Opinion: What China’s DeepSeek breakthrough really means for the future of AI
- Share via
Last week, the Nasdaq stock exchange — which lists significant U.S. tech stocks — experienced a big drop. This resulted from the Chinese startup DeepSeek announcing that it had developed an artificial intelligence model that performs as well as OpenAI and Meta’s AI technology, but at a fraction of the cost and with less computing power.
AI chip designer Nvidia lost nearly $600 billion of its market capitalization (the total dollar value of its outstanding shares of stock) — the largest single-day drop experienced by a company in U.S. market history. Although Nvidia’s share price has recovered some ground, analysts continue to second-guess ambitious AI infrastructure plans, including the company’s specialized graphics processing unit chips as well as massive data centers like those built and operated by Amazon.
DeepSeek’s creators claim to have found a better way to train their AI by using special parts, improving how the AI learns rules and deploying a strategy to keep the AI running smoothly without wasting resources. According to the company’s report, these innovations drastically reduced the computing power needed to develop and run the model and therefore the cost associated with chips and servers. This sharp cost reduction has already attracted smaller AI developers looking for a cheaper alternative to high-profile AI labs.
Irony of ironies: Authors and artists have accused OpenAI of stealing their content to ‘train’ its bots -- but now OpenAI is accusing a Chinese company of stealing its content to train its bots.
At first glance, reducing model-training expenses in this way might seem to undermine the trillion-dollar “AI arms race” involving data centers, semiconductors and cloud infrastructure. But as history shows, cheaper technology often fuels greater usage. Rather than dampen capital expenditures, breakthroughs that make AI more accessible can unleash a wave of new adopters, including not only tech startups but also traditional manufacturing firms and service providers such as hospitals and retail.
Microsoft Chief Executive Satya Nadella called this phenomenon a “Jevons paradox” for AI. Attributed to the 19th century English economist William Stanley Jevons, the concept describes how making a technology more efficient can raise rather than lessen consumption. Steam and electrical power followed this pattern: Once they became more efficient and affordable, they spread to more factories, offices and homes, ultimately increasing use.
Nadella is right: Today’s plummeting development costs for generative AI are poised to generate a similar expansion. That means the sky is not falling for Big Tech companies that supply AI infrastructure and services. Major tech players are projected to invest more than $1 trillion in AI infrastructure by 2029, and the DeepSeek development probably won’t change their plans all that much.
While training costs may drop, the long-term hardware requirements for massive machine learning workloads, data processing and specialized AI software remain enormous. Although chip prices might fall as model training becomes more efficient, AI-based applications — such as generative chatbots and automated industrial controls — demand powerful servers, high-speed networks to transmit massive data flows and reliable data centers to handle billions of real-time queries. Regulatory, security and compliance demands further complicate implementation, requiring advanced, sometimes costly solutions that can store and process data responsibly.
Artificial intelligence could achieve sentience in 10 years. We should prepare for these systems to have their own subjective experiences, including sensing pain caused by humans.
General-purpose technologies that transform economies typically spread in two stages. First, during a long gestation period, well-funded organizations experiment, refining prototypes and processes. Later, once standards stabilize and ready-to-use solutions emerge, more cautious firms jump in. In the case of electricity, the first stage saw factories spending years reorganizing production floors and adopting new workflows before electrification spread widely; in the case of AI, it has consisted of big banks, retailers and manufacturers making slow, piecemeal use of the technology.
A century and a half ago, when the Bessemer process introduced the use of hot air to blast impurities out of molten iron and mills figured out how to produce standardized steel products, manufacturers pivoted. Steel prices plummeted and consumption soared, eventually increasing spending in that sector despite steelmakers’ more efficient use of iron ore.
Now that DeepSeek and other innovations promise lower costs, more companies may be ready to embrace or at least try AI, and the demand for AI infrastructure is likely to increase. A more affordable, cutting-edge model could also encourage industries, startups and entrepreneurs to use AI more widely, increasing its adoption in logistics, customer service and more.
Imagine, for example, a 200-person law firm specializing in commercial real estate. Initially, it uses ChatGPT sometimes to produce quick contract summaries, but its partners grow uneasy about inconsistent quality and confidentiality risks. After testing a contracts-focused model provided by a reputable vendor, the firm adopts technology that integrates directly with its document management system. This allows associate attorneys to auto-summarize hundreds of pages in seconds, rely on AI “clause suggestions” tailored to real estate precedents, and limit the need to seek guidance from senior partners to cases of especially ambiguous or high-stakes language. Moreover, the system design prevents client data from leaving the firm’s domain, increasing security.
Over time, the firm adds AI modules for advanced litigation research and automated billing notes, steadily reducing administrative tasks and letting human experts focus on strategic legal insight. It sees quicker contract turnaround, standardized billing and a new willingness among partners to explore AI-based tools in other areas.
In short, AI’s capital demands won’t shrink thanks to DeepSeek; they will become more widely distributed. We’ll see this spur expansion in power grids, cooling systems, data centers, software pipelines and infrastructure that enables more devices to use AI, including robots and driverless cars. The trillion-dollar infrastructure push may persist for years to come.
Victor Menaldo is a political science professor at the University of Washington and is writing a book on the political economy of the fourth industrial revolution.
A cure for the common opinion
Get thought-provoking perspectives with our weekly newsletter.
You may occasionally receive promotional content from the Los Angeles Times.
