'International Perspective on GenAI Views from the Frontiers' | RAISE Summit 2024 | Paris

Geopolitical Shifts and Sovereign AI Trends

• Global investment strategies are increasingly aligning with "sovereign AI" concepts, driven by the perception that AI infrastructure requires specific geopolitical control.
• Saudi Arabia is rumored to be establishing a $40 billion fund dedicated specifically to AI investments.
• The US and EU are heavily investing in government and corporate AI initiatives, with the EU AI Act creating a framework intended to establish Europe as a "global AI trust hub."
• National governments are prioritizing the development of "national champions" in large language models (LLMs) to control domestic data and culture.
• Singapore and the Middle East are explicitly framing AI strategies to align with national investments, viewing AI as a tool to uplift human potential in sectors like healthcare.

Technical Challenges in Low-Resource and Multilingual Environments

• English dominates web content (>50%), yet represents less than 20% of the global population, creating a significant data imbalance for global AI deployment.
• India contributes only ~1% of global digital data despite representing 18% of the world's population and hosting 22 official languages and thousands of dialects.
• Standard tokenizers fragment non-Latin languages inefficiently; for example, a single Hindi word may split into 10 tokens compared to 2 for English, increasing inference costs and slowing generation speeds.
• Developers in non-English regions face a lack of local evaluation benchmarks and must build custom tokenizers and test sets from scratch to achieve usable performance.
• Multimodal challenges include significant "code-mixing" (switching between three languages in a single conversation), which is rare in training data and difficult for models to interpret naturally.

Strategic Responses from Industry Leaders

• Temasek (Benjamin): Identifies AI as a dual-use technology with risks of societal disruption (e.g., disinformation, deepfakes) and predicts a future of 27+ potential EU LLMs that may struggle against global scale economics.
• Scale AI (Michael): Emphasizes that the future geopolitical choice is between Western vs. CCP-aligned base models; countries must choose LLMs that reflect their specific cultural values and language.
• Samba Nova (Rodrigo): Built a full-stack solution (chip to model) to enable private, secure deployment; recently released "Samba Lingo" open-source models achieving world records in nine low-resource languages (e.g., Hungarian, Arabic, Thai) using techniques that leverage cross-language learning.
• Criterium (Chandra): Developed a custom tokenizer and full AI stack to address India's specific linguistic fragmentation and lack of data; notes that regulatory constraints were initially proposed but removed to allow startup agility.
• Qtai (Patrick): Established an open science lab to counter the trend of closed models, focusing on high-quality, trustworthy open models and ensuring data creators benefit from AI advancements.

Market Dynamics and Investment Outlook

• Western Europe is experiencing slower AI adoption rates compared to the US and Middle East, potentially due to regulatory uncertainty stifling entrepreneurship.
• The UAE's rapid development of the Falcon open-source model demonstrated that with sufficient data and compute, new regional leaders can emerge quickly.
• Investors anticipate a market structure defined by a proliferation of local LLM champions alongside a strong open-source sector.
• Future growth from hundreds of millions to billions of users depends on overcoming technical barriers related to language barriers and cultural biases.
• While B2C adoption is currently high, long-term large-scale adoption in the public sector and defense requires addressing specific security, safety, and robustness requirements.
• Long-term market conviction is being held loosely due to the rapid pace of architectural and data availability changes, with a 5-10 year horizon expected for significant shifts.