Why DeepSeek Is Rattling Markets — If Not the Future of AI
In an era when new AI research and technical breakthroughs are constantly upending assumptions and business models, it’s truly extraordinary when something happens that stops everyone in their tracks.
But that’s exactly what is happening. DeepSeek-R1 — a new open source foundation model from the AI research unit of High-Flyer, a China-based hedge-fund manager with $8 billion in assets that is known for leveraging AI to trade — is rivaling the performance of top models and causing a big dip in chip and Magnificent Seven stocks, particularly NVIDIA.
Why Is DeepSeek Unique?
Debuting on January 20th, DeepSeek-R1 is now in the top five on leaderboards like Chatbot (LMSYS) Arena and performs relatively well on benchmarks like the newly-released Humanity’s Last Exam (HLE) and at chain-of-thought reasoning. What’s remarkable is that it does all of this at a modest training cost while leveraging H100 Nvidia Chips, while industry leaders spend orders of magnitude more building AI with NVIDIA’s latest Blackwell chips.
DeepSeek also leverages several innovative techniques — like skipping supervised fine-tuning in favor of pure reinforcement learning and using “Aha moments” as pivot tokens — and is one of the only models this powerful with an MIT license, one of the most permissive and flexible open source licenses.
What Are the Longer-Term Implications of DeepSeek?
Open Source Parity
As Meta’s Yann LeCun notes, DeepSeek’s performance is as much a breakthrough in years’ worth of open research and open source technology as it is a marvel for any one company or country. While many Fortune 500 companies and startups dealing with sensitive data may not ever leverage DeepSeek’s API for various reasons, its debut points toward a future where open models can power more use cases affordably.
Expect More
Given that DeepSeek is open source, it’s reasonable to expect others to build upon this and match it — and quickly. Little stands in the way of other AI and technology giants from reverse-engineering DeepSeek’s chain-of-thought and other techniques — and then using their own brand, distribution, customer support and scale to be the lowest-cost provider. In the long run, this competition is probably good for open source and private offerings alike.
This May Not Be As Disruptive To Chipmakers As Many Think
With NVIDIA losing over $400 billion in market cap in under 24 hours, it’s clear that many are concerned about how important chipmakers are to the AI revolution since performant AI can clearly run on cheaper, less powerful hardware. While predicting the future here is potentially fraught, one important thing to keep in mind is that models — like software — will continue to get more efficient even as the hardware gets more capable, similar to the PC. In that sense, DeepSeek’s gain isn’t really NVIDIA’s loss in the same way better software does not necessarily slow down chipmakers.
AI Investment Will Not Likely Slow Down
Will hyperscalers or big model builders like OpenAI or Anthropic slow down investment to get costs down? Not likely in this environment. Game theorying it out, the costs of not moving are greater than the costs of moving. What’s more likely is a doubling down on all of the above, with new methodologies and techniques of DeepSeek combined with other emerging techniques from academia enriching existing efforts.
It’s also reasonable to expect that we forget all about this in a month when it’s old hat and a similar leap forward emerges — especially at scale with the Blackwell chip and an increase in data center capacity and energy production.
AI Diversification Matters
AI is about much more than chipmakers and data infrastructure. 62.6% of Swift’s recently-launched AI Index — which leverages a fine-tuned LLM to analyze earnings and regulatory filings as well as data from a variety of sources to determine levels of AI investment and activity — is composed of companies building or betting on AI-powered applications, from Recursion Pharmaceuticals leveraging AI in drug discovery to Lemonade’s AI-powered digital insurance.
Conclusion
With the S&P 500 equity risk premium recently turning negative for the first time since 2002, ultimately some market correction amid an AI boom is inevitable. One thing is clear, however: enterprise adoption of AI is likely to accelerate, propelled by innovation and lower costs. The frenzied reaction to DeepSeek and what it will spur is just an exciting opening chapter to the next phase.
