← Home
The All-In Podcast #200 · January 31, 2025 · 100m

DeepSeek Shockwave, AI Costs Plummeting, China's AI Strategy

China's DeepSeek releases an AI model comparable to GPT-4 at a fraction of the training cost. The besties debate whether this invalidates the US AI infrastructure investment thesis, whether China's AI strategy is working, and whether compute moats are real.

Canon

US export controls denied China access to NVIDIA's best GPUs. Instead of stopping Chinese AI, this constraint forced DeepSeek to innovate algorithmically — producing more efficient models. The restriction became the advantage.

Highlights

DeepSeek proves that compute moats are weaker than assumed — algorithmic innovation can substitute for brute-force compute
DeepSeek achieved GPT-4-level performance with reportedly $6M in training costs versus OpenAI's hundreds of millions. If true, the entire thesis of 'whoever has the most GPUs wins' is wrong — algorithmic efficiency can substitute for compute scale.