Last month, the Chinese artificial-intelligence startup DeepSeek released a new large language model (LLM) that it claims performs at least as well as those of leading American AI firms, despite having cost far less to train, and using a fraction of the computing power. Almost immediately, US tech stocks plunged, with chipmaker Nvidia losing a record-breaking $589 billion in value in one day.

It is a shakeup that America’s scale-obsessed AI industry “arguably needed,” writes Nobel laureate economist Daron Acemoglu. DeepSeek’s rise has exposed the industry’s “groupthink,” which left it blind to “alternative, cheaper, more promising approaches.” It has also raised difficult questions, including “whether the US industry has other, even more dangerous blind spots,” and whether China, with its “extractive institutions,” is “leapfrogging the US,” despite US export controls.

Tech investor Charles Ferguson agrees, viewing a “powerful, innovative Chinese model achieving parity with US products” as the “predictable result of a major US and Western policy failure.” Since AI will drive “enormously consequential economic and military transformations,” democratic countries must seek to maintain AI leadership, through an “enormous public-private strategic mobilization on the scale of the Manhattan Project, NATO, various energy-independence efforts, or nuclear-weapons policies.”

But Angela Huyue Zhang of the University of Southern California thinks the focus on the US-China rivalry “misses the point.” This is a case not of China challenging the US, but of an “ingenious startup” challenging “oligopolistic incumbents,” which undoubtedly knew that “lower-cost models were feasible,” but made a “deliberate business choice” not to invest in developing them. Given this, DeepSeek’s rise should be welcomed, as it may well inject “badly needed” competition into the US AI sector.

Similarly, Columbia University’s Amar Bhidé argues that “seeing transformational advances abroad as a threat to domestic well-being is misguided,” because the blind pursuit of “technological or scientific leadership” is a “foolish enterprise.” While DeepSeek’s approach “cannot overcome the unavoidable limitations of backward-looking statistical models,” it could make their “price performance good enough for wider use.” In this sense, DeepSeek’s technology could be to LLMs what “Nicola Tesla’s breakthroughs with alternating current were to electrification.”

BOSTON – After the release of DeepSeek-R1 on January 20 triggered a massive drop in chipmaker Nvidia’s share price and sharp declines in various other tech companies’ valuations, some declared this a “Sputnik moment” in the Sino-American race for supremacy in artificial intelligence. While America’s AI industry arguably needed shaking up, the episode raises some difficult questions.

The US tech industry’s investments in AI have been massive, with Goldman Sachs estimating that “mega tech firms, corporations, and utilities are set to spend around $1 trillion on capital expenditures in the coming years to support AI.” Yet for a long time, many observers, including me, have questioned the direction of AI investment and development in the United States.

With all the leading companies following essentially the same playbook (though Meta has differentiated itself slightly with a partly open-source model), the industry seems to have put all its eggs in the same basket. Without exception, US tech companies are obsessed with scale. Citing yet-to-be-proven “scaling laws,” they assume that feeding ever more data and computing power into their models is the key to unlocking ever-greater capabilities. Some even assert that “scale is all you need.”

Before January 20, US companies were unwilling to consider alternatives to foundation models pretrained on massive data sets to predict the next word in a sequence. Given their priorities, they focused almost exclusively on diffusion models and chatbots aimed at performing human (or human-like) tasks. And though DeepSeek’s approach is broadly the same, it appears to have relied more heavily on reinforcement learning, mixture-of-experts methods (using many smaller, more efficient models), distillation, and refined chain-of-thought reasoning. This strategy reportedly allowed it to produce a competitive model at a fraction of the cost.

Although there is some dispute about whether DeepSeek has told us the whole story, this episode has exposed “groupthink” within the US AI industry. Its blindness to alternative, cheaper, more promising approaches, combined with hype, is precisely what Simon Johnson and I predicted in Power and Progress, which we wrote just before the generative-AI era began. The question now is whether the US industry has other, even more dangerous blind spots. For example, are the leading US tech companies missing an opportunity to take their models in a more “pro-human direction”? I suspect that the answer is yes, but only time will tell.

Then there is the question of whether China is leapfrogging the US. If so, does this mean that authoritarian, top-down structures (what James A. Robinson and I have called “extractive institutions”) can match or even outperform bottom-up arrangements in driving innovation?

