Khan Capital | January 2025
Key Takeaways
- DeepSeek’s R1 model matched OpenAI’s frontier reasoning model at roughly 90% lower cost, triggering a $589 billion single-day loss for Nvidia, the largest in US stock market history.
- The sell-off swept across the entire AI infrastructure chain, with the Nasdaq losing approximately $1 trillion as chipmakers, power companies, and equipment manufacturers fell sharply.
- The reported $5.6 million training cost is likely understated, but the efficiency gains are real and may accelerate AI adoption through Jevons’ Paradox, increasing total demand even as unit costs fall.
- The locus of AI value creation is shifting from infrastructure providers toward application developers, a transition that may favour software companies over hardware-dependent plays.
- US export controls did not prevent DeepSeek’s achievement, raising strategic questions about the efficacy of technology restrictions as a tool for maintaining competitive advantage.
On Monday 27 January 2025, Nvidia lost $589 billion in market capitalisation in a single trading session: the largest single-day destruction of value for any stock in the history of the US equity market. The catalyst was not an earnings miss, a regulatory action, or a macroeconomic data release. It was the emergence of a one-year-old Chinese startup called DeepSeek, which had released an AI model apparently matching the performance of the best Western systems at a fraction of the cost.
The implications reverberated far beyond Nvidia. The Nasdaq Composite fell 3.1%. The S&P 500 dropped 1.5%. Broadcom plunged 17%. Constellation Energy, which had recently signed a deal to restart Three Mile Island to power AI data centres, collapsed 20%. The entire AI infrastructure complex was caught in a violent reassessment of the assumptions that had underpinned the most powerful investment theme of the past two years.
| Stock / Index | 27 Jan Move | Key Concern |
|---|---|---|
| Nvidia (NVDA) | -16.9% (-$589bn) | GPU demand thesis questioned |
| Broadcom (AVGO) | -17% (-$200bn) | AI chip demand repricing |
| Constellation Energy | -20% | AI data centre power demand questioned |
| Siemens Energy / GE Vernova | -20% / -21% | Power equipment orders at risk |
| Nasdaq Composite | -3.1% (~$1tn lost) | Broad AI infrastructure repricing |
| S&P 500 | -1.5% | Concentrated AI exposure |
| Dow Jones | +0.7% (+289 pts) | Rotation into healthcare, staples |
What Happened: The Model That Shook Silicon Valley
DeepSeek’s story begins in Hangzhou, where the startup was founded by Liang Wenfeng, a quantitative hedge fund manager. In late December 2024, DeepSeek released V3, an open-source large language model that the company claimed was developed in just two months at a computing cost of approximately $5.6 million. For context, OpenAI’s GPT-4 reportedly cost over $100 million to train, and Meta had just announced plans to spend upward of $65 billion on AI development in 2025 alone.
Then, on 20 January 2025, DeepSeek released R1, a reasoning model designed to compete with OpenAI’s cutting-edge o1. In third-party benchmarks, R1 performed on par with o1 across a range of tasks, yet it was nearly twice as fast and approximately 90% cheaper to operate. The model was open-source, free to use, and quickly became the most downloaded app on Apple’s iOS store, displacing ChatGPT from the top position.
The technical achievement was striking. DeepSeek had built R1 using Nvidia’s H800 GPUs, a deliberately restricted chip that Nvidia was permitted to export to China under US export controls. The entire thesis of Western AI dominance rested on two pillars: superior access to the most advanced semiconductors, and the ability to deploy massive capital to train frontier models. DeepSeek appeared to have circumvented both constraints. Marc Andreessen called it “one of the most amazing and impressive breakthroughs I’ve ever seen,” dubbing it “AI’s Sputnik moment.”
