The rapid early expansion of frontier AI capabilities was largely driven by a few powerful, U.S.-based labs. In recent years, however, open-weight models released by labs in China — notably DeepSeek — and later by Moonshot, Z.ai, and others have partly disrupted that dominance. Growing concerns about concentrated power have prompted many governments and institutions to talk about the need for some degree of AI sovereignty.
AI sovereignty does not necessarily mean full exclusive control of an AI stack. For states or organizations it promises localized security and privacy, better compliance with local laws (for example EU AI and data regulations), more reliable service, and AI outputs that may be more culturally specific. Interdependence can be acceptable, but it is important that choices go beyond only open-weight models or commercial offerings.
Big labs offering sovereignty solutions
Unsurprisingly, the same large AI labs that helped create dependency are also marketing their own sovereignty solutions. While some local institutions consider or adopt these initial offerings, open-source AI technologies are presented as an alternative that could provide more flexibility and a way to manage dependencies. Tim O’Reilly has argued that open-source AI could open the door to greater participation in AI’s future development.
From the chip technology needed for training and inference to the cloud and data infrastructure enabling model development, companies such as NVIDIA, OpenAI, Google, Microsoft, and Amazon Web Services are staking claims to remain relevant in the global push for AI sovereignty. Stanford University’s Human-Centered Artificial Intelligence Lab (HAI) details these approaches in its report The Commercial Landscape of AI Sovereignty Offerings, noting that while these labs “promise that countries will own their AI stack,” they often deepen dependencies on U.S. Big Tech.
There are commercial alternatives outside the U.S. offering full-stack solutions for particular layers of the stack: companies in Europe, the Gulf region, Asia and elsewhere are positioning themselves as local alternatives. The HAI report highlights that many of the most mature providers are actively backed by their governments; in such cases sovereignty is a stated policy objective, with governments directing funding, shaping procurement, and sometimes selecting companies to build domestic AI capacity.
But such collaborations can also open the door to political intervention. Claims about censorship and control were levelled soon after DeepSeek’s initial 2025 release, and there have been probes into whether U.S. models limit certain types of discourse. The HAI report also points out that many alternative providers still rely on foundational technologies — particularly chips and cloud infrastructure — that are often tied to U.S. suppliers.
Open source as a diversification strategy
The proliferation of commercial sovereignty offerings gives room for diversification and potential leverage to negotiate with dominant players. Open-source AI also plays an important role in creating opportunities for broader diversification and greater sovereignty. HAI warns that sovereignty strategies ignoring open-source AI risk normalizing fragmentation and political overreach.
For open-source AI to counter the diversification of commercial offerings, it must expand beyond model weights. Tim O’Reilly argues for a federated system of open-source AI — a federation of models, protocols, code and capacity — to ensure both broad deployment and broad participation across the stack.
Orchestration and agent harnesses
A key element in expanding the open-source stack is the orchestration layer above model weights, often implemented as agent harnesses. Raffi Krikorian, Chief Technology Officer at Mozilla, has argued that capability is concentrating at this orchestration layer and that closed labs are seeking to lock it down; thus, building open harnesses is as important as open models.
Commercial products that have dominated include Claude Code and Codex. OpenClaw briefly disrupted the open-source space in late 2025 before its creator joined OpenAI. NousResearch’s self-improving Hermes agent harness has attracted substantial attention, with over 230,000 stars on GitHub. More recently, harnesses such as Pi and DeepSeek Harness are expanding the open-source offering in line with Krikorian’s call.
Less visible but critical layers: inference, storage, interoperability
Some of the strongest open-source projects operate in less visible but critical layers. Inference engines such as vLLM, SGLang, llama.cpp and ONNX Runtime enable efficient serving of a range of models across data centers, regional clouds, personal computers and edge devices. Ray — developed by researchers at UC Berkeley — distributes demanding AI workloads, while Ollama lowers the barrier to running models locally. Together, these projects allow institutions to swap models, hardware and hosting providers without rebuilding their stack around a proprietary platform.
Other fast-growing projects fill out data, interoperability and accountability layers. The Model Context Protocol and Agent2Agent Protocol provide open standards for agents to connect to tools and to one another. Qdrant, Chroma, Milvus and LanceDB offer open infrastructure for storing and retrieving institutional knowledge. MLflow, Opik and OpenLLMetry let developers evaluate, trace and monitor AI applications without surrendering operational data to a closed dashboard. The AI Potluck Gap Map classifies inference, deployment and agent protocols as mature open ecosystems, while identifying resilience gaps in storage and observability, pointing to needed investments.
Sovereignty will depend less on finding a single open replacement for Big Tech and more on maintaining interoperable public alternatives across every consequential layer of the stack.
Acquisition risk and the need for robustness
Despite these open alternatives, the threat of acquisition and absorption remains. The fintech company Stripe recently acquired OpenRouter, a platform that enables developers to access and route various models and has become a central hub for open-weight models. NVIDIA has acquired Hugging Face, a key player in the open-source AI ecosystem, and recently concluded a licensing deal with Poolside, which develops open-weight coding models — reportedly to avoid the oversight of a full acquisition.
AI sovereignty may in practice mean calibrating interdependence with U.S. Big Tech or temporarily replacing a foreign dependency with a domestic one. But genuine choice requires building technical capacity, open infrastructure and participatory institutions across the stack. Open-source projects remain vulnerable to commercial absorption, and efforts to protect and sustain them must become more robust.
Survey
The article’s source invites professionals involved in cybersecurity to answer a short 11-question survey for a report.
