Shares in major technology companies fell sharply Monday after a little-known startup in China released powerful new artificial intelligence models built for cheap with equipment that is not considered top of the line, sparking worries about U.S. dominance in a crucial emerging sector.

The development is also raising questions about the hundreds of billions of dollars that companies such as OpenAI, Meta and Google are spending to develop AI, even as some experts cast doubt on the Chinese company’s claims.

Nvidia Corp. NVDA-Q +1.45% increase, the leading maker of chips that are used to power AI, fell 17 per cent as of Monday afternoon, shaving roughly US$550-billion from its market cap, more than the combined total market cap of Canada’s four most valuable public companies. Taiwan Semiconductor Manufacturing Co. Ltd. dropped 14 per cent, as did Oracle Corp. Alphabet GOOGL-Q +0.66% increase and Microsoft MSFT-Q +0.46% increase dropped 3.2 per cent and 2.3 per cent respectively.

Canadian company Celestica Inc. CLS-T +4.64% increase, which builds equipment for data centres and has seen its shares soar on demand for AI hardware, plummeted by 26 per cent. Some energy companies, which power data centres, did not escape the rout either, including Canada’s Capital Power Corp. CPX-T +1.80% increase, which shed 15 per cent of its market value.

The selloff was prompted by DeepSeek, a company in China. The startup released a large language model called V3 in December and followed it up with another called R1. The latter is known as a reasoning model that can handle more complex questions and problems than standard chatbots.

The company has flown under the radar, with discussion about it limited mostly to AI developers. But talk hit a frenzy over the weekend, especially as Silicon Valley leaders picked up on a technical report from the company that said it spent just US$5.6-million to train one of its models. U.S. companies are spending hundreds of millions of dollars to train individual models, while research firm Epoch AI has estimated that building the largest AI models will cost more than US$1-billion by 2027.

Not only did DeepSeek appear to dramatically undercut development costs, it claimed to have done so using a relatively small number of Nvidia chips that are not even state-of-the-art, as the U.S. has attempted to restrict the export of advanced chips to China in recent years to thwart the country’s progress in AI. The benchmarks used to assess the prowess of DeepSeek’s models show that R1 is roughly on par with some of the latest releases from U.S. companies, such as OpenAI and Meta META-Q +1.52% increase. The models are also open-source, meaning the code is available for free. To top it off, DeepSeeek unveiled an image generation model on Monday that it says outperforms its competitors.

DeepSeek did not respond to interview requests.

“Constraints create innovation,” said Boris Wertz, founder of Version One Ventures in Vancouver. “In this case, it was more a whole industry that got a little drunk,” he said, referring to the money AI companies have spent to acquire graphics processing units and build data centres.

Just last week, OpenAI, Oracle and others announced a plan to spend up to US$500-billion to construct data centres to power AI in the U.S., while Meta is spending up to US$65-billion on AI infrastructure this year. Elon Musk’s xAI bought 100,000 GPUs to build what Nvidia has described as the world’s largest AI supercomputer.

The performance and costs of DeepSeek’s models, however, are throwing those huge numbers into a new light. “What happens with these hundreds of billions of dollars of capital that are being raised and funneled into the space?” said John Ruffolo, founder of Maverix Private Equity. “If you’re an investor devoted to this space, you must be panicking today.”

The release from DeepSeek comes at a time when adoption of AI “is still incredibly low, as people and businesses grapple with understanding how to best leverage the technology,” said Nicole Janssen, co-founder at AltaML in Edmonton.

The release of ChatGPT from OpenAI in 2022 kicked off a frenzy of development and investment, as businesses try to dominate the market for generative AI, which refers to applications that produce and interpret text and other media, even as concerns remain about accuracy and reliability.

Lower cost could help drive AI adoption, but some experts are skeptical about DeepSeek’s claims.

