DeepSeek shows alternative models of AI are possible
So long as you don’t ask it about Winnie the Pooh
It was a tale that would have baffled Ayn Rand. A scrappy, young, Chinese Communist Party-backed startup squared off against the monopolistic behemoths of Silicon Valley and delivered one of the biggest upsets in recent AI history. On 20 January 2025, Chinese AI startup DeepSeek released its new model, R1, along with a research paper explaining its innovative training approach. The model was free, open-source, and outperformed U.S. counterparts across a range of metrics.
Without access to Nvidia’s latest chips, DeepSeek was forced to innovate with older hardware, developing more efficient training techniques. In doing so, they upended the long-held belief that more computational power always equals better results. This disruption erased nearly $600 billion from Nvidia’s market cap in a single day of trading.
It’s a Market Correction, Not a Crash
Despite Nvidia’s dramatic market cap drop and a seismic shift in how AI development is viewed, this isn’t the end of profitability in the AI industry. DeepSeek’s success, though groundbreaking, still relied on Nvidia’s older chips—proving their products remain critical to the sector. What’s changed is that DeepSeek’s efficiency has forced competitors to rethink their pricing strategies. Why spend so much on compute when similar performance can be achieved at a fraction of the cost? The result will likely be a more competitive market, but not a less lucrative one. Innovation and efficiency are now just as valuable as computational power, signaling a recalibration of the industry rather than its collapse. Companies that can adapt to these shifting dynamics will continue to thrive.
Compute Still Matters; Export Controls Still Matter
DeepSeek’s achievements don’t completely eliminate the importance of compute power or the challenges posed by export controls. Access to cutting-edge chips and greater compute resources could still improve DeepSeek’s performance, highlighting the limits of efficiency without advanced hardware. Export controls, while not as debilitating as some had feared, remain a bottleneck for Chinese companies, slowing progress in certain areas. Policy analysts have argued that there may not have been enough time for these policies to fully take effect and demonstrate their impact. Even so, DeepSeek’s ingenuity proves that innovation can sometimes leapfrog these constraints, reducing dependency on compute power and opening alternative paths to success. The narrative that China is significantly limited in AI development has been fundamentally shaken.
China Proves the Viability of Its Own AI Development Model
For years, experts claimed China was 12 to 18 months behind the U.S. in AI development. DeepSeek’s R1 has shattered this myth. China’s government plays a dual role in its AI ecosystem: enforcing strict controls on data usage and output while providing financial incentives and infrastructure to support companies. This "dancing in shackles" [dai liaokao wudao] approach has created a uniquely high-pressure environment that forces innovation. Data annotation, for example, is often handled in-house or by partnering with government-backed facilities, ensuring quality control while driving local economic activity. This combination of state control and incentives has proven for the moment that China can compete with and even surpass global players in AI development.
We Are the Winners
DeepSeek’s breakthrough isn’t just a win for China; it’s a win for everyone outside Big AI. While the Chinese Communist Party benefits, so too does anyone seeking access to world-leading models without the monopolistic pricing that tech giants have relied on. DeepSeek has proven that high-performing AI models can be built at a fraction of the cost, lowering barriers to entry for competitors and making development more accessible. The full ripple effects of this disruption remain unclear—how will Big AI respond? What new directions will companies explore with downstream apps? One thing is certain: DeepSeek has loosened Silicon Valley’s stranglehold on the industry, paving the way for more innovation and alternative approaches to AI development.
But, don’t ask it about Tiananmen Square
This does not mean that the best approach to AI development is the one sanctioned by the Chinese Communist Party. Political control and censorship over AI means that DeepSeek’s model refuses to answer questions of political significance to China such as Taiwan, internal rebellions and the machination of the Communist Party. When prompted, the model currently responds with: ‘Sorry, that’s beyond my current scope. Let’s talk about something else.’ Free expression and the open discussion of political beliefs should be baked into future open-source models.