CHAPTERS
DeepSeek arrives: a cheap Chinese model rattles Big Tech and markets
Kara introduces DeepSeek as a high-performing, low-cost Chinese AI model that appears to beat major Western models in some benchmarks while using fewer high-end NVIDIA chips. She frames the story as both a Silicon Valley competitiveness shock and a Wall Street shock, citing immediate stock declines.
Scott’s take on the NVIDIA selloff: dramatic headline, modest reset in context
Scott contextualizes NVIDIA’s one-day drop as enormous in dollars but less extreme relative to the company’s recent run-up. He argues the market was primed for an excuse to deflate an overinflated valuation.
Export controls and unintended consequences: the ‘workaround’ argument
Scott suggests NVIDIA may use DeepSeek as evidence that restricting chip sales encourages rivals to innovate around U.S. constraints. The implication is that policy designed to slow competitors can accelerate alternative approaches that undermine incumbents.
Training costs and the brute-force vs efficiency debate
The conversation turns to DeepSeek’s claimed training economics, contrasting the ‘buy more chips’ approach with architectural and systems efficiency. Scott highlights reported numbers that suggest similar capability at far lower training cost.
Second-order shock: energy and nuclear stocks wobble as compute assumptions change
Scott notes that AI’s next bottleneck was expected to be power generation, helping drive up nuclear and energy names. If models can deliver similar performance with less energy, those trades unwind and the broader AI supply-chain thesis gets questioned.
Walmart vs Tiffany AI: bifurcation into cheap and premium model layers
Scott relays analyst thinking that AI will segment into low-cost commoditized models and high-end frontier systems that still require massive compute. DeepSeek may represent the ‘Walmart’ layer while premium players push sophisticated capabilities.
Kara reads Yann LeCun: the real story is open source beating proprietary
Kara cites LeCun’s argument that DeepSeek doesn’t prove China is surpassing the U.S., but that open-source ecosystems are out-innovating closed models. He points to shared tooling and published research enabling rapid iteration and collective progress.
“Talking his own book”: Meta’s incentives and the politics of openness
Scott and Kara acknowledge LeCun’s perspective is self-serving given Meta’s open-source strategy with LLaMA. The discussion touches on how open sourcing affects competitive advantage and fuels debate about whether it undermines U.S. leadership.
Infrastructure isn’t just training: inference at scale and who pays for it
Kara highlights another LeCun point: much of the spending is aimed at inference—running AI for billions of users—rather than training. Even if training gets cheaper, advanced features could push inference costs up, making monetization the key question.
China’s ‘more for less’ playbook: constraint-driven innovation using open source
Scott argues DeepSeek fits China’s long-running pattern of delivering comparable outcomes with fewer resources. With limited access to top-tier chips, the incentive to find software/system workarounds increases—especially leveraging open-source foundations.
Guardrails and risk: open models, fewer restrictions, and misuse concerns
The discussion pivots to safety and governance, with Scott noting Meta-style open models can be run with minimal guardrails. He contrasts more restrictive systems (e.g., Anthropic) with models that may more readily provide sensitive or harmful guidance.
Is this the start of a broader tech correction—or a buying opportunity?
Scott closes by questioning whether DeepSeek is the trigger for a long-awaited major market correction or a temporary narrative-driven dip. He also frames the moment as a free-trade lesson: restrictions may have sped up the very innovation that spooked markets.
