Feb 2, 2025 · 12m · allin
AI Czar David Sacks Explains the DeepSeek Freak Out
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the All-In Podcast, David Sacks and his co-hosts analyze the technical breakthroughs, financial realities, and geopolitical implications of Chinese AI firm DeepSeek's new reasoning model. They debunk viral cost myths while exploring how open-source innovation and extreme compute efficiency are shifting value across the global tech landscape.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 96.9% of the talking time here. How this is scored →
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Chamath forcefully dismisses external claims about training costs by arguing that semiconductor analysts and market participants are driven by inherent biases toward Nvidia.
Hardest push from the hosts ▶ 6:15 TK challenging Sacks on white paper stress testingTK interrupts Sacks' cost debunking to point out that DeepSeek's published white paper allows developers to stress-test and verify their cost efficiency claims directly.
Biggest teaching moment ▶ 6:51 Sacks breaking down DeepSeek's true hardware clusterSacks educates the panel on the distinction between single training run costs and overall R&D infrastructure by detailing DeepSeek's 50,000 GPU cluster.
The host holds their own ▶ 9:40 Chamath detailing GRPO and PTX architectural innovationsChamath demonstrates substantial technical expertise by breaking down how DeepSeek bypassed CUDA via PTX assembly programming and implemented GRPO reinforcement learning.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| The DeepSeek Reaction & Drivers of the Story | 0 | 0 | 1 | 0 | David Sacks delivers an uninterrupted monologue synthesizing the global reaction to DeepSeek, breaking down open-source dynamics and comparing base LLMs to reasoning models. Because no hosts speak or interact during this opening section, all host-side scores remain zero. | |
| Debunking the $6 Million Cost Myth & Real Compute Hardware Costs | 5 | 6 | 2 | 4 | Hosts Jason Calacanis, TK, and Chamath Palihapitiya interject with questions regarding training costs, white paper validity, and GPU ownership. Sacks educates the group by citing semiconductor analyst estimates that DeepSeek utilizes a 50,000-hopper GPU cluster worth over a billion dollars. | |
| DeepSeek's Algorithmic Innovations & Bypassing CUDA | 9 | 3 | 3 | 6 | Chamath Palihapitiya takes the lead, reframing the narrative around algorithmic breakthroughs forced by compute constraints. He demonstrates high technical expertise by detailing DeepSeek's use of GRPO over standard PPO reinforcement learning and their bypass of Nvidia CUDA using low-level PTX code. | |
| Market Impact, Value Chain Shift & The Electricity Analogy | 7 | 2 | 1 | 1 | David Friedberg introduces an economic macro perspective, highlighting how model commoditization shifts value upstream or to downstream users, drawing an analogy to the early U.S. electricity market. The tone is entirely collaborative and exploratory. |