small modular reactors as another possible route for securing power.
This is changing the traditional data center model. Instead of simply finding a suitable location for servers, companies are increasingly considering where computing and energy can work together.
Local Computing Is Becoming a Strategic Asset for Regulated Markets
Large, traditional data centers can take years to plan and construct. Sovereign AI needs are creating demand for faster alternatives.
Modular edge data centers, for example, can place high-density computing closer to local energy sources or within restricted enterprise environments.
Telecommunications companies can also become infrastructure partners by using their existing fiber networks, edge locations, and regulatory relationships to host domestic AI workloads.
The approach is particularly relevant to sensitive applications. Defense systems, core financial transactions, and public administration workloads may require physically isolated infrastructure or sovereign cloud regions, while less sensitive applications can continue to use global public clouds.
Gartner recommends treating compute placement as a multitiered decision rather than putting every AI workload into one geographically agnostic cloud.
Energy also needs to become part of vendor selection. Organizations evaluating sovereign infrastructure providers may need to look at the physical location of facilities, local grid capacity, energy sources, and the availability of dedicated or renewable power.
For companies that depend on AI around the clock, the question is not simply whether a provider has enough GPUs today. It is whether those resources will remain available when demand rises, grids become constrained, or international supply chains are disrupted.
AI Agents Are Creating a New Kind of Security Problem
The move from AI assistants to autonomous agents is adding another layer of complexity. An agent can retrieve information, communicate with other systems, make decisions, and execute a series of actions without waiting for a person at every stage.
Traditional security reviews are not designed for that kind of machine-speed activity. A problem can occur while an application is running, rather than during a scheduled compliance review.
Gartner places Bleu, KNTRA, Prem, Qubittum, S3NS, Saptiva AI, Votal AI, and Zama in the deterministic governance and trusted execution category.
The companies are approaching the problem from different directions. Prem, for example, provides encrypted, self-hosted workspaces. Bleu and S3NS focus on sovereign-cloud environments designed to keep sensitive workloads under national oversight. Other platforms use runtime controls, deterministic logic, or cryptographic techniques such as fully homomorphic encryption.
Votal AI is highlighted for securing enterprise LLM and agentic deployments through real-time guardrails and continuous agentic red teaming, including work connected to the NVIDIA-led Open Secure AI Alliance.
The broader change is straightforward: AI governance is moving closer to the point where an AI system actually takes action.
The Cloud Stack Is Becoming More Flexible and More Fragmented
Another pressure point is dependence on a single cloud provider. While a closed environment can make an AI project easier to launch, moving its models, data, and applications later can be difficult.
Gartner identifies CerebrixOS, Joule from SAP, Prime Intellect, Stack8s, and Together AI in the multicloud and decentralized orchestration category. These companies are developing software layers that connect different computing resources, allowing workloads to run across public clouds, private infrastructure, on-premises systems, and other GPU sources.
Together AI and Prime Intellect work across diverse GPU resources, while Stack8s and CerebrixOS provide multicloud control capabilities. Joule from SAP supports enterprise workflows within a secure business environment.
This approach gives organizations greater flexibility when pricing, capacity, performance, or regional regulations change. Gartner recommends model-agnostic systems, open-standard APIs, and multicloud orchestration so that models or workloads can be moved without rebuilding the entire system.
Regional Models Are Filling the Gaps Left by Global AI
There is also a less visible side to AI sovereignty: language. Gartner notes that English makes up nearly half of the content in major open web crawl repositories, while Arabic accounts for less than 1%. This imbalance can leave local dialects, cultural references, and legal or administrative contexts underrepresented in AI systems.
Gartner identifies GenerIA, Krutrim AI Labs, Mistral, Sakana AI, and Sarvam AI in the sovereign models and regional languages category. Their approaches vary. Mistral focuses on open-weight models and flexible deployment, while Sakana AI uses evolutionary model merging to develop and optimize models without relying entirely on the cost of training massive systems from scratch.
In India, Sarvam AI and Krutrim AI Labs are developing models around Indian languages and domestic infrastructure, while GenerIA focuses on applications for public administration and regulated environments.
Open-weight models also give organizations more control. They can be fine-tuned for local needs and deployed within their own infrastructure or secure environments, reducing the need to send sensitive information through closed external APIs.
Sovereign AI Is Becoming an Infrastructure Decision
Gartner’s research points to a market that is still taking shape, but its direction is becoming clearer. By 2027, Gartner expects 35% of countries to rely on region-specific AI platforms using localized data. By 2030, it projects that European organizations will shift 20% of their technology budgets toward European-registered sovereign services.
For organizations, the focus now is on practical choices such as hybrid deployment, portable architectures, local fine-tuning, decentralized compute, and real-time controls for autonomous systems. Sovereign AI does not mean moving every workload into a private data center. It means identifying which systems require tighter control and building the right infrastructure around them.
As AI becomes more deeply connected to finance, healthcare, public services, and industrial operations, control over the technology behind it will matter just as much as the capabilities of the models themselves.
