Key ideas in the global race for AI supremacy

Explosive new entrants like DeepSeek lower the barriers to participation in AI, forcing a rethink of how and where value can be created.

When the Chinese firm DeepSeek unveiled its new large language model (LLM) in January 2025, it caught U.S.-based tech giants by surprise. The high performance and relatively low training costs were in strong contrast to the OpenAI model, which has required billions of dollars to create. Tech stocks went into a tailspin. Nvidia, the California-based company producing graphics processing units (GPUs) — essential for training neural networks — was particularly hard hit, with stock prices plunging 17% in a single day. Although stock prices have since largely recovered, nervousness around DeepSeek lingers.

In a session at IESE Business School, professor Sampsa Samila addressed some of the myths and concerns surrounding DeepSeek, analyzing the potential market disruption and geopolitical impact. Is the hype justified when other models, such as Elon Musk’s Grok 3, are also carving out new spaces in the market?

We’ll have to wait and see what happens with DeepSeek, but there are three main implications of its arrival that go beyond the technology itself.

1. AI costs are falling across the board

What DeepSeek did wasn’t so innovative, as costs for all LLMs have been dropping fast. Nevertheless, the precipitous drop sent OpenAI and Grok scrambling.

DeepSeek claims to have trained its base model for just $5.6 million — a staggering 95% cost reduction compared with the hundreds of millions or even billions spent by U.S. firms. But is it true?

Samila cautioned that DeepSeek’s reported costs only cover the final training run, omitting salaries, capital expenditures and prior research, meaning the true price tag remains unclear.

Up until now, the perceived high costs associated with AI development have prevented competitors from following suit. But DeepSeek has shown that when costs plummet, the barrier to replication disappears. This throws traditional business models into turmoil.

It’s also a reminder that necessity is the mother of invention. Strict import restrictions on GPUs drove Chinese developers to efficiencies that U.S. companies, flush with funding, haven’t been forced to pursue.

2. Research remains fundamental — and increasingly geopolitical

China has long been seen as an imitator rather than an innovator, but DeepSeek founder Liang Wenfeng believes it’s time for China to lead. His DeepSeek-R1-Zero model exhibits conceptual reasoning skills and even taught itself to reason without supervised fine-tuning — removing human intervention and pushing AI closer to artificial general intelligence or AGI. This progress comes despite intense lobbying from major AI firms to limit competition and psychologically intimidate new entrants.

DeepSeek’s success signals that China is rapidly gaining ground in the AI space, thanks to investment in research. According to a Nature ranking, 7 of the top 10 research institutions are based in China. The Emerging Technology Observatory charts the growing number of English-language artificial intelligence articles that have at least one confirmed author from a Chinese organization.

The Global AI Talent Tracker shows China nearly matching the U.S. in producing elite AI researchers, with increasing numbers choosing to remain in or return to China. It appears AI is developing along increasingly nationalistic lines, which begs the question: Does it matter who wins the race to AGI?

Many believe it does, with concerns ranging from AI-driven weaponry to propaganda. Vinod Khosla, one of the big Silicon Valley VC investors, believes the United States should be very careful about open weight or open source models. His views are echoed by U.S. senators such as Josh Hawley (R-Mo.), who introduced a bill to decouple U.S. AI development from China.

However, a hard separation could severely hinder scientific progress. It is reminiscent of World War I, when researchers lost access to discoveries across geopolitical divides, with lasting effects. Though a “digital iron curtain” may seem appealing to some, it risks a dystopian future, while tensions between private, state-run and open-source models continue to grow.

3. The value is no longer in the price

As entry barriers vanish and costs approach zero, the idea of monopolizing AI becomes increasingly absurd.

What does one ultimately need to develop an AI model? Data, human capital and training compute — three things that aren’t too hard to come by. The real competitive edge may lie in commoditizing the complement; in other words, rather than seeking to monetize AI directly, businesses should use it to enhance other areas of value.

Platforms with large user bases now hold the power, and companies that control distribution and engagement will extract the most value from AI, regardless of whether they own or develop the models themselves.

This may explain why Meta, with its massive user ecosystem, remained relatively stable in the days following the arrival of DeepSeek, while Nvidia took a hit. Similarly, Grok 3 may lack an inherent competitive advantage, but integrated into Musk’s X ecosystem, it becomes far more valuable.

The question, therefore, is no longer how much can AI be sold for, but rather, what new possibilities does it unlock?