DeepSeek, the Chinese startup that roiled AI stocks and the broad market this week, threw a wrench in U.S. efforts to stay ahead of China in artificial intelligence.
While DeepSeek's advances have been eyed in Silicon Valley since December, its announcement of a new AI model on Jan. 20 and launch of a ChatGPT-like app triggered a sell-off in AI stocks as word of its advances spread. Nvidia stock plunged 17% on Jan. 27 as other AI chipmakers also took big hits.
DeepSeek's thrifty approach to building large language models raised worries over demand for high-end AI chips and for power to run AI-centric data centers. The Chinese lab used creative software techniques and far less computing power than U.S. rivals in AI.
"There's plenty of uncertainty over what the true demand for state-of-the-art chips, semiconductor fabrication plants and energy will be," economist Ed Yardeni said in a note.
The rout on Wall Street upended Wall Street's view of AI's future, highlighting growing worries that massive investments in AI infrastructure may not be worth it. DeepSeek, a little-known Chinese startup, claimed it spent under $6 million on model development.
Questions remain on DeepSeek's technology and its ability to compete with U.S. companies like startups OpenAI and Anthropic as well as tech giants like Google-parent Alphabet and Meta Platforms.
AI Stocks: The Jevons Paradox
Yet much cheaper methods for building AI systems could spur wider adoption by U.S. companies.
"Jevons paradox strikes again!" Microsoft Chief Executive Satya Nadella posted on X recently, referring to the impact of technological advancements.
At the heart of the Jevons paradox is the idea that as a technology becomes more energy efficient, the cost of using that technology declines.
"As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of," Nadella said.
Yardeni echoed this view, saying, "The negative consequence of DeepSeek is that it challenges the business models of American companies that expected to use their exclusive access to Nvidia's most expensive and powerful chips to dominate and profit from the AI revolution."
"The good news is that they should be able to follow DeepSeek's lead in lowering the cost of AI infrastructure spending," he added. "Also good: more competition in the AI economy will give business and individual consumers more bang for their AI bucks."
Arms Race To 'AGI' Impact
One possibility is that capital spending will shift from training AI models to setting up capacity for AI applications. Meta, for example, last week guided to higher capital spending in 2025 even though it was aware of DeepSeek's progress, analysts say.
In the wake of DeepSeek's efficiency gains, any changes in spending plans by cloud computing giants will be a key area for investors to watch.
And the pursuit of artificial general intelligence (AGI) by OpenAI and tech giants could be key. AGI capability is considered the highest form of AI, featuring humanlike or better cognitive capabilities.
"With the added cost efficiency of AI training/inference from DeepSeek, the industry has two choices to choose from: altogether spend less on AI going forward versus beginning an arms race to reach AGI," said a Bank of America report.
"We believe the latter is the more likely scenario, with hyperscalers and AI developers increasing their capex outlook."
Open AI said that's where it's heading.
"Look forward to bringing you all AGI and beyond," OpenAI CEO Sam Altman tweeted in response to DeepSeek's Jan. 20 announcement.
Meanwhile, Wall Street analysts generally stuck with a theme that the commoditization of AI models would spur product innovation.
DeepSeek: Potential Winners And Losers
Analysts said the best AI stocks may shift from infrastructure-related "picks and shovels" to "apps and services."
"Long term, the continued pressure to lower the cost of compute — and the ability to reduce the cost of training and inference using new, more efficient algorithmic techniques — could result in lower capex than previously envisioned and lessen Nvidia's dominance, especially if large-scale GPU clusters are not as critical to achieve frontier-level model performance," said a William Blair report.
Broadcom and Marvell Technologies are among other AI chipmakers that took a hit on DeepSeek.
The William Blair report added: "We see DeepSeek news as largely positive for the software and the application layer as it will make AI cheaper and more accessible and enable it to be more cost-effectively integrated into a larger swath of existing and new applications."
"Cheaper AI models should speed up and pull forward development of new products, services, and the expected gains in productivity, even if it comes at the expense of equipment providers in the short term," said a JPMorgan report.
