After DeepSeek Episode 2

Key Concepts

• DeepSeek Model: A fully open-source AI model released in January 2025, considered a turning point in China’s AI development.
• Sputnik Moment: An analogy used to describe DeepSeek’s impact, comparing it to the Soviet Union’s launch of Sputnik and its challenge to US dominance.
• AI Efficiency: Focus on reducing computing resource requirements for AI models, particularly through techniques like low-bit training (FP8).
• Decentralized Generative AI: Training AI models locally at the data source (e.g., hospitals) to address data privacy and access challenges.
• AGI (Artificial General Intelligence): Hypothetical AI with human-level cognitive abilities.
• AI Plus Strategy: Beijing’s roadmap launched in August 2025 to integrate AI across various sectors (manufacturing, agriculture, services).
• FP8 (Floating Point 8): A calculation format using 8 bits for numerical data in AI models, reducing computing resource needs.
• Transformer Model/BERT: Foundational architecture for modern large language models, developed by Google.

China’s After DeepSeek Era: A Technological Shift

Introduction: The DeepSeek Disruption

The release of the DeepSeek AI model in January 2025 is widely regarded as a pivotal moment for China’s artificial intelligence landscape. This event, likened to the Soviet Union’s launch of Sputnik, challenged the prevailing assumption that only US companies with substantial computing power could achieve world-class AI capabilities. This podcast episode explores the technical, economic, and geopolitical ramifications of DeepSeek, examining the new realities of China’s AI development and its focus on efficiency and self-reliance.

The Challenge to US AI Dominance

Mark Andreessen, a Silicon Valley entrepreneur, first coined the “Sputnik moment” analogy, highlighting the significant challenge DeepSeek posed to American AI dominance. Prior to DeepSeek, the US was perceived as the sole source of true AI power. However, DeepSeek demonstrated China’s ability to produce a powerful alternative AI system, prompting a re-evaluation of the global AI landscape. Kochan, a researcher at the Brookings Institution, emphasizes that DeepSeek is “the tip of the iceberg” representing years of underlying trends finally coming to fruition. The question has shifted from can China innovate to how China innovates.

US Export Controls and China’s Self-Reliance

The discussion addresses the impact of US export controls, particularly those restricting access to high-end chips from Nvidia. While China had already initiated a national semiconductor program in 2014 (and even earlier efforts), many argue that these controls inadvertently accelerated China’s push for self-sufficiency. Kochan believes the export controls have both motivated and accelerated China’s chip self-reliance efforts, creating a more closed-off, yet rapidly developing, Chinese ecosystem. This is despite the fact that China had been pursuing semiconductor independence for years.

Divergent AI Strategies: US vs. China

A key distinction lies in the differing approaches to AI development between the US and China. The US is heavily invested in achieving Artificial General Intelligence (AGI) and building massive supercomputers. In contrast, China prioritizes integrating AI into everyday products – cars, toys, machines, and industrial automation. This difference is highlighted by the observation that the entire US economy seems to be betting on AGI, while China focuses on practical, immediate applications. This focus on real-world impact is seen by some as a potential advantage for China, as American AGI systems currently lack widespread practical application.

The Next Wave of Innovation: Beyond Large Language Models

Beyond large language models, several areas are poised for significant innovation in China. These include:

• Industrial Robotics: A rapidly growing sector with substantial potential.
• Biotech: A thriving private sector driving advancements.
• Space Industry: A burgeoning private space industry.
• Autonomous Vehicles: Significant progress being made by Chinese startups.

Despite China’s growing technological prowess, concerns remain regarding domestic issues like weak consumption, middle-class job anxiety, and the prevalence of the gig economy. The long-term sustainability of this model is a key question, as China transitions away from a period of high-speed growth.

Technical Deep Dive: The Last Mile of Generative AI

Professor Yang, from Hong Kong Polytechnic University, provides a technical perspective on AI development, dividing it into three stages:

1. Zero to One (2017): The birth of the Transformer model by Google.
2. One to Ten (2022): The launch of ChatGPT-4 and the development of search engines to replace Google’s.
3. Ten to One Hundred (Present): Generative AI integrated into every application.

Professor Yang emphasizes the importance of efficiency and accessibility, particularly in the context of China’s vast consumer and manufacturing base. She highlights the need to move beyond the Silicon Valley approach of centralized models based on general web data. Her team is pioneering a decentralized generative AI approach, training models locally at the data source (e.g., hospitals) to address data privacy and access concerns. This involves training models in situ without moving sensitive data.

FP8 Low-Bit Training: A Key to Efficiency

A crucial innovation discussed is FP8 (Floating Point 8) low-bit training. By using only 8 bits to store numerical data in AI models (compared to the more common 16 bits), FP8 can reduce computing resource and memory storage requirements by approximately 50%. Professor Yang asserts that low-bit training “will be the future,” representing a significant step towards more efficient AI. This demonstrates a shift towards out-engineering rivals on algorithms and architecture when outspending them on hardware is not feasible.

Data Accessibility and the AGI Challenge

Professor Yang argues that the current limitations in achieving AGI are not due to a lack of sophisticated models, but rather a problem of data accessibility. While models like DeepSeek utilize 20 terabytes of data from the general web, high-end applications like cancer treatment require access to petabytes of specialized, private data that cannot be easily obtained from public sources. One petabyte equals 1,000 terabytes.

The Rise of Open Source and Collaboration

Professor Yang advocates for open-source AI development, believing it fosters collaboration and accelerates innovation. She acknowledges the contributions of US companies like Google and OpenAI, stating that “all the technical people worldwide should thank OpenAI.” She also suggests a mutually beneficial relationship, noting China’s strength in production and application, and advocating for collaboration between the two nations. INFEX.AI, Professor Yang’s startup, was valued at $250 million US after its initial funding round, demonstrating the investment in this decentralized AI approach.

Navigating the US-China AI Ecosystem

Kochan notes that third-party countries are adopting a “mix and match” approach, utilizing Nvidia chips where possible and customizing Chinese open-source AI models. He anticipates increasing pressure from both the US and China on these countries to choose sides, creating a difficult position given their reliance on both nations for technology and trade.

Conclusion: A Rewritten Tech Race

The podcast concludes that the rules of the tech race are being rewritten. China is demonstrating a focus on algorithm innovation, efficiency, and practical applications, rather than solely relying on brute-force hardware scaling. The release of advanced Chinese models is expected to continue, and the shift towards a more accessible, tailored AI utility is gaining momentum. China’s “AI Plus” strategy, combined with state and private investment, is defining the “after DeepSeek era.” The future of AI will likely be shaped by the interplay between the US and China, with both nations solidifying their dominance and potentially fostering collaboration despite geopolitical tensions.