AI Sovereignty: Bargaining with Big Tech and the Promise of Full Stack Open Source AI

Nicole Butterfield · rss · 2026-09-22 Reader 🖥️

Agent Harness Agent-to-Agent Protocols AI Regulation & Governance (External)Compute & Energy Infrastructure Data Sovereignty MCP (Model Context Protocol)Open-Weight Models

Genuine AI sovereignty requires building and sustaining an interoperable, full-stack open-source AI ecosystem — spanning agent harnesses, inference engines, protocols, data stores, and observability tooling, not just open-weight models or vendor-branded 'sovereignty' offerings — because dependency and acquisition pressure persist at every layer that stays closed or commercially controlled.

practical · expository — Argues for a course of action (institutions and the open-source community should build/adopt full-stack open AI infrastructure) rather than merely describing the sovereignty landscape; structured as a survey of stack layers in service of that recommendation.

Structure

Part	Role
Definition of AI sovereignty as calibrated interdependence, not total stack control	context
Critique of Big Tech labs' own 'sovereignty' offerings (Stanford HAI report: these 'deepen dependencies on U.S. Big Tech')	evidence
Non-US government-backed 'full stack' alternatives, and the political-intervention risk they also carry	evidence/complication
Layer-by-layer inventory of the open-source AI stack (agent harnesses, inference engines, orchestration/data protocols, vector stores, observability)	evidence
Acquisition case studies (Stripe/OpenRouter, NVIDIA/Hugging Face, NVIDIA/Poolside) showing continued absorption pressure on open infrastructure	complication
Closing reframe: sovereignty as calibrating interdependence and sustaining interoperable public alternatives across every layer	conclusion

Key terms

AI sovereignty
Not total control of one's AI stack, but having enough localized security/privacy, jurisdictional compliance, service reliability, and cultural fit — plus enough alternatives beyond commercial offerings — to negotiate interdependence from a position of leverage rather than being locked in.
architecture of participation
(Quoting Tim O'Reilly) A federated system — of models, protocols/code, and capacity — designed so that both 'being everywhere' and 'everyone having a say' are structurally possible, as opposed to a single dominant stack.
agent harness Agent Harness
The orchestration layer sitting above model weights, cited (via Mozilla CTO Raffi Krikorian) as where capability is now concentrating and which closed labs are 'welding shut' — making open harnesses (not just open weights) a key sovereignty battleground.

Propositions

explicitAI sovereignty does not require total control of the AI stack — it means having localized security/privacy, better legal-jurisdiction fit, more dependable service, and options beyond purely commercial offerings.conf: high
explicitThe same AI labs whose dominance created the demand for AI sovereignty are also selling their own 'sovereignty' offerings, which a Stanford HAI report finds still 'deepen dependencies on U.S. Big Tech' even while promising local ownership.conf: high
explicitSome non-US 'full stack' AI alternatives are actively government-backed, making sovereignty for them 'not just a marketing claim but a stated policy objective.'conf: high
implicitGovernment-backed alternative AI providers can create openings for political intervention and censorship, just as dependency on US labs can.conf: medium
explicitReal AI sovereignty requires open-source technology to proliferate beyond open-weight models into every layer of the stack: agent harnesses, inference engines, orchestration/data protocols, vector/knowledge stores, and observability/evaluation tooling.conf: high
extendsHow a Culture of Data-Driven Conversations Can Support Platform Engineering · p8 — Extends the abstract claim that digital sovereignty 'must be a first-class design constraint' with a concrete, layer-by-layer inventory of what open-source sovereignty actually requires (harnesses, inference, protocols, storage, observability).
extendsAI models can disappear overnight. Is your engineering team built to survive it? · p10 — Extends the claim that running open-weight models grants a company 'digital autonomy' by arguing that self-hosting open weights alone is insufficient sovereignty — the harness, inference, protocol, storage, and observability layers must also stay open, or the same acquisition/vendor risk simply reappears one layer up.
explicitThe orchestration layer above model weights (agent harnesses) is where capability is now concentrating, and closed labs are already 'welding it shut' — making open harnesses strategically as important as open weights.conf: high
explicitStorage and observability are the weakest/least mature layers of the open-source AI stack, while inference, deployment, and agent protocols are comparatively mature open ecosystems.conf: high
explicitOpen-source AI infrastructure remains persistently vulnerable to commercial acquisition and absorption, so proliferation of open projects alone does not guarantee durable independence.conf: high
supportsI've been wearing two AI recorders. Only one belongs near a client. · p7 — Supports the broader pattern that big-tech ownership consolidation of independent AI products is itself a standing risk distinct from any single product's engineering — there Meta/Limitless and Amazon/Bee, here Stripe/OpenRouter and NVIDIA/Hugging Face.
implicit'Calibrating interdependence' — not achieving full independence — is the realistic goal of AI sovereignty; sovereignty is a matter of degree and leverage, not a binary state.conf: medium
explicitThe Model Context Protocol and Agent2Agent Protocol are the open standards currently filling the orchestration/data-interoperability layer of the open-source AI stack, letting agents connect to tools and to one another without a single vendor controlling the interface.conf: high
supportsThe Agentic P&L: Beyond the Empire of Headcount · p15 — Supports the prediction that an MCP-style protocol layer becomes de facto federated-enterprise infrastructure ('TCP/IP of the federated enterprise') by naming MCP and Agent2Agent as the two protocols already doing exactly that federating job in the open-source AI stack.

