How China Could Dominate U.S. AI | Dr. Michael Power on Open Source and "The Three Assassins" of Moore's Law

Episode Summary

Executive Summary: Michael Power argues China’s open-source, low-cost AI stack is structurally better positioned than the U.S. closed, capital-intensive model. He says U.S. AI faces limits from physics, materials, economics, and power, while Chinese models and chips are becoming more efficient and widely adopted, threatening U.S. margins, valuations, and the current AI investment boom.

Main Topics: Chinese open-source AI as the winning architecture (Priority: 5/5): Power argues China’s AI model is based on open source/open weight and is effectively free, making it more flexible, widely adopted, and better suited for global diffusion than closed U.S. models. U.S. AI as a capital-intensive, margin-dependent model (Priority: 5/5): He says the U.S. stack depends on monetization, huge capex, and high margins from chips, clouds, and models, which makes it vulnerable if AI commoditizes. NVIDIA and semiconductor vulnerability (Priority: 5/5): Power argues NVIDIA’s dominance is challenged by hyperscalers building custom chips, Chinese diversification, and architecture changes that reduce reliance on ever-smaller chips and CUDA. DeepSeek as the catalyst for efficiency breakthroughs (Priority: 5/5): He highlights DeepSeek’s recent technical advances as evidence that Chinese models can dramatically reduce compute/memory requirements and erode the need for expensive U.S. hardware. Physics, materials science, and economics as constraints (Priority: 4/5): Power frames Moore’s Law and chip miniaturization as hitting hard limits: electron leakage, material degradation, and uneconomic costs at the smallest nodes. Power, data centers, and the AI energy bottleneck (Priority: 4/5): He says U.S. AI is constrained by electricity and data-center power needs, while China has invested heavily enough in energy to avoid similar bottlenecks. Investment implications and winners/losers (Priority: 5/5): He is cautious on U.S. AI equities like NVIDIA, Oracle, Microsoft, OpenAI-linked assets, while favoring relatively stronger names like Google and some Chinese platforms such as Alibaba, Tencent, and Baidu.

Key Arguments: China’s AI is built as a utility-like, open-source system, so value accrues in applications rather than model monetization; this encourages adoption and global spread. U.S. AI is a closed, expensive ecosystem that needs to recoup massive capex through high margins, but AI software is likely to commoditize and compress those margins. NVIDIA’s moat is weakening because hyperscalers are developing in-house chips (TPU, Trainium) and software alternatives that reduce dependence on CUDA and high-end GPUs. DeepSeek’s latest technical work suggests training/inference can be done with far less compute and memory, undermining the assumption that U.S.-style scale spending is necessary. The bottlenecks in leading-edge chips are physical: at very small nodes, electrons leak, materials degrade, heat rises, and the economics worsen. China is moving away from a 3nm arms race and toward “cognitive towers” and 14–18nm systems that can deliver adequate performance at lower cost. China’s broader industrial base gives AI more real-world deployment points, especially in factories, energy systems, robotics, drones, and logistics. U.S. AI demand is being inflated by unprofitable customers and a prisoner’s-dilemma capex race, which may end in a bust even if individual firms survive. Google appears relatively well positioned because of its TPU strategy, Android distribution, Gemini model, and possible alignment with Apple. OpenAI, Oracle, and other standalone model/data-center plays look more exposed because their economics depend on continued hypergrowth and heavy external funding. Chinese AI’s openness means foreign users can adopt it without necessarily sharing data with China, and offline use further reduces data-transfer concerns. The likely future is a bifurcated world: reinforced U.S. islands of profitability, and a much broader global adoption of Chinese AI tools because they are free and customizable.

Data Points: Estimated value of U.S. AI ecosystem: ~$15 trillion - Power’s rough estimate for publicly traded securities plus VC-funded companies in the U.S. AI ecosystem Estimated share of Chinese models using user fees: <5% - He says only a small minority of Chinese AI models charge end users, usually for specialized use cases DeepSeek memory requirement reduction: from 100% to 7% - He says the new architecture can produce the same output with roughly 7% of the prior memory usage Training efficiency improvement: 15x - He describes DeepSeek’s training-side architecture as increasing effective power of a small chip by roughly 15 times DeepSeek R1 spending claim: $1.5 billion (disputed by speaker) - A cited estimate for DeepSeek’s first R1 model, which Power says he doubts but uses as a high-end comparison OpenAI Stargate capital contribution: $19 billion - Mentioned as OpenAI’s own equity contribution toward Stargate, contrasting with DeepSeek’s lower costs Chinese manufacturing share by 2030: 45% of world manufacturing - Power cites this as part of China’s industrial base supporting AI deployment U.S. manufacturing share by 2030: 10% - Used to illustrate China’s larger industrial footprint and more deployment venues for AI chips Robot installations in China last year: More than the rest of the world combined - Cited as evidence of China’s rapid smart-factory adoption Top supercomputers using Linux: 100 out of 100 - Used to support the case that open-source infrastructure tends to dominate performance-critical systems NVIDIA chip size progression: 30nm to 50nm to 7nm to 3nm - Used to illustrate how far miniaturization has advanced and why it is nearing limits Energy capacity increase needed by OpenAI: 125x over 8 years - Used as an example of unsustainable compounding in U.S. AI scaling assumptions Hyperscaler capex pattern: Higher in 2025 than 2024; higher again in 2026 - Described as a prisoner’s-dilemma / Red Queen race where no firm can stop spending

Pivotal Quotes: "The essence of the Chinese AI approach is that it’s free." — Michael Power: He explains why open-source/open-weight distribution gives China structural advantage "We are seeing the diseconomies of physics, the diseconomies of chemistry, and the diseconomies of economics." — Michael Power: His summary of why leading-edge chip scaling is reaching practical limits "The bubble is technological, but the financial bubble is a symptom of that technological bubble." — Michael Power: He argues investment excess follows from underlying technical constraints and commoditization

Implications: If Power is right, AI returns will compress, U.S. chip/model leaders face valuation risk, and global adoption may shift toward free Chinese architectures. Investors should distinguish real winners from capex-heavy momentum stories and expect a more fragmented AI world.