From Dependency to Autonomy: The Case for Open Source AI | AI Impact Summit 2026

Contents

Executive Summary

This panel discussion from the AI Impact Summit 2026 examines how open source AI can break dependency on monopolistic commercial models and build resilience for nations, enterprises, and individuals. Panelists argue that rather than pursuing nationalist "sovereignty," the focus should be on decentralized AI development, local models, equitable access, and cross-border collaboration to ensure competitive advantage, cultural representation, and democratic control over AI technology.

Key Takeaways

Reframe "Sovereignty" as "Autonomy through Access": The real goal isn't isolation—it's ensuring nations, enterprises, and individuals can access, understand, modify, and run AI systems locally. This requires open source as infrastructure, not nationalist tech stacks.
Small Models Are Strategic Infrastructure: Invest in commoditized, domain-specific smaller models and open standards (similar to how Linux became the foundation for everything above). Frontier model races are won by hyperscalers; real-world resilience is built on lean, adaptable tools.
Data Licensing and Transparency Are Non-Negotiable: Any credible open-source AI project must be transparent about training data provenance and attribution. This differentiates genuine open source from "open washing" and prevents repeating the pattern of unconsented scraping.
Open Source Under Geopolitical Pressure Requires Vigilance: The political climate is pushing nationalist framing of open projects. Communities must actively defend open principles (interoperability, cross-border collaboration) against co-optation into "sovereign tech" or "local source" rebranding.
Technologists Have Power to Shape Values: Engineers, maintainers, and organizations can collectively resist pressure to build unfairly or inequitably—but only if they organize openly and collaboratively. The alternative (individual capitulation) strengthens centralized control.

Key Topics Covered

Dependency vs. Sovereignty: The distinction between geopolitical sovereignty (nationalistic) and technological autonomy (access, control, resilience)
Local and Small Language Models: Why small, domain-specific models matter more than frontier models for real-world deployment
Open Source as Infrastructure: How commoditization and open standards enable innovation without centralized control
Data Sovereignty and Ethics: Training data transparency, attribution, and the problem of AI companies using copyrighted material without consent
Geopolitical Pressures on Tech: How political shifts (US policy on diversity/inclusion, EU "sovereign" AI initiatives, China's open-source strategy) affect collaborative tech development
Public Sector AI Adoption: How open-source models enable governments and public services to serve citizens equitably
Labor, Power, and Decentralization: GPU access, compute constraints, and who controls the means of AI production
"Open Washing" and "Local Source": The risk that nationalist framing will undermine genuine open-source principles
Diversity in AI Development: Cultural representation, language models tailored to local contexts, and loss of diversity under monoculture models
Funding and Ecosystem Sustainability: How open communities are funded and maintained when hyperscalers extract value

Key Points & Insights

Sovereignty ≠ Isolationism: True technological autonomy comes through cross-border collaboration in open communities, not nationalist protectionism. Open source has historically succeeded because it operates across geopolitical boundaries.
The Small Model Imperative: Not every use case requires frontier models. Small, domain-specific models (7B parameters, running on 8GB RAM) serve local needs better and are more deployable at scale. This is especially critical for public sector adoption where a small government has 6 million employees but only 4 GPUs to serve them.
Data is the New Battleground: The training data is more important than model weights or code. Current AI companies are training on scraped, unconsented data. Open-source AI projects should transparently document their training data, and data licensing frameworks (like India's proposed dataset license) are essential.
GPU Access Concentration Threatens Decentralization: AI development is shifting from code/IP battles to hardware constraints. If only entities with massive data centers can train models, power consolidates. Miniaturization and edge computing become political necessities, not engineering choices.
"Local Source" vs. "Open Source": EU and some national initiatives are rebranding protectionist tech programs as "sovereign AI" or "open source." This "open washing" or "local source" approach undermines genuine open-source values while appearing to embrace them—a warning sign detected by Russia's experience with isolation.
Geopolitical Realism is Necessary: Balkanization of the internet and competing national tech stacks are already happening. Open-source communities should acknowledge this while refusing to weaponize open source itself. Parallel systems (different models for different regions) will exist—the question is whether they remain interoperable.
Diversity and Bias in AI Require Distributed Governance: Homogeneous teams make worse decisions (87% worse by one cited study). A single, globally-dominant model trained on the same data amplifies cultural bias and eliminates regional context (e.g., understanding "tea" as both drink and meal in British English).
Business Model Innovation in Open AI: Companies can compete on domain adaptation, system integration, and inference optimization—not just model size. Open models allow enterprises to avoid vendor lock-in while still generating competitive advantage through specialization.
Funding Models for Open Communities are Under Strain: Wikipedia, OpenSSL, and other critical commons are underfunded. When hyperscalers extract value (via scraping or training), it costs the open community disproportionately to maintain. Licensing agreements and contribution requirements need rethinking.
Building Technology is Building Power: Technologists and open-source maintainers have agency to reject political pressure (e.g., Python Foundation's rejection of US government funding restrictions). Open collaboration allows communities to collectively define what "fair and equitable AI" means and enforce it through transparency and shared standards.