Arguments

Genuine AI sovereignty requires a federated, full-stack open-source ecosystem, not just open-weight models or vendor-branded 'sovereignty' offerings. strength: medium

Explicit premises:

Big Tech's own sovereignty offerings still deepen U.S. dependency (per HAI report).
Open-weight models alone leave the harness, inference, protocol, storage, and observability layers under commercial or closed control.
Open infrastructure projects already exist across most layers (vLLM, Ray, Ollama, MCP/A2A, Qdrant, MLflow, etc.), but resiliency gaps remain in storage and observability (per AI Potluck Gap Map).

Implicit premises:

Sovereignty is fundamentally about maintaining genuine choice and negotiating leverage, not about eliminating dependency entirely.
A single closed or immature layer in the stack can recreate a chokepoint even when every other layer is open.

Inferences:

Because storage and observability remain the least mature open layers, they are the most likely near-term battleground for further commercial acquisition — a prediction the piece's own examples (Hugging Face, OpenRouter) already partly illustrate at the model-routing/hub layer.

Narratives

DeepSeek's 2025 release prompted claims of censorship and control in Chinese models; more recently, probes have examined how US models may also limit certain discourse.
→ Implicitly supports the proposition that political-control risk is not unique to one geopolitical bloc — both US-dependency and non-US alternative-dependency carry it.
Stripe's acquisition of OpenRouter, NVIDIA's acquisition of Hugging Face, and NVIDIA's licensing deal with Poolside ("purportedly to avoid the oversight of complete acquisition")
→ Supports the proposition that open-source AI infrastructure remains persistently vulnerable to commercial absorption, and grounds the piece's closing warning in concrete, near-term examples.

Questions addressed

✅ What does AI sovereignty actually require, beyond simply adopting open-weight models?
A federated, full-stack open infrastructure spanning agent harnesses, inference engines, protocols, data/knowledge stores, and observability/evaluation tooling.
✅ Which layers of the open-source AI stack are currently underserved?
Storage and observability, per the AI Potluck Gap Map, versus more mature inference/deployment/agent-protocol ecosystems.
⭕ How much interdependence should an institution accept, and by what criteria should it 'calibrate' that interdependence?

Gaps

incompleteThe piece never quantifies how much of the stack must be open before an institution can credibly claim sovereignty — no threshold, benchmark, or worked case study of an institution that achieved durable, measurable independence using the named open projects.
Read next: Look for a case study or audit framework (e.g. a specific EU or Nordic public-sector AI deployment) that operationalizes 'how open is open enough.'
incompleteNo discussion of the practical switching costs and operational friction of actually adopting the named open alternatives (vLLM, Ray, Qdrant, MLflow, etc.) at production scale for an institution accustomed to hyperscaler tooling.
Read next: Read a migration/build-vs-buy account that weighs the engineering cost of assembling an open stack against the convenience of a managed commercial one.
incompleteThe piece stays at the global/US-vs-alternatives level and never engages the EU/Nordic public-sector sovereignty debate already present elsewhere in this corpus (e.g. Norwegian NAV/Entur procurement, or the Heis podcast's coverage of AI in the Norwegian public sector) — a natural connection it leaves unmade.
Read next: Connect to extract-jannik-william-frederiksen-post and extract-heis-11-ai-spesial-vilde-opsal-anders-sveen for a concrete Nordic instance of the same calibration problem.

Closes gaps elsewhere

16) Inga Strümke om hvor AI er på vei, den outsourcede problemløsningen og datamanifolden · g4 — Doesn't resolve the specific Llama/Mistral-vs-Copilot question, but reframes it: the piece argues the open-weight-vs-closed-API binary is the wrong lever entirely, since sovereignty depends on which layers of the whole stack (harnesses, inference, protocols, storage, observability) stay open, not on model weights alone.
How a Culture of Data-Driven Conversations Can Support Platform Engineering · g5 — That gap flagged sovereignty as 'bolted on, not integrated' — mentioned as a must-have without elaboration. This piece supplies exactly that missing integration: a full-stack breakdown of which open-source layers sovereignty actually depends on, and where the gaps in that stack currently sit.