New Lex Fridman Insight: DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters
Sent May 30, 2026
Key Insights
- DeepSeek's R1 model is 27 times cheaper than OpenAI's o1 model, costing $2 per million tokens.
- NVIDIA's H20 chip, despite having lower FLOPS than the H100, performs better on reasoning tasks due to higher memory bandwidth.
- China's AI development is bolstered by a trillion RMB subsidy, but US export controls limit their access to cutting-edge GPUs.
- OpenAI's joint venture with Oracle involves a $100 billion infrastructure investment, but funding remains uncertain.
- DeepSeek's mixture of experts model activates only 37 billion of its 600 billion parameters, reducing compute costs.
How the conversation moved
The episode begins with a discussion on DeepSeek's innovative AI models, particularly focusing on the cost efficiency and technical advancements of the DeepSeek R1 model compared to OpenAI's offerings. Nathan Lambert outlines how the mixture of experts model allows DeepSeek to reduce compute costs significantly by activating only a subset of parameters during training and inference. This approach not only makes DeepSeek's models more accessible but also positions them as a competitive alternative in the AI landscape dominated by OpenAI.
Dylan Patel then shifts the conversation to NVIDIA's H20 chip, emphasizing its superior performance in reasoning tasks despite having lower FLOPS than the H100. The H20's architecture, which prioritizes memory bandwidth over sheer computational power, exemplifies how specific design choices can enhance performance for particular tasks. This insight challenges the conventional focus on FLOPS as the primary metric for evaluating AI hardware, suggesting that memory architecture can be equally, if not more, important.
The conversation takes a geopolitical turn as the speakers discuss the implications of US export controls on China's AI capabilities. Patel argues that while these controls aim to maintain US technological dominance, they could inadvertently encourage China to develop independent technological capabilities. This tension highlights the complex interplay between technological advancement and geopolitical strategy, with significant implications for global AI development.
The episode concludes with an exploration of OpenAI's financial challenges, particularly in relation to its joint venture with Oracle for AI infrastructure. Despite the ambitious $100 billion investment plan, funding remains uncertain, raising questions about OpenAI's ability to scale its operations and maintain its competitive edge. This financial uncertainty underscores the broader challenges faced by AI companies in balancing innovation with the practicalities of funding and infrastructure development.
Surprising moments
In-depth
DeepSeek's Competitive Edge
- DeepSeek's R1 model is significantly cheaper than OpenAI's o1 model.
- The mixture of experts model reduces compute costs by activating only a subset of parameters.
- DeepSeek's open weights and permissive licensing offer flexibility in AI development.
NVIDIA's Chip Architecture
- NVIDIA's H20 chip excels in reasoning tasks due to its memory bandwidth.
- The H20's architecture highlights the importance of memory over sheer FLOPS for specific tasks.
US-China AI Geopolitics
- China's AI development is supported by significant subsidies but hampered by US export controls.
- These controls could push China to develop its own technological capabilities independently.
OpenAI's Infrastructure Investment
- OpenAI's joint venture with Oracle involves a massive infrastructure investment.
- Funding for this venture is uncertain, impacting OpenAI's future scalability.
Notable Quotes
DeepSeek is doing fantastic work for disseminating understanding of AI.
Still open
- Nathan Lambert questioned whether the cost of deploying AGI capabilities at scale would delay their widespread implementation.