The State of AI: Rise of Reasoning, Surge in Chinese Open Source, Sovereign AI , Investing in AI Today | Nathan Benaich, Air Street Capital
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
Nathan Benaich, founder of Air Street Capital and author of the State of AI Report, shares insights from his eighth annual deep dive into the most significant developments in AI. (03:39) The conversation explores the biggest takeaways from his latest report, including the revolutionary shift toward reasoning and tool calling capabilities, the surge in China's open source models challenging Western dominance, the emergence of real AI revenue at scale, and the rise of sovereign AI initiatives globally. (36:36) Nathan also discusses whether we're in an AI bubble, his investment strategy at Air Street Capital, and where he believes value will ultimately accrue in the AI ecosystem.
- Core themes: AI's evolution from simple input-output models to sophisticated reasoning systems, geopolitical shifts in AI development, and the maturation of AI as a commercially viable technology
Speakers
Nathan Benaich
Nathan Benaich is the founder of Air Street Capital and author of the State of AI Report, which he has been producing for eight years. He holds a PhD and previously worked at venture firms including Point Nine Capital and Playfair Capital before launching Air Street in 2019. Nathan has been actively investing in AI companies since the early days of the current wave, with a focus on vertical software, DevTools, defense/security, and tech bio sectors.
Turner Novak
Turner Novak is the founder of Banana Capital and host of The Peel podcast. He focuses on early-stage venture investing and has built a platform exploring the world's greatest startup stories through in-depth conversations with founders and investors.
Key Takeaways
AI Has Evolved from Simple Lookup to Complex Reasoning
The most significant development in AI over the past year has been the transition from models that simply retrieve memorized information to systems capable of step-by-step reasoning and tool calling. (06:22) This fundamental shift enables AI systems to access real-time information through web searches, use APIs, and interface with software products rather than relying solely on training data. The capability is visible when users see "thinking" indicators in ChatGPT, showing the model's reasoning process. However, research suggests these reasoning traces may not always be truthful - models might be showing what humans expect to see rather than their actual thinking process.
China's Open Source Models Are Reshaping Global AI Competition
Chinese companies have emerged as major players in open source AI development, with DeepSeek and other firms challenging the dominance of American closed-source models. (13:01) This shift represents a significant change from just 12 months ago when Meta's LLaMA was the leading open source initiative. Chinese models particularly excel in world modeling and vision capabilities, especially for generating pictures and long-form videos. Two Chinese AI companies, Minimax and Knowledge Atlas Company (GLM), have already gone public on the Hong Kong Exchange for several billion dollars each, marking the first pure-play model companies to reach public markets.
AI Revenue Has Reached Massive Scale Across Major Players
The AI industry has transitioned from minimal revenue just two years ago to tens of billions of dollars across major players. (26:46) Previously, companies like Jasper were generating more revenue than OpenAI for similar use cases, raising questions about whether value would accrue to model providers or application layers. This dynamic has flipped as core models improved and became interfaces for everything, with model vendors creating their own SaaS wrappers. The rapid revenue growth demonstrates genuine market demand rather than speculative investment.
Sovereign AI Has Become a Global Marketing Phenomenon
Nation states worldwide are investing over $100 billion in data center capacity under the banner of "AI sovereignty," driven heavily by NVIDIA's marketing efforts. (27:51) The concept promises countries the ability to train and run AI models domestically without foreign interference, but faces practical challenges around hardware improvement cycles, software dependencies, and vendor control. Despite these limitations, the sovereign AI narrative has created significant customer diversification opportunities for hardware providers and represents a major geopolitical shift in how nations view AI capabilities as critical infrastructure.
Successful AI Companies Focus on Product Experience Over Model Ownership
Building a successful AI company doesn't necessarily require training your own models - the key factors are product experience, taste, and user data collection. (100:05) Many successful AI applications can be built using existing foundation models, allowing companies to focus resources on product-market fit and growth rather than expensive R&D. The defensibility comes through superior customer relationships, proprietary data, and user preference feedback rather than model architecture. Companies should only train custom models when existing systems cannot solve their specific problem or when competing directly in the model race.
Statistics & Facts
- Over $100 billion in global commitments have been made by different countries to build data center capacity under sovereign AI initiatives. (27:24) This represents a massive shift in national infrastructure investment priorities driven by geopolitical concerns about AI dependency.
- OpenAI and Anthropic have added more new revenue since ChatGPT's launch than every other publicly traded software company combined during the same period. (40:13) This statistic was from approximately six months ago, so the differential has likely grown even larger.
- Some of the best AI model companies are achieving 60-70% gross margins on serving their models, a significant improvement from previous periods when margins were unclear or negative. (41:05) This improvement comes from better training efficiency and higher model quality that commands premium pricing.