Notable Quotes or Statements

"I don't like sovereignty because I'm a globalist and a member of the open source community. We've achieved what we've achieved by working across geopolitical boundaries."
— Amanda Brock, CEO, Open UK

"You wouldn't really want to use an AI trained only on X. It would be very angry and would only fight with you."
— Jimmy Wales, Founder, Wikipedia (on the importance of diverse training data)

"Sovereignty doesn't mean solitude. You don't have to operate by yourself. You can build in collaboration in communities."
— Mishi Choudhary, General Counsel, Software Freedom Law Center

"If we do not act now in terms of decentralization, we might finish in a society where the only political power is with those who own GPUs for data centers."
— Anastasia Stenko, CEO & Co-founder, Pers (Paris-based AI company)

"The models are actually becoming the interface through which we interact with the digital world and the web. If we want individuals to be truly empowered, we need to give communities and individuals the power to develop their models."
— Anastasia Stenko

"When a large hyperscaler's bot crawls Wikipedia, it hits every obscure page and we need more database servers. The cost is disproportionate. Our donors aren't donating to subsidize huge AI companies."
— Jimmy Wales (on the economic externalities of AI scraping)

"We are losing competitiveness by introducing the same exact tools which completely get out all of our data which is very specific to some of the business processes."
— Anastasia Stenko (on enterprise dependency on proprietary closed models)

"I sometimes call it local source, not open source. This is open washing—a different kind of open washing that's just evolving."
— Amanda Brock (on nationalist AI framing)

"People have to make decisions, and our minds are ours. How our minds are used are our choices. That's how I see sovereignty."
— Mishi Choudhary

"We have the power to say, 'We hear you don't care about bias, but we still do, and we're going to build it that way.'"
— Laura Gilbert, formerly UK Government, now Tony Blair Institute (on technologist agency)

"Never stop building."
— Anastasia Stenko (closing statement)

Speakers & Organizations Mentioned

Name	Role / Organization	Context
Amanda Brock	CEO, Open UK	Moderator; framed discussion around dependency vs. sovereignty
Jimmy Wales	Founder, Wikipedia	Discussed local LLMs, privacy, data reuse, and training data economics
Mishi Choudhary	General Counsel, Software Freedom Law Center (Founder)	Legal expert on open source; discussed collaboration, India's AI strategy, and global majority leadership
Anastasia Stenko	CEO & Co-founder, Pers (Paris)	Built open-source small language models; discussed edge computing, business competition, and decentralization
Laura Gilbert	Formerly UK Government (Number 10), now Tony Blair Institute	Public sector AI policy; discussed resilience, inequality reduction, and technologist agency
Mishi Choudhary (also)	General Counsel, DC-based cybersecurity company	Additional context on commercial open-source use
Deb Nicholson	Executive Director (?), Python Foundation	Mentioned for rejecting $1M in US government funding over political compliance requirements

Organizations/Initiatives Mentioned:

Open UK
Software Freedom Law Center
Wikipedia
Pers (Paris-based AI company)
Tony Blair Institute
Number 10 (UK Prime Minister's Office)
Python Foundation
Linux ecosystem
Apache Software Foundation
Debian
Open Atom Foundation (China)
Mozilla Foundation
NScale (UK data center work)
Vector Institute (Canada)
Munich Security Conference
African Union (55 countries)

Governments/Policy Contexts:

UK (Brexit impacts on EU collaboration)
India (open-source strategy, data licensing, leadership in Global South)
China (2017 open-source pivot, five-year plans, lean AI engineering)
France (sovereign AI initiatives, EU framing)
Russia (cautionary example of isolation and loss of open-source engagement)
US (policy shifts on diversity/inclusion, tariffs, political pressure on tech companies)
EU (sovereign AI funding, nationalistic framing)

Technical Concepts & Resources

Concept / Tool	Context / Significance
Local Large Language Models (LLMs)	Running models on personal hardware (128GB RAM laptops) to avoid cloud dependency; enables privacy and offline use
Small Models / Smaller Parameter Models	7B parameter models, 8GB RAM requirements; domain-specific, deployable at scale without massive data centers
Model Fine-tuning	Domain adaptation via fine-tuning on Llama, Qwen; not always sufficient for specialized needs (e.g., legal, medical)
Edge Computing / Edge Deployment	Running inference on edge devices (Android, IoT); critical for decentralization
Frontier Models	Large, generalist models from OpenAI, Google, Anthropic; resource-intensive, closed-source or restrictive license
Training Data Transparency	Open documentation of data sources, provenance, licensing; differentiates genuine open source from "open washing"
Falcon & Llama (Llama 2, Llama 3)	First commercial-license open-source models; used in France's sovereign AI proof-of-concept
Model Weights & Code	More easily shared than training data; data is the strategic layer
FIPS 140-2 Module	US federal government cryptographic standard; discussed in context of open-source community supporting critical infrastructure
Open Standards (N2CP Protocol, others)	Commoditization of underlying technical layers; enables plug-and-play interoperability
Hardware Constraints (GPU, Memory, Energy)	Physical bottlenecks to decentralization; GPU access concentration is a political economy issue
Inference vs. Training	Inference (running models) is cheaper and more distributed; training (creating models) remains centralized
Agentic Turn	Recent architectural shift toward autonomous agent models; led by OpenAI, Google, Anthropic
Claude, GPT, Gemini	Commercial closed-source models; used as comparison points for open alternatives
Wikipedia (Dataset, Training Source)	Heavily used in LLM training; brings bias toward WEIRD (Western, Educated, Industrialized, Rich, Democratic) perspectives; economics of open knowledge commons
Hugging Face / Model Repositories	Implied infrastructure for sharing open-source models
Cross-Border Open Communities	Kernel, Apache, Debian as examples of geopolitically distributed development teams

Additional Context & Themes

Geopolitical Landscape Shifts:

US policy moves away from diversity/inclusion ("safety") toward "security" only
EU framing of "sovereign AI" with funding, but risk of rebranding protectionism as open source
China's 2017 pivot to open source and continued investment in lean, small models (manufacturing-focused strategy)
India positioning as leader of the Global Majority; attention to smaller nations' needs
Russia's historical example: isolation in tech led to loss of participation in global open-source ecosystem

The "Year of Ontology" Problem:

Terms like "civic tech," "public good," "digital public good," "sovereign AI," "open source" are being redefined in politically motivated ways
Lack of shared definition creates confusion and enables co-optation (open washing)
Panelists advocate for precision and fidelity to core open-source values (transparency, interoperability, collective governance)

Labour & Economic Dimensions:

Hyperscalers' strategy: eliminate labor through AI automation
Open-source alternative: enable distributed small-company ecosystems that create local jobs and services
Public sector adoption of open source can spin up regional SMEs, improve service delivery, reduce inequality
Funding asymmetry: open communities are underfunded while hyperscalers extract value from their work

Diversity & Representation:

Single global models homogenize culture, language, and knowledge representation
Small, local models allow regional/cultural context (tea, history, institutional knowledge)
Diverse teams make better decisions; monoculture models encode monoculture values
Risk of quiet, invisible loss of autonomy (not dramatic, but cumulative)