Developers from top American AI companies are expressing admiration for the emerging DeepSeek AI models, which have gained significant attention recently. At the same time, they’re questioning the belief that their multimillion-dollar technologies have been outperformed by a low-cost alternative from a Chinese startup.

On Monday, Chinese startup DeepSeek triggered a stock market decline, with its free AI assistant surpassing OpenAI’s ChatGPT as the leading app on Apple’s App Store in the United States. DeepSeek claims it developed its model using Nvidia’s H800 chips, known for their lower capabilities, at a cost of less than $6 million.

As concerns about rising competition echo throughout the U.S. stock market, some AI specialists praised DeepSeek’s skilled team and current research but seemed largely unaffected by the news. These insights came from individuals at four leading AI laboratories who chose to remain anonymous as they weren’t permitted to speak publicly.

OpenAI CEO Sam Altman acknowledged on X that R1, one of the several models released recently by DeepSeek, is “impressive, particularly regarding its performance at such a low price.” Nvidia also stated that DeepSeek’s success highlights the demand for more of its chips.

In response to a surge of customer interest, software company Snowflake has decided to include DeepSeek models in its AI model marketplace.

Snowflake’s executive vice president of product, Christian Kleinerman, mentioned that while employees have praised DeepSeek’s models as “incredible,” the company carefully weighed potential risks associated with utilizing AI technology developed in China before ultimately offering it to clients.

“We determined that as long as our customers are informed, we see no issues supporting it,” he said.

As the U.S. AI sector races to evaluate DeepSeek’s V3 model, it’s worth noting that DeepSeek published a research paper in December explaining the model behind its widely-used application. However, the document leaves several questions unanswered, such as the total cost of development.

A source indicated that China has significantly narrowed the gap in AI model development, now lagging by only six months compared to the United States, down from 18 months. Nevertheless, with the excitement generated by DeepSeek’s free model offerings, there are concerns that the company might soon struggle to secure enough chips to satisfy demand.

It’s important to recognize that DeepSeek’s success isn’t just about its $6 million budget, which is minor compared to the projected $250 billion that large U.S. cloud companies are expected to invest in AI infrastructure this year. The research paper clarified that this figure pertains specifically to the chip utilization during the final training run, not the overall development costs.

Experts from two prominent AI labs informed Reuters that the expenses incurred for designing that training run could be significantly more substantial. They highlighted that the training run’s cost is merely the beginning of a much larger financial undertaking.

The report indicated that the V3 training was executed using 2,048 of Nvidia’s H800 chips, designed to align with U.S. export regulations set in 2022, which sources believe would have minimal impact on China’s AI advancement.

Two AI labs speculated that earlier development stages probably required a much larger number of chips, with one individual estimating that such an investment might exceed $1 billion.

Despite these concerns, some American AI leaders praised DeepSeek’s choice to make its models open source, allowing other companies or individuals to access and modify them freely.

Venture capitalist Marc Andreessen remarked on X, “DeepSeek R1 is one of the most remarkable breakthroughs I’ve encountered – and being open source, it’s a tremendous gift to society.”

The recognition received by DeepSeek’s models highlights the potential of open-source AI technology as a viable alternative to expensive and closely monitored systems like OpenAI’s ChatGPT, according to industry analysts.

In recent years, the most valuable companies on Wall Street have thrived on the belief that they alone possess the extensive capital and computational resources necessary to create and expand new AI technologies. This week and next, as many American tech giants prepare to release their quarterly earnings reports, those assumptions might face increased scrutiny.

Reporting by Anna Tong, Jeffrey Dastin and Kenrick Cai in San Francisco and Katie Paul in New York; Edited by Noel Randewich and Christopher Cushing