The end of AI’s artificial scarcity

The release of DeepSeek R1 shook the financial markets, triggering a $1 trillion sell-off in tech stocks, according to Business Insider. The DeepSeek AI model, developed outside of Silicon Valley and China’s tech giants, showed that next-generation AI no longer required billions of dollars—or an insatiable supply of electricity—to train or operate. Investors panicked, and with good reason.

The market’s reaction reflects a deeper realization: The perceived scarcity of AI was never real. The notion that access to top-tier models would remain tightly controlled has fueled the AI investment boom, but that assumption no longer holds. DeepSeek R1 rivals OpenAI's GPT-4o1 in performance, yet its API is much cheaper- $0.55 per million input tokens compared to OpenAI's $15 for the same number of tokens. Imagine paying $15 for a million grains of rice, when you could get the same amount of rice for $0.55. DeepSeek can run locally on a workstation. And unlike other large AI systems, it has no rate limits, meaning that they aim to serve every user request.

The cracks were already forming

For the past two years, investors and analysts have treated AI as a winner-take-all game, believing that training frontier models requires billions of dollars, massive compute clusters, and high-end Nvidia GPUs restricted by U.S. export controls. AI was expected to remain centralized, with dominance concentrated in a few companies such as OpenAI, Anthropic, and Meta.

DeepSeek R1 is forcing a rethink of these beliefs.

There is some skepticism of DeepSeek’s claim of spending only $6 million training cost as reported by the New York Times. This figure does not seem to include the cost of prior research, extensive model refinement, and access to large-scale computing infrastructure.

More importantly, DeepSeek almost certainly benefited from exposure to American AI models, probably through distillation techniques that allow smaller models to absorb knowledge from larger ones like GPT-4o1.

Even if the true cost was much higher than $6 million, one thing is clear: AI training and inference is becoming dramatically more efficient. The idea that cutting-edge AI always requires massive capital investment is no longer valid.

This is not an industry collapse, but a correction. AI remains a transformative technology, but the idea that only a few American companies can control its future no longer holds.

The moat is shrinking—but not gone

DeepSeek R1 exposes weaknesses in the current AI business model, but it does not automatically dethrone AI giants like OpenAI. If anything, the situation is reminiscent of the early search engine market. In the beginning, there were many alternatives to Google, but Google continues to dominate because even a small performance advantage is enough to cement user loyalty.

In AI, these small differences matter even more. Users—whether enterprise customers or individual developers—are not just looking for a model that performs well on benchmarks. They want reliability, consistency, ecosystem integration, and long-term support. AI is not just a product; it is a tool that users need to trust.

OpenAI has a formidable technical advantage that continues to grow, which translates into user confidence. The company’s integration with Microsoft provides a built-in distribution channel through Windows, Azure, and enterprise software. While DeepSeek R1 rivals GPT-4o1 in some raw capabilities, OpenAI's brand, infrastructure, and partnerships make it the default choice for most users.

AI is becoming increasingly specialized. While OpenAI may remain the leader in general-purpose AI, other models are likely to dominate in specific domains—whether mathematics, programming, music composition, movie production, or medicine. DeepSeek or future competitors do not need to outperform OpenAI across the board to gain market share in targeted applications.

AI is the great leveler

For much of human history, power, wealth, and opportunity were tied to physical strength. The fastest runner, the strongest warrior, or the most resilient worker were the people who frequently determined the course of history. But machines wiped out those advantages. A kung-fu master is no match for a gun. A strong worker is no match for a forklift. What was once a defining human trait became secondary to technology.

AI is on the verge of doing the same for intelligence.

For centuries, cognitive ability has been a key determinant of success. Some people were naturally better at solving problems, reasoning, and absorbing knowledge. These differences have shaped corporate hierarchies, scientific advances, and financial markets.

AI changes this equation. Just as machines have compensated for physical differences, AI will compensate for cognitive differences. Complex tasks that once required years of training and elite analytical skills—discovering drugs, designing software, diagnosing diseases—are now within reach of anyone with a good AI model.

This does not mean that human intelligence will become obsolete, just as physical strength did not disappear when machines took over work. But intelligence will not be as important a determinant of success as it has been in the past. The advantage will shift from raw cognitive ability to knowing how to use intelligence wisely and for meaningful purposes.

The business model no longer holds

The biggest threat to OpenAI, Anthropic, and Meta is not DeepSeek R1 itself, but what it represents. These companies built their businesses on the assumption that AI inference would remain expensive and cloud-based, with users paying a premium for API access.

But what happens when high-quality models can run locally on consumer hardware?

DeepSeek R1 shows that AI is moving away from the cloud. If today’s best models can run on high-end workstations, similar models will soon run seamlessly on smartphones. This transition would undermine the cloud-based AI access model, just as local processing disrupted cloud gaming.

Even if AI services remain in demand, margins will erode. Hosting AI models is quickly becoming a commodity business, and premium pricing for general-purpose AI is unlikely to be sustainable.

The policy failure no one wants to discuss

Beyond its business implications, DeepSeek R1 exposes fundamental flaws in U.S. efforts to contain China’s AI progress.

For the past two years, the U.S. has restricted exports of Nvidia’s most powerful chips, betting that without access to cutting-edge hardware, China’s AI development would slow. DeepSeek R1 suggests that the bet isn’t paying off.

The model almost certainly benefited from distillation techniques, learning from American AI systems without direct access to restricted GPUs. And as China’s own AI models advance—fed in part by their exposure to Western systems—simply blocking hardware sales won’t be enough to keep China from catching up.

This raises difficult questions. If cutting off Nvidia chips isn’t working, what should U.S. policy focus on instead? Should restrictions target model architectures and training methods, not just hardware? 

Is containment already impossible?

For now, Washington doesn’t have an answer.

AI isn’t collapsing—It’s growing up

The market’s reaction to DeepSeek R1 was dramatic, but it was also inevitable. Investors had been pricing AI as if the scarcity would last forever. That was never realistic.

None of this means that AI is slowing down. Quite the opposite—AI is accelerating, just in a way that financial markets have not priced in.

The competitive landscape is shifting. AI is getting cheaper, more efficient, and more accessible. The moat around large-scale AI training is shrinking, and the real power now lies in applications, distribution, and deployment.

OpenAI remains dominant, just as Google still leads search despite credible alternatives. But as AI becomes more specialized, new leaders will emerge in targeted areas. The days of assuming that one or two companies will control the entire industry are over.

AI isn’t collapsing. It’s entering its next phase.