AI Czar David Sacks Explains the DeepSeek Freak Out

Summary

David Sacks gives his synthesis of the DeepSeek reaction, explaining why an open source reasoning model from a Chinese lab became a global story tied to a one day trillion dollar market cap decline. He argues the six million dollar training figure should be debunked because it covers only the final training run and ignores cumulative R&D and a compute cluster of roughly 50,000 Nvidia Hopper GPUs that cost well over a billion dollars. Chamath Palihapitiya focuses on the engineering, noting DeepSeek replaced PPO with the lighter GRPO algorithm and compiled around CUDA using PTX, which he treats as evidence that Nvidia software lock-in can be bypassed under constraint. David Friedberg closes by arguing that cheaper, faster, commoditizing models push value away from the model layer toward applications, users and the broader economy.

• Sacks says the story landed hard because it combined US versus China competition with the open source versus closed source debate.
• R1 compressed industry estimates of China AI lag from six to twelve months down to three to six months.
• The cited six million dollar figure covers only the final training run, not fully loaded R&D or hardware.
• American labs final training runs were already in the tens of millions of dollars nine to ten months earlier.
• Dylan Patel estimates roughly 50,000 Nvidia Hoppers, 10,000 H100s, 10,000 H800s and 30,000 H20s, implying over a billion dollars of compute.
• Those chips were accumulated before and around export controls through the founder affiliated hedge fund.
• Chamath points to GRPO replacing PPO and to PTX bypassing CUDA as constraint driven innovations that also expose Nvidia lock-in risk.
• Friedberg expects commoditizing models to shift value downstream to applications, users and the wider economy.