My bias is to think that top-down control hampers innovation, as Robinson and I argued in Why Nations Fail. While DeepSeek’s success appears to challenge this claim, it is far from conclusive proof that innovation under extractive institutions can be as powerful or as durable as under inclusive institutions. After all, DeepSeek is building on years of advances in the US (and some in Europe). All its basic methods were pioneered in the US. Mixture-of-experts models and reinforcement learning were developed in academic research institutions decades ago; and it was US Big Tech firms that introduced transformer models, chain-of-thought reasoning, and distillation.

What DeepSeek has done is demonstrate success in engineering: combining the same methods more effectively than US companies did. It remains to be seen whether Chinese firms and research institutions can take the next step of coming up with game-changing techniques, products, and approaches of their own.

Moreover, DeepSeek seems to be unlike most other Chinese AI firms, which generally produce technologies for the government or with government funding. If the company (which was spun out of a hedge fund) was operating under the radar, will its creativity and dynamism continue now that it is under the spotlight? Whatever happens, one company’s achievement cannot be taken as conclusive evidence that China can beat more open societies at innovation.

Another question concerns geopolitics. Does the DeepSeek saga mean that US export controls and other measures to hold back Chinese AI research failed? The answer here is also unclear. While DeepSeek trained its latest models (V3 and R1) on older, less powerful chips, it may still need the most powerful chips to achieve further advances and to scale up.

Nonetheless, it is clear that America’s zero-sum approach was unworkable and ill advised. Such a strategy makes sense only if you believe that we are heading toward artificial general intelligence (models that can match humans on any cognitive task), and that whoever gets to AGI first will have a huge geopolitical advantage. By clinging to these assumptions – neither of which is necessarily warranted – we have prevented fruitful collaboration with China in many areas. For example, if one country produces models that increase human productivity or help us regulate energy better, such innovation would be beneficial to both countries, especially if it is widely used.

Like its American cousins, DeepSeek does aspire to develop AGI, and creating a model that is significantly cheaper to train could be a game changer. But bringing down development costs with known methods will not miraculously get us to AGI in the next few years. Whether near-term AGI is achievable remains an open question (and whether it is desirable is even more debatable).

Even if we do not yet know all the details about how DeepSeek developed its models or what its apparent achievement means for the future of the AI industry, one thing seems clear: a Chinese upstart has punctured the tech industry’s obsession with scale and may have even shaken it out of its complacency.

SAN FRANCISCO – The release of the Chinese DeepSeek-R1 large language model, with its impressive capabilities and low development cost, shocked financial markets and led to claims of a “Sputnik moment” in artificial intelligence. But a powerful, innovative Chinese model achieving parity with US products should come as no surprise. It is the predictable result of a major US and Western policy failure, for which the AI industry itself bears much of the blame.

China’s growing AI capabilities were well known to the AI research community, and even to the interested public. After all, Chinese AI researchers and companies have been remarkably open about their progress, publishing papers, open-sourcing their software, and speaking with US researchers and journalists. A New York Times article from last July was headlined, “China Is Closing the AI Gap with the United States.”

Two factors explain China’s achievement of near parity. First, China has an aggressive, coherent national policy to reach self-sufficiency and technical superiority across the entire digital technology stack, from semiconductor capital equipment and AI processors to hardware products and AI models – and in both commercial and military applications. Second, US (and EU) government policies and industry behavior have exhibited a depressing combination of complacency, incompetence, and greed.

It should be obvious that Chinese President Xi Jinping and Russian President Vladimir Putin are no friends of the West, and that AI will drive enormously consequential economic and military transformations. Given the stakes involved, maintaining AI leadership within democratic advanced economies justifies, and even demands, an enormous public-private strategic mobilization on the scale of the Manhattan Project, NATO, various energy-independence efforts, or nuclear-weapons policies. Yet the West is doing the opposite.

In the US, government and academic research in AI are falling behind both China and the private sector. Owing to inadequate funding, neither government agencies nor universities can compete with the salaries and computing facilities offered by the likes of Google, Meta, OpenAI, or their Chinese counterparts. Moreover, US immigration policy toward graduate students and researchers is self-defeating and nonsensical, because it forces highly talented people to leave the country at the end of their studies.

Then there is the US policy on regulating Chinese access to AI-related technology. Export controls have been slow to appear, wholly inadequate, poorly staffed, easily evaded, and under-enforced. Chinese access to US AI technologies through services and licensing agreements has remained nearly unregulated, even when the underlying technologies, such as Nvidia processors, are themselves subject to export controls. The US announced stricter licensing rules just a week before former President Joe Biden left office.

