DeepSeek’s big question: Where does AI’s true value reside?

The given news that OpenAI is investigating whether DeepSeek trained its chatbot by repeatedly querying the U.S. company’s AI models to exfiltrate large volumes of data in a technical process called distillation.

Even if that is the case, however, the competitive threat to U.S. tech companies is rising, as is the challenge of protecting their intellectual property.

Questions about distillation of models notwithstanding, DeepSeek has earned the attention—and the respect—of U.S. technologists. “I would say, personally, we have seen the Chinese today, this week establish themselves as first-rate software developers," Jared Spataro, chief marketing officer of Microsoft’s AI at Work group, told me.

Venture capital investors and the startup founders of Silicon Valley have long defined themselves as disrupters, attacking the flanks of established companies. In a cycle going back more than half a century, one startup emerged after another, more or less from Silicon Valley, San Francisco or Seattle, the northern capital of the U.S. tech empire.

When promising startups emerged elsewhere, they were seen as farm-team players to elevate to the big leagues through investment and alliances. Microsoft’s partnership with French AI startup Mistral, itself founded by Google DeepMind and Meta veterans, sparked some resentment in Europe because it seemed to follow that model. It has been funded by U.S. investors including Lightspeed General Partners, General Catalyst and former Google CEO Eric Schmidt.

DeepSeek upended Silicon Valley in a more profound way, because the disruption came from outside its circle of venture firms, entrepreneurs, tech giants and universities.

DeepSeek was started in Hangzhou in 2023 by Liang Wenfeng, and is owned entirely by Chinese hedge fund High-Flyer, which Liang launched in 2015 with two college friends.

Liang met on Jan. 20—Inauguration Day for President Trump—with China’s premier and discussed how homegrown companies could narrow the gap with the U.S., The Wall Street Journal reported.

DeepSeek’s rise shows companies from China and elsewhere can shake the tech foundations of the U.S., which is already mired in battles over control of China-based TikTok and the export of technologies such as Nvidia’s leading-edge AI chips.

That’s why venture capitalist Marc Andreessen called DeepSeek’s R1 “AI’s Sputnik moment."

The high-end theory

But domination doesn’t necessarily follow from disruption.

The market for AI instead could face bifurcation along the lines of the mobile phone market, where global Android market share was 82% at the end of the third quarter of 2024, according to researcher IDC. Apple captures value, meanwhile, by excelling in the high-profit business of high-end smartphones.

Even if DeepSeek is roughly comparable to the best currently on offer in the U.S., it apparently needed to build on the underlying work of U.S. companies, and top AI developers such as OpenAI have more powerful models in the works.

DeepSeek could play the Android role and help commoditize much of the AI market, especially because R1 is optimal for small devices, according to Stephen Messer, co-founder of enterprise AI company Collective[i]. “If that is the case, AI becomes cheaper and more accessible and it is built into everything," Messer said.

History shows that as AI’s price performance improves, usage goes up. Two years after OpenAI introduced ChatGPT, for example, its most cost-efficient small model, 4o mini, can process a token, or unit of data, for 1/150th of the original price, according to company documents reviewed by the Journal.

Demand has also increased; ChatGPT had 300 million weekly average free and paid users by December, up from 100 million in November 2023.

“As these models commoditize, don’t think that that means it’s a bad business to be in," Microsoft’s Spataro said. “That just means there’s going to be more usage."

DeepSeek’s efficiency could accelerate demand, to the potential benefit of U.S. chip giant Nvidia and other key chip makers and suppliers. And while there will always be demand for leading edge chips and models, more of the value in tech could shift away from foundation models, infrastructure and traditional software-as-a-service toward AI-based agents. U.S.-based companies Salesforce and ServiceNow have staked early claims there.

“We feel like the value of AI is going to end up migrating up to essentially the application layer," Spataro said. “You’re going to see the value of taking a great technology and applying it to business problems."

These factors could help big U.S. tech companies retain leadership in the upper end of the AI market—and most of the value created by AI—although make no mistake, they will face competition there as well.

The next level

OpenAI’s forthcoming o3 reasoning model is expected to be a significant advance over its current models. And that’s how OpenAI and other pioneer model developers hope to maintain their edge: by producing a flow of models setting new standards for performance, which quickly becomes a commodity.

That business model, which also drives Nvidia, anticipates enough companies will find it essential to have the best models and chips, and will keep paying for them. If such assumptions about ongoing demand for cutting-edge models are correct, DeepSeek’s shocking emergence could infuse the AI industry with innovation, help U.S. tech giants speed toward goals such as the development of artificial general intelligence, and expand their market. But competitive pressure has also suddenly reached a new level, and there’s no going back.

Write to Steven Rosenbush at steven.rosenbush@wsj.com