AI Czar David Sacks Explains the DeepSeek Freak Out

DeepSeek’s R1 materially narrowed perceptions of China’s AI lag, because it became the first widely released reasoning model after OpenAI’s o1. The speakers said industry estimates shifted from China trailing by roughly 6–12 months to perhaps 3–6 months after R1 matched a model OpenAI had released only months earlier.

The viral claim that DeepSeek built R1 for just $6 million badly confuses one training run with the total cost of creating the system. Even accepting the figure, the comparable American final-training costs were described as tens of millions, while billion-dollar estimates often include hardware, research, failed experiments and years of development.

DeepSeek was not a tiny laboratory working with almost no computing power. Semiconductor analyst Dylan Patel was cited estimating access to roughly 50,000 Nvidia Hopper GPUs—about 10,000 H100s, 10,000 H800s and 30,000 H20s—with the cluster’s hardware value alone estimated above $1 billion.

DeepSeek’s most consequential achievement may be technical efficiency rather than simply cheaper replication of Western models. The discussion highlighted its GRPO reinforcement-learning approach, which uses less memory than conventional PPO, and low-level PTX programming that bypassed much of Nvidia’s CUDA abstraction to extract more from constrained hardware.

Restrictions on Chinese access to top-end chips may have unintentionally encouraged engineering choices Western labs had little reason to pursue. The speakers argued that abundant capital and compute can reduce pressure to optimize, while DeepSeek’s constraints forced approaches that better-funded teams had not discovered or prioritized.

DeepSeek’s founder had accumulated substantial Nvidia hardware through his AI-driven hedge fund before later export restrictions tightened. The speakers stressed that known H100 and H800 holdings did not by themselves imply sanctions violations, while also acknowledging outsiders cannot reliably determine DeepSeek’s complete inventory.

If powerful models become rapidly cheaper and interchangeable, the durable economic value may migrate away from the model makers themselves. The discussion compared AI with electricity: enormous productivity gains could accrue chiefly to applications, users and businesses throughout the economy rather than to the companies producing the underlying intelligence.