Oppenheimer said in a report: "We think this will help with AI application growth, but will likely be volatile and remains too early to predict winners."
Who Owns DeepSeek, And What Is Its AI Model?
Chinese hedge fund High Flyer, founded by Liang Wenfeng, controls DeepSeek. DeepSeek's R1 model, released Jan. 20, mimics the way humans reason, much like OpenAI's latest model.
Within the U.S. tech industry, observers generally credited DeepSeek with innovative approaches in training AI models that improved computing efficiency. Still, many believe DeepSeek borrowed from preexisting models, possibly from OpenAI or Meta's open-source models using a technique called "distillation."
Some observers are also skeptical about DeepSeek's claim that its AI model uses far less computing power than its U.S. counterparts. One view is that Chinese companies have been stockpiling thousands of Nvidia chips, despite increased export controls, by purchasing from third parties.
"DeepSeek created an awesome LLM (large language model) model," said Wedbush analyst Daniel Ives in a report. "However, this Chinese AI lab/LLM model is not bringing down the entire U.S. tech ecosystem with it."
Ives compared DeepSeek to Temu, the e-commerce site owned by China-based PDD Holdings. Temu was viewed as an "Amazon model destroyer a few years ago." But "Amazon's team adjusted, and now look where Temu and Amazon both are sitting," Ives said.
AI Stocks: The Trump Factor
DeepSeek's emergence could have big implications for U.S. policy toward AI development as well as China. The market sell-off came a few days after President Donald Trump unveiled "Stargate," a program to build out data centers for OpenAI. The Stargate program founders initially target $100 billion capex and say that could reach $500 billion in a few years.
Japan's SoftBank Group, Oracle, MGX and OpenAI are the initial equity funders in Stargate. Arm Holdings, Microsoft and Nvidia are technology partners.
Stargate underlines Trump's focus on AI. He revoked President Joe Biden's AI executive order, which required companies training large AI models to share information about them with the federal government.
Trump also ordered Silicon Valley venture capitalist David Sacks, his new artificial intelligence and cryptocurrency czar, to come up with an AI "action plan" in 180 days.
Is DeepSeek AI's Sputnik Moment?
Meanwhile, a government commission has proposed that Congress fund a Manhattan Project-like program dedicated to acquiring an Artificial General Intelligence capability.
DeepSeek has had such a huge impact on the perceived U.S. dominance in AI that venture capitalist Marc Andreessen called the launch of its R1 open source model "AI's Sputnik Moment." He was referring to the 1957 launch of a Soviet satellite that spurred the space race to the moon.
AI clearly is an important battleground and the U.S. lead could not be taken for granted. In a recent interview, Alphabet President Ruth Porat said, "It isn't a foregone conclusion that the U.S. will keep its advantage." She noted that with Nvidia and others, the U.S. leads in AI chips, but China "is on par and may even be a bit ahead on what's called diffusion of basic capabilities."
China clearly has emerged as a formidable challenger.
"In the U.S. and during the Biden administration, the general take has been that bigger models are going to be better, that you want to have more compute, you want to have more data, you want to spend a lot of resources to make these big models," said Sam Bresnick, a research fellow at Georgetown's Center for Security and Emerging Technology (CSET).
He added: "In China, I think because they are more compute constrained, and potentially for other reasons, they've focused more on smaller models, which are tailored to specific industries."
AI Stocks: Chip Export Rules
It remains to be seen whether Trump will follow through on tougher export rules.
"The people surrounding Trump are all generally China hawks in that they are distrustful of China. They fear Chinese power, whether that is military power or technological power," Bresnick added. "I think people in the administration are going to be pretty concerned about cutting a deal on AI with China revoking export controls because there's very good reason to believe the Chinese AI industry would get a big boost from that. DeepSeek's CEO said the number one issue they faced was limited access to compute resources."