Finally, US policy ignores the fact that AI R&D must be strongly supported, used, and, where necessary, regulated throughout the private sector, the government, and the military. The US still has no AI or IT equivalent of the Department of Energy, the National Institutes of Health, NASA, or the national laboratories that conduct (and tightly control) US nuclear-weapons R&D.

This situation is partly the result of sclerotic government bureaucracies in both the European Union and the US. The EU technology sector is severely overregulated, and the US Departments of Defense and Commerce, among other agencies, need reform.

Here, the tech industry is somewhat justified in criticizing their governments. But the industry itself is not blameless: Over time, lobbying efforts and revolving-door personnel appointments have weakened the capabilities of critically important public institutions. Many of the problems with US policy reflect the industry’s own resistance or neglect. In critical ways, it has been its own worst enemy, as well as the enemy of the West’s long-term security.

For example, ASML (the Dutch maker of state-of-the-art lithography machines used in chip manufacturing) and the US-based semiconductor-equipment supplier Applied Materials both lobbied to weaken export controls on semiconductor capital equipment, thus assisting China in its effort to displace TSMC, Nvidia, and Intel. Not to be outdone, Nvidia designed special chips for the Chinese market that performed just slightly below the threshold set by export restrictions; these were then used to train DeepSeek-R1. And at the level of AI models, Meta and the venture capital firm Andreessen Horowitz have lobbied fiercely to prevent any limits on open-source products.

At least in public, the industry’s line has been: “The government is hopeless, but if you leave us alone, everything will be fine.” Yet things are not fine. China has nearly caught up with the US, and it is already ahead of Europe. Moreover, the US government is not hopeless, and must be enlisted to help. Historically, federal and academic R&D compare very favorably with private-sector efforts.

The internet, after all, was pioneered by the US Advanced Research Projects Agency (now DARPA), and the World Wide Web emerged from the European Organization for Nuclear Research (CERN). Netscape co-founder Marc Andreessen created the first web browser at a federally funded supercomputer center within a public university. Meanwhile, private industry gave us online services like CompuServe, Prodigy, and AOL (America Online) – centralized, closed, mutually incompatible walled gardens that were justly obliterated when the internet was opened to commercial use.

The challenges of AI R&D and China’s rise require a forceful, serious response. Where government capacity falls short, we need to bolster it; not destroy it. We need to pay competitive salaries for government and academic work; modernize US (and EU) technology infrastructure and procedures; create robust R&D capacity within the government, particularly for military applications; strengthen academic research; and implement rational policies for immigration, AI R&D funding, safety testing, and export controls.

The one truly difficult policy problem is openness, particularly open-source licensing. We cannot let everyone have access to models optimized for hunter-killer drone attacks; nor, however, can we stamp “top secret” on every model. We need to find a pragmatic middle ground, perhaps relying on national defense research laboratories and carefully crafted export controls for intermediate cases. Above all, we need the AI industry to realize that if we don’t hang together, we will hang separately.

LONDON – In 1957, the Soviet Union launched the world’s first artificial satellite into orbit, sparking fears in the United States that, unless it took radical action to accelerate innovation, its Cold War adversary would leave it in the technological dust. Now, the Chinese startup DeepSeek has built an artificial intelligence model that it claims can outperform industry-leading American competitors, at a fraction of the cost, leading some commentators to proclaim that another “Sputnik moment” has arrived.

But the focus on the US-China geopolitical rivalry misses the point. Rather than viewing DeepSeek as a stand-in for China, and established industry leaders (such as OpenAI, Meta, and Anthropic) as representatives of the US, we should see this as a case of an ingenious startup emerging to challenge oligopolistic incumbents – a dynamic that is typically welcomed in open markets.

DeepSeek has proved that software ingenuity can compensate, at least partly, for hardware deficiencies. Its achievement raises an uncomfortable question: Why haven’t leading US industry leaders achieved similar breakthroughs? Nobel laureate economist Daron Acemoglu points the finger at groupthink, which he says prevented Silicon Valley incumbents from adequately considering alternative approaches. He might have a point, but it is only half the story.

DeepSeek’s success didn’t happen overnight. In May 2024, the firm launched its V2 model, which boasted an exceptional cost-to-performance ratio and sparked a fierce price war among Chinese AI providers. Moreover, over the last year or so, Chinese firms – both giants (including Alibaba, Tencent, and ByteDance) and startups (like Moonshot AI, Zhipu AI, Baichuan AI, MiniMax, and 01.AI) – have all developed cutting-edge AI models with remarkable cost efficiency.

