Interview, Fireside Chat, Panel
AI Czar David Sacks Explains the DeepSeek Freak Out
- Industry perceptions of China's AI lead relative to OpenAI's O1 are expected to shift from a 6-to-12-month gap to a 3-to-6-month gap, with O1 having been released approximately four months prior to the analysis.
- The narrative surrounding DeepSeek is predicted to be driven by international sentiment opposing U.S. dominance and supporters of open-source models seeking to undercut OpenAI's pricing.
- Claims that DeepSeek spent only $6 million on training are anticipated to be debunked as misleading comparisons against full American costs, with analysts citing a cluster of roughly 50,000 Hopper-class GPUs valued at over $1 billion.
- This cluster composition is specified as approximately 10,000 H100s, 10,800 H100s (or similar variants), and 30,000 H20s, raising potential concerns regarding compliance with 2022 and 2023 export bans on specific chip models.
- Potential regulatory risks include the uncertainty of whether DeepSeek obtained additional chips beyond the 50,000-unit estimate in violation of export restrictions.
- Innovations such as custom algorithms (potentially GRPO) and direct hardware control (PTX) are projected to be driven by compute constraints rather than resource surplus, prompting questions on whether Western entities can adopt similar constraint-driven innovation.
- If model performance improves while costs decline, the industry may face accelerated commoditization, shifting value creation from the model layer to users or the broader economy similar to the electricity sector.
- Analyst sentiment regarding training costs may be influenced by bias, with semiconductor experts potentially favoring high cost narratives to support bullish views on Nvidia, while disruptors leverage low-cost claims.
- Historical cost data, such as Anthropic's final training run costing in the tens of millions approximately nine to 10 months ago, is noted to contextualize current expenditure debates.