DeepSeek claims to have used Nvidia's H800 chips, which were designed to comply with U.S. export controls released in 2022. But some claim it used more advanced H100 AI chips from Nvidia.
"DeepSeek's work illustrates how new models can be created leveraging widely available (open source) models and compute that is fully export control compliant," Nvidia said in a statement.
China's Tech Giants Also A Potential Target?
Chinese companies such as Huawei's HiSilicon, meanwhile, continue to develop more powerful homegrown computer chips.
If chip export rules do not thwart China's development of AI systems, Trump has other options. In Foreign Affairs, Rhodium Group analyst Reva Goujon suggested the U.S. could expand trade controls to target China's tech giants such as Alibaba, Baidu and Tencent.
Alibaba on Jan. 29 unveiled a new version of its AI model called Qwen2.5 Max that it claims has performance comparable or better than OpenAI, DeepSeek, Google and Meta models.
Also, Baidu reportedly developed a technology enabling the integration of GPUs from different suppliers into a single AI model training cluster, mitigating chip shortages.
Alibaba provides global cloud computing services similar to those of Amazon. Its cloud services include access to its own large language model, called Qwen.
Alibaba operates data centers in about 30 countries, while Tencent has data centers in 21 countries.
AI Stocks: Global Data Center Investments
One view is that U.S.-China AI competition will be global, with both countries aiming to find partners in the Middle East and Asia to build out energy-intensive data centers. The new Stargate AI project includes Abu Dhabi's MGX, for example.
Jared Cohen, cohead of the Goldman Sachs Global Institute, commented on global AI competition at a recent conference.
"Saudi Arabia, Qatar, and UAE are what I would describe as geopolitical swing states," he said. "They can either lean into China on AI or they can lean into the U.S. on AI."
He added that the U.S. has an edge because it produces the most powerful computer chips and has export controls. Many countries in the Middle East as well as Asia want to build their own sovereign AI data centers. Still, China is a big trading partner of Middle Eastern countries.
Both OpenAI and Microsoft have stepped up lobbying. Microsoft is among AI stocks to watch.
"There's an estimated $175 billion sitting in global funds awaiting investment in AI projects, and if the U.S. doesn't attract those funds, they will flow to China-backed projects — strengthening the Chinese Communist Party's global influence," OpenAI said in its "blueprint for U.S. AI infrastructure" released Jan. 13.
AI investments in the U.S. were high even before the Stargate project was announced. Microsoft and BlackRock launched a $30 billion fund to invest in AI infrastructure last year.
The U.S. has led the global AI private investment cycle, pouring over $430 billion into it from 2014 to 2024, far outpacing China and Europe, according to a JPMorgan report.
AI Stocks: China, U.S. Cooperation?
Despite hawkish views that U.S./China AI competition is a zero-sum scenario with military applications, some observers see opportunities for cooperation.
Bresnick at Georgetown's CSET noted that the two superpowers have been careful to stay at the negotiating table. In November, the Biden administration and Chinese President Xi announced a deal that addressed a nuclear military risk. They agreed that artificial intelligence will not be allowed to make decisions involving the use of nuclear weapons.
An AI treaty could be similar to agreements on nuclear deployment. An agreement could focus on AI safety standards and governance frameworks.
An AI treaty would be harder to enforce than nuclear pacts, said Oren Etzioni, a University of Washington professor and founding CEO of the Allen Institute for Artificial Intelligence.
But Etzioni told IBD that U.S. and Chinese research universities could play an important role in AI cooperation for issues like climate change or managing pandemics.
"My belief is that it's multifaceted and while we may need to restrict certain things, we can still maintain cooperation, particularly in academia," Etzioni said.
"You have to look at the specific arena, whether it's chips or open source LLMs, and decide (on restrictions). When the Trump administration talks about something like (an AI) Manhattan Project, there are reasons to make major investments. That's not inconsistent with having continuing academic interchange."
Follow Reinhardt Krause on Twitter @reinhardtk_tech for updates on artificial intelligence, cybersecurity and cloud computing.