Even within the US, researchers have long explored ways to improve the efficiency – and thus lower the costs – of AI training. For example, in 2022, former Meta researcher Tim Dettmers, now at the Allen Institute for Artificial Intelligence, and his co-authors published research on optimizing AI models to run on less computing power. DeepSeek cited their research in the technical paper it released along with its V3 model.

Put simply, it would have been impossible for any AI firm – especially an industry leader – not to realize that lower-cost models were feasible. But American AI developers showed much less interest than their Chinese counterparts in pursuing this line of innovation. This was not a matter only of insularity or hubris; it appears to be a deliberate business choice.

AI development has so far been defined by the “scaling law,” which predicts that more computing power leads to more powerful models. This has fueled demand for high-performance semiconductor chips, with more than 80% of the funds raised by many AI companies going toward computing resources.

That is why the biggest winner has been the advanced chipmaker Nvidia, which claimed 90% of the market for AI graphics processing units by the end of last year. Thanks to this virtual monopoly in the hardware layer, Nvidia could control the foundations of generative AI. The cloud-computing sector, which provides the on-demand computing power AI models require, is similarly concentrated, with Amazon, Google, and Microsoft dominating the market.

But these upstream players aren’t just passive suppliers. They have strategically positioned themselves across the AI value chain by acquiring, investing in, or forming alliances with leading AI model developers. Nvidia has invested in OpenAI, Mistral, Perplexity, and others. Google not only develops its own AI models, but also holds a stake in Anthropic, OpenAI’s main competitor. And Microsoft, an early OpenAI investor, recently backed Inflection AI in the US and expanded overseas, with investments in France’s Mistral and the United Arab Emirates’ G42.

Taking this approach has ensured that the entire AI industry depends on a few giant firms and entrenched a dynamic whereby rising demand for computing power across the sector increases these firms’ profits. As dominant players, they had less incentive to improve cost efficiency downstream, which could cut into their upstream profits.

Chinese AI firms have been operating within an entirely different reality, as US-led trade restrictions have prevented them from purchasing the most advanced chips. The goal of US export controls has always been to cripple China’s AI sector. But, as DeepSeek has shown, they have had the opposite effect, spurring precisely the innovations that will enable Chinese firms to challenge American AI oligopolies. Already, DeepSeek’s rise triggered a stock-market selloff of AI-related US companies, not least Nvidia.

This is surely unwelcome news for US President Donald Trump’s administration. Trump has made no secret of his determination to contain China, including by fulfilling his promise to impose a 10% across-the-board import tariff on Chinese goods. And he has heavily courted Silicon Valley bosses – once aligned with the Democratic Party – who have eagerly embraced the prospect of lax regulation.

But that does not mean that DeepSeek’s rise is bad news for the US or the AI industry more broadly. Over the past five years, calls to rein in America’s tech giants have been growing louder. Despite the best efforts of former President Joe Biden’s administration, however, the US Congress has failed to introduce any meaningful legislation on this front. Ironically, thanks to US policies designed to constrain China’s AI ambitions, the US AI sector seems set to get some of the market competition that it so badly needs.

Geopolitics might have contributed to DeepSeek’s rise. But the firm’s disruption of the AI industry is about market – not great-power – competition.

CAMBRIDGE – Thomas Edison, the autodidactic telegraph operator turned entrepreneur, is often considered the greatest inventor of all time, while Nikola Tesla, who worked for an Edison company in Paris before emigrating to the United States, is barely remembered, except through Elon Musk’s electric-vehicle company. Yet it was Tesla’s breakthrough with alternating current (AC), not Edison’s direct current (DC) technology, that made mass electrification affordable. The prohibitive costs of DC would have kept Edison’s urban electrification a plaything of the rich, like many of his other inventions.

Could the Chinese investor Liang Wenfeng’s DeepSeek AI models represent a similar breakthrough in AI, or are they scams like cold fusion and room-temperature superconductivity? And if they are confirmed, should the US treat them as a mortal threat, or as a gift to the world?

Like many transformative technologies, AI had evolved over many decades before OpenAI’s release of ChatGPT in late 2022 triggered the current mania. Better algorithms, complementary devices such as mobile phones, and cheaper, more powerful cloud computing had made the technology’s use widespread but barely noticed. Trial and error had shown where AI could or could not outperform human effort and judgment.

