After DeepSeek Episode 3

Key Concepts

• Deepseek R1: A Chinese AI reasoning model launched in early 2025, considered a turning point for China’s AI capabilities.
• Low-Bit Training: A method of AI model training that prioritizes efficiency and cost-effectiveness over brute force computational power.
• Self-Sufficiency (in Tech): China’s strategic goal to reduce reliance on foreign technology, particularly in critical areas like semiconductors.
• Hardware Constraints: Limitations imposed by sanctions and access to advanced chips, driving Chinese innovation towards efficiency.
• Commercialization of Tech: A shift in China’s tech strategy to focus on creating commercially successful industries alongside national security applications.
• Centralization (in Tech Policy): Increased control and coordination by the central government in China’s tech sector.
• Edge Processing: Performing AI computations on devices themselves rather than relying on centralized supercomputers.

China’s AI Future: Navigating Constraints and Defining a New Path – An Analysis of the After Deepseek Series

I. The Deepseek Moment & A Shift in Confidence

The launch of Deepseek’s R1 reasoning model in early 2025 marked a pivotal moment for China’s AI industry. Prior to this, China faced significant challenges due to hardware constraints, largely stemming from international sanctions. Deepseek R1 demonstrated that China could compete not through sheer computational power, but through innovation in efficiency and cost-effectiveness. As Tilly Jang of Gaffico Draggonomics explains, “What Deepseek proved is that Chinese AI can maybe find its own path of cost effectiveness…through like a very reasonable cost.” This success significantly boosted confidence within the Chinese AI community – both among developers and users – fostering a belief in the viability of a local AI industry. This shift moved AI development from a purely policy-driven goal to a market-driven reality, fueled by user curiosity and national pride.

II. Innovation Under Constraints: The Rise of Low-Bit Training & Alternative Pathways

The core strategy emerging from these constraints is a focus on “low-bit training,” a technique prioritizing efficiency over raw computational power. This approach allows Chinese developers to build “good enough” models at a reasonable cost, carving out a commercial space for themselves. This isn’t about competing directly with the most powerful US models, but offering viable alternatives. The US strategy of restricting access to advanced chips, while intended to hinder China’s progress, has inadvertently spurred innovation.

Huawei serves as a prime example. Despite sanctions, they have demonstrated the ability to produce advanced chips (7nm) and employ strategies like extending the lifespan of existing chips and aggressive optimization to mitigate hardware limitations. There are even reports of idle chip capacity in some Chinese government AI data centers, suggesting the constraints aren’t absolute roadblocks. This pressure is driving “a new kind of engineering ingenuity,” as the host, Siman, notes. Furthermore, innovation is extending beyond hardware, with a growing focus on “edge processing” – embedding intelligence directly into devices rather than relying solely on centralized supercomputers.

III. Talent Pipeline & The Growing Ecosystem

China possesses a substantial and growing talent pipeline of skilled engineers and ambitious entrepreneurs. This has been a long-term investment, and Deepseek simply brought this capacity “to a boiling point,” making it visible to the rest of the world. Rayma, a China tech analyst and investor, highlights the increasing sophistication of the talent pool and the surprising level of candor and intellectual generosity she’s encountered during tech tours in China. Interestingly, recruitment for these tours is shifting, with increasing interest from American participants following the Deepseek launch.

IV. US-China Divergence: Different Goals, Different Approaches

The series highlights a fundamental divergence in strategic goals between the US and China regarding AI development. The US prioritizes achieving “frontier AI” – recursively improving superintelligence – believing that the first to reach this milestone will achieve global dominance. China, conversely, views AI more as an “assistive technology” and is focused on widespread deployment across the economy.

Rayma frames this as two athletes training for different competitions: “One’s training for a marathon, the other one’s training for like a decathlon.” This difference in approach leads to different priorities and potentially duplicated effort, although both ecosystems stand to benefit from R&D advancements.

V. Structural Factors: Capital & Energy as Key Determinants

Beyond chips and talent, two often-overlooked variables are crucial: capital and energy. The US maintains a significant advantage in capital markets, with access to far greater liquidity for both private and public company funding. Valuations for comparable companies are often 10x higher in the US, enabling greater risk-taking and innovation.

However, China holds a critical edge in energy capacity. A Macquarie Investment Bank report suggests that China would need to add only 1-5% to its recent power generation build rate to meet its AI-related demand over the next five years, compared to a much higher 70% for the US. This energy advantage is vital, as AI training and operation are incredibly power-intensive.

VI. China’s Five-Year Plan: Commercialization & Centralization

Tilly Jang’s analysis of China’s next five-year plan reveals two key shifts in strategy. First, a greater emphasis on commercialization – moving beyond national security applications to create commercially successful tech industries, mirroring the success of the EV and solar sectors. Second, increased centralization – a move towards greater control and coordination by the central government to avoid redundant investments and ensure alignment with national goals. This represents a departure from previous approaches where local governments often pursued independent initiatives.

VII. The Future Landscape: A Splitting of Standards & Supply Chains

The After Deepseek series concludes that the global AI race is not about one side “catching up,” but about the emergence of two fundamentally different AI markets. Ultimately, demand will dictate which models prevail. As Tilly Jang predicts, “people would choose what seal them the best regardless who produced it.”

The long-term advantage will be determined by structural factors – US capital markets versus China’s energy capacity – and a likely splitting of technological standards and supply chains. This creates both challenges and opportunities for China, the US, and the rest of the world.

Notable Quotes:

• Tilly Jang (Gaffico Draggonomics): “What Deepseek proved is that Chinese AI can maybe find its own path of cost effectiveness…through like a very reasonable cost.”
• Rayma (China Tech Analyst & Investor): “I think what we really have to…thank DeepSeek for is that it made legible or visible to the rest of the world what a lot of many Chinese scientists and engineers have been working on for the past decade.”
• Tilly Jang (Gaffico Draggonomics): “people would choose what seal them the best regardless who produced it.”

This analysis provides a detailed and specific overview of the After Deepseek series, preserving the original language and technical precision of the transcript. It focuses on actionable insights and specific details, offering a comprehensive understanding of China’s evolving AI landscape.