The Market Reaction: Repricing the AI Capex Thesis
The sell-off on 27 January was not simply a reaction to competitive threat; it was a fundamental repricing of the capital expenditure assumptions underpinning the entire AI investment chain. The bull case rested on a simple cascading proposition: frontier AI requires enormous computing power, which requires enormous quantities of Nvidia’s most advanced GPUs, which requires massive data centre buildouts, which requires vast amounts of electricity. DeepSeek’s achievement threatened to invert this logic.
The market punished every link in the chain. Among the hyperscalers, Microsoft fell 2%, Alphabet 4%, and Meta nearly 3%. The carnage was most severe in AI-adjacent infrastructure: Constellation Energy fell 20%, Vistra Energy and NRG Energy dropped sharply, and Siemens Energy plunged 20%.
Notably, the Dow Jones Industrial Average actually rose 289 points on the day. The divergence underscored the rotational nature of the sell-off: this was not a broad market panic but a targeted reassessment of the AI infrastructure premium.
What the Market Is Misunderstanding
The $5.6 million figure is misleading. DeepSeek’s reported training cost covered only the final training run’s computing costs. It excluded extensive prior research, architectural experimentation, and data preparation. As Bernstein analyst Stacy Rasgon noted, the models “look fantastic, but we don’t think they are miracles.”
Efficiency gains may increase demand, not reduce it. There is a well-established economic principle, Jevons’ Paradox, which holds that improvements in the efficiency of resource use often lead to increased total consumption of that resource. If AI becomes dramatically cheaper to deploy, the range of viable applications expands enormously. Jensen Huang later argued that investors fundamentally misread the situation, emphasising that inference and post-training still require massive computing power.
The export control question is more complex than it appears. The controls did not prevent DeepSeek from training a competitive model, but they likely constrained what DeepSeek could achieve. Without access to Nvidia’s most advanced H100 chips, DeepSeek was forced to innovate around hardware limitations.
Distillation concerns add uncertainty. OpenAI confirmed it had seen “some evidence” that DeepSeek may have used distillation, a technique where a smaller model learns from the outputs of a larger, more capable model. This raises questions about the reproducibility of the achievement.
Structural Interpretation: What DeepSeek Means for the AI Trade
The DeepSeek shock should be understood not as the death of the AI investment thesis but as its maturation. It introduces three structural shifts: the locus of AI value creation is moving from infrastructure to applications; the geographic concentration of AI capability is less absolute than assumed; and the capital efficiency of AI development is improving faster than expected.
Implications for Investors
Nvidia and the semiconductor complex face a valuation reset but not an existential threat. The key metric to watch is hyperscaler capex guidance in upcoming earnings reports.
AI application and software companies are the potential beneficiaries. Cheaper inference costs mean higher margins for companies deploying AI in their products.
Power and energy infrastructure names face the most acute near-term risk as the thesis that AI data centres would require enormous new power generation capacity is questioned.
Chinese technology stocks may attract renewed interest as DeepSeek’s achievement challenges the narrative that Chinese AI companies are permanently disadvantaged by export controls.
Concentration risk in the AI trade is now undeniable. The Magnificent Seven’s dominance of S&P 500 returns was built substantially on the AI capex narrative.
Conclusion
The DeepSeek shock is not the end of the AI trade; it is the end of the easy AI trade. The assumption that frontier AI requires limitless capital, that only the largest Western hyperscalers can compete, and that the entire infrastructure chain deserves premium multiples has been challenged by a small Chinese startup with $5.6 million in reported computing costs and a model that topped the iOS download charts in days.
Sources: CNBC, CNN, NBC News, Fortune, Fast Company
Related Reading
DeepSeek disrupted the AI investment thesis built on the foundations covered in Nvidia’s AI Supercycle and The AI Bubble Debate. For how AI spending continued despite the shock, see AI Capex Boom: $600 Billion and Counting. For the broader tech volatility that followed, see Tech Volatility in 2026. The trade rivalry that preceded the tech competition is traced through the Phase 1 trade deal. The Huawei ban that catalysed China’s drive for tech self-sufficiency is covered in the May 2019 trade war escalation.
Leave a Reply