The magical glibness of ChatGPT and other large language models (LLMs) created the illusion that generative AI was a brand-new breakthrough. ChatGPT had a million users within five days of its release, and 300 million weekly users two years later. High-tech behemoths like Microsoft, Meta, and Alphabet placed multibillion-dollar bets on AI products and data centers, quickly forgetting their earlier enthusiasm for virtual and augmented reality.

In 2024, Nvidia, which had invested $2 billion in its Blackwell AI chip, became the world’s most valuable company, with its market capitalization having risen ninefold in two years. Its chief executive, Jensen Huang, predicted that $1 trillion would be invested in data centers using such chips in the next few years. All of this made Apple’s cautious, wait-and-see approach to AI seem quaintly old-fashioned.

Never mind that the new AI did not provide value to end users remotely commensurate with the monumental investment (not to mention its insatiable demand for electricity). Investments continued to grow under the assumption that hyper-scaled data centers would reduce AI costs, and increased use would make the models smarter.

But underneath their shiny new hoods, LLMs, like many of the decades-old AI models, still use pattern recognition and statistical predictions to produce their output, which means that their reliability rests on the future being like the past. This is an important limitation. Humans can imaginatively interpret historical evidence to anticipate what might happen differently in the future; they can also improve their predictions through imaginative discourse with each other. AI algorithms cannot.

But this flaw is not fatal. Since processes that obey the laws of nature are naturally stable, the future is like the past in many ways. Given unambiguous feedback, AI models can be made more reliable through training, and even if the underlying process is unstable – or the feedback ambiguous – statistical predictions can be more cost-effective than human judgment. Wildly off-the-mark ads served up by Google’s or Meta’s algorithms are still superior to advertising blindly. Dictating texts to a mobile phone can produce howlers, but it is still quicker and more convenient than pecking away on a small screen.

By 2022, resourceful innovators had discovered innumerable cases where statistically based AI was good enough or better than alternatives that relied on human judgment. As computer hardware and software improved, cost-effective use cases were bound to expand. But it was delusional to think that LLMs were a great leap forward simply because they could converse like humans. In my personal experience, LLM applications have been worse than useless for doing research, producing summaries, or generating graphics.

Nonetheless, reports of DeepSeek’s prowess have sent shock waves through financial markets. DeepSeek claims to have achieved OpenAI- and Google-quality AI performance using only low-end Nvidia chips and at a fraction of the training and operating costs. If true, demand for high-end AI chips will be lower than anticipated. That is why the DeepSeek news erased about $600 billion from Nvidia’s market capitalization in a single day, as well as hammering the stocks of other semiconductor companies and companies that have invested in data centers or sell electricity to those centers.

To be sure, DeepSeek’s claims may turn out to be inaccurate. Many of Tesla’s claims about his inventions after his AC breakthrough were wildly exaggerated, even fraudulent, and the Soviet propaganda machine routinely fabricated scientific and technological breakthroughs alongside real advances. But frugal, out-of-the-box innovations can be transformative. Just look at Musk’s low-cost reusable rockets. India’s successful Mars mission cost a mere $73 million, less than the budget of the Hollywood sci-fi movie Gravity.

If vindicated, DeepSeek’s technology could be to LLMs what Tesla’s AC inventions were to electrification. While it cannot overcome the unavoidable limitations of backward-looking statistical models, it could make their price performance good enough for wider use. Those developing LLM models will no longer have to depend on subsidies provided by large operators with an interest in locking them in. Less resource-hungry models could reduce demand for data centers or help direct their capacity toward more economically justifiable uses.

What about geopolitics? Last spring, a report by the Bipartisan Senate AI Working Group called for $32 billion in annual “emergency” spending on non-defense AI, supposedly to compete better with China. Venture capitalist Marc Andreessen described the arrival of DeepSeek as “AI’s Sputnik moment.” US President Donald Trump thinks that the Chinese AI model is a “wake-up call for US industries,” which should be “laser-focused on competing to win.” He has announced plans to impose new tariffs on semiconductor imports from China, and his predecessor had imposed export controls on high-end AI chips.

In my book The Venturesome Economy, I argued that seeing transformational advances abroad as a threat to domestic well-being is misguided. Blindly pursuing technological or scientific leadership is a foolish enterprise. What matters most is the willingness and ability of businesses and consumers to develop and use products and technologies stemming from cutting-edge research, wherever it might come from. This is true of DeepSeek’s open-source AI models as well.

Of course, we need to control hostile regimes’ menacing military uses of cutting-edge Western technologies. But this is a different and difficult challenge. If addressing it through export controls were possible, we would have stopped worrying about North Korean or Iranian nuclear weapons a long time ago.