Alexandr Wang, the founder of U.S. company Scale AI, told CNBC recently that DeepSeek could have as many as 50,000 high-powered GPUs that it cannot publicly disclose owing to U.S. export controls.

“We need to take statements from DeepSeek about their compute resources with a grain of salt. We have no way to verify these numbers,” said Daniel Roy, a professor at the University of Toronto and research director at the Vector Institute. DeepSeek has also not been transparent about the data that it used to train its models, he added.

Thanos Moschopoulos, a technology analyst at BMO Capital Markets, advised investors in a research note to be “cautious” about claims about training costs, writing that more testing is needed to fully understand the uses and limitations of DeepSeek’s models.

“Ultimately, while this creates incremental uncertainty regarding the longer-term size and growth of the market for AI infrastructure, we expect this to remain a large total addressable market,” he wrote.

The implications of both cheaper and more efficient AI development could affect some of Canada’s largest investors, including pension funds, that have placed major bets on data centre companies and the energy producers that power these facilities. Such bets are typically long-term, and major pension funds don’t often react to market volatility.

The Canada Pension Plan Investment Board (CPPIB) is expected to soon own a major stake in Baltimore-based Constellation Energy Corp., which was one of the stocks that was hit hardest in Monday’s sell-off, falling 20 per cent by mid-afternoon.

CPPIB also owns a nearly $900-million stake in U.S.-based data centre operator Equinix Inc., which had dropped more than 5 per cent by mid-afternoon.

The Public Sector Pension Investment Board, which manages pensions for the federal public service, has backed Toronto-based Cohere, a builder of AI models that was valued at US$5.5-billion last year. Cohere and DeepSeek have different approaches, however. Cohere focuses on building proprietary AI models for use by enterprise customers, where security and privacy are paramount. Implementing open-source models, such as those made by DeepSeek, can require more technical expertise.

“The next revolution in LLMs won’t come from throwing money at the problem,” said Cohere cofounder Nick Frosst in a statement. “This is a moment of clarity, where the obvious becomes undeniable in how we will pursue better LLMs.”

Gennady Pekhimenko, co-founder of Toronto-based CentML and an expert in machine learning efficiency, said the few million dollars in training costs cited by DeepSeek are likely not the entire picture, and that the company relied on other AI models as part of its process. “Their training utilizes the existing models to get very high-quality, representative data, so they needed to train way less,” he said. “It’s not like you just start the process and get everything from scratch and you’re done.”

Mr. Pekhimenko said there remains a need for GPUs and data centres to build large, foundational models. Plus, running AI models after training – such as asking a question of ChatGPT – requires a lot of infrastructure. “It doesn’t mean you shouldn’t build data centres,” he said.

DeepSeek is sparking geopolitical concerns, too. The U.S. and China are locked in a race for supremacy in AI, which could bring enormous economic benefits and military advantages. Silicon Valley investor Marc Andreesen wrote on X that DeepSeek’s latest model is “AI’s Sputnik moment.”

“AI is the front for the new Cold War. Superiority of technology in the digital world is correlated to superiority militarily,” Mr. Ruffolo said.

DeepSeek has limitations, notably that its responses are in line with the views of the Chinese Communist Party. “Sorry, that’s beyond my current scope,” the DeepSeek app responded when The Globe and Mail asked about the 1989 Tiananmen Square protests. “Let’s talk about something else.” The app also said that Taiwan has “always been an inalienable part of China’s territory since ancient times.”

Because DeepSeek has published its methods, other companies could try to replicate its approach to building cheaper, more powerful models. Vijay Gadepally, the chief technology officer at Canadian AI cloud computing company Radium, noted the potential benefits. “It lowers the barrier to entry,” he said, “forces us to rethink the investment ratio between infrastructure and algorithms for AI progress, and is a step forward in reducing the energy and environmental impact of AI.”

Joe Castaldo
Sean Silcoff, Technology Reporter
James Bradshaw, Institutional Investing Reporter
The Globe and Mail, January 27, 2025