At the prospect of a paradigm-shifting AI model, the markets reacted violently. DeepSeek precipitated a widespread selloff in the tech sector on Monday. Stocks for Silicon Valley’s biggest players like Alphabet, Amazon, Meta, and Microsoft tumbled. Across the board, chipmakers such as Broadcom and Micron Technology plummeted. Nvidia’s stock fell 17% yesterday. The chipmaker shed $600 billion in market cap—the largest single-day drop ever. 

Monday’s meltdown was a cold plunge on what had otherwise been a bull market for tech on the belief that AI would usher in a new golden age of worldwide productivity and sector-wide returns. But DeepSeek’s release threw a wrench in the works—even if only temporarily—because when new information comes to light that could disprove virtually all the assumptions underpinning a given investment thesis, investors do what is rational: They exit that investment. 

One day later, investors were convinced the market would rebound after Monday’s thrashing. 

“We view DeepSeek’s release as part of an ongoing evolution, not revolution, and think that this market reaction is largely overdone,” Jefferies wrote in a note to investors on Monday.  

Across Wall Street that seemed to be the prevailing sentiment. Long-term optimism on artificial intelligence remained, resting largely on the view that hundreds of billions in investment into the technology would continue. DeepSeek’s innovative training approach, investors believe, could expand the market for AI to new enterprise customers. At the same time, the U.S. will shovel money toward government-funded AI research in an effort to maintain its lead over China. That’s to say nothing of the headwinds DeepSeek could now face as its new-kid-on-the-block hype fades to be replaced by the same level of scrutiny its more established competitors regularly face. 

DeepSeek will grow the overall size of the AI market

More than anything, DeepSeek seemed to offer evidence that there is a way to more efficiently develop AI models. DeepSeek’s training and inference costs were one-hundredth those of a recent OpenAI model, and one-tenth those of Meta’s Llama 3, according to data cited in an analyst note from Mizuho. 

The previous conventional thinking had been that training AI systems required inordinate amounts of compute and energy. That rationale drove hyperscalers like Meta and OpenAI to invest billions in data centers to power their research efforts. Firms like Nvidia invested just as much into developing chips that would deliver the compute their largest customers needed. Meanwhile, cloud companies like Amazon and Microsoft devoted their financial resources to expanding AI infrastructure. But Wall Street sees the new hyperefficient school of AI development embodied by DeepSeek as good news for megacap tech stocks. 

“On the flip side, a more efficient training, lower token costs could imply AI can now be made available to a larger total addressable market, a bigger audience, and hence can drive more scale if it drives significant efficiency, productivity, and value,” a Mizuho analyst wrote.

That rationale hinges on the idea that as costs to develop AI drop, more businesses will be able to afford it, thus creating new customers for semiconductor and cloud companies. In that scenario, the semiconductor companies whose stocks fell on Monday would expand their customer base beyond Big Tech firms like Alphabet, Amazon, and Meta. 

Bull investors are also relying on the hyperscalers to maintain the existing levels of capital expenditures they have already committed to AI, under the assumption they would welcome the opportunity to get more bang for their buck as they spend it. 

“Initially this ‘more can be done with less’ approach may seem like a net negative, but we view this efficiency gain as actually serving as an accelerant for AI workloads,” Raymond James analyst Ed Mills wrote in an analyst note. “In this context, we view DeepSeek’s innovations as suggesting that efficiencies in model training and deployment could ultimately drive greater demand for compute resources as AI applications become more widespread and accessible.”

The added efficiency could also lower overall costs for the hyperscalers. In that case, they could reach a positive return on invested capital for their AI capex more quickly, according to Raymond James.  

DeepSeek will kick-start U.S. government investment in AI

On Monday President Donald Trump referred to DeepSeek as a “wake-up call” for the U.S tech industry, which, he said, should be “laser-focused on competing.”

Wall Street believes DeepSeek’s success will spur the U.S. government to funnel more investment toward AI in an effort to ensure American companies don’t fall behind their Chinese counterparts.  

“In the near term, DeepSeek’s achievement is likely to pressure the U.S. into increased support for domestic AI development, most likely leading to increased federal investment in AI research and infrastructure,” wrote Mills of Raymond James. The firm added that it could envision the U.S. passing an AI version of the CHIPS and Science Act.  

The government is likely to dedicate additional funding toward military and defense applications of AI in the wake of DeepSeek’s success, according to Evercore ISI. 

In the earliest days of his second term in office, Trump has already announced a major $500 billion public-private initiative, dubbed the Stargate Project, in collaboration with OpenAI, Oracle, and SoftBank. The plan will see the three companies build a sprawling network of data centers to power AI development across the country. Trump’s announcement of the project included an explicit reference to the ongoing AI race with China. 

“It’ll ensure the future of technology,” Trump said last week when announcing Stargate. “What we want to do is keep it in this country. China is a competitor, and others are competitors.”

After the release of DeepSeek, Stargate could see its focus expand to include research on training efficiency, according to Raymond James. “There may be a reevaluation of how these resources are allocated, with a potential shift towards funding more diverse and innovative approaches to AI development,” the firm wrote. 

Investors wonder if DeepSeek is all it’s cracked up to be

While DeepSeek undoubtedly presented a new methodology for training AI models that splits large language models into multiple specialized versions, its work is already being dissected and parsed by Silicon Valley’s brightest. Investors, too, scrutinized the upstart model carefully in order to understand exactly what sort of impact it could have on the rest of the industry. 

So far the view is that the initial excitement won’t last. “DeepSeek is high in novelty claims, but time will tell whether it creates a disruption in current AI models and AI architectures—we strongly suspect it does not,” Baird senior research analyst Tristan Gerra wrote on Monday. 

DeepSeek’s model has outperformed those of some U.S. firms on inference, which is the process by which already trained models make predictions. However, the exact nature of the cost associated with developing its new model is disputed. DeepSeek said it spent $6 million to train its new model. Most analysts believe the number is far higher. That’s in large part because DeepSeek built some of its work on open-source models from OpenAI and Meta. Because DeepSeek’s model was built on existing research, the costs between its work and a ground-up model are “not remotely comparable,” Baird’s Gerra wrote.  

It is also not clear how many chips and how much computing power DeepSeek uses or where it got them. Alexandr Wang, CEO of Scale AI and the world’s youngest self-made billionaire, said DeepSeek had Nvidia’s H100 chips, which it wasn’t supposed to have. Those are Nvidia’s most powerful chips and are subject to export controls that ban their sale in China. “My understanding is that DeepSeek has 50,000 H100 chips, which they can’t talk about obviously because it’s against the export controls the United States has put in place,” Wang told CNBC last week. 

There is also the added fact that if DeepSeek’s new methods are as efficient as they seem to be, then there is no reason its competitors won’t be able to copy them. 

As Oppenheimer analysts noted: “Regardless of cost and/or hardware used, DeepSeek appears to have introduced some innovative AI inference techniques which cloud service providers and hyperscalers are likely to adopt and internalize if proven out.”