The Wire Staff 
New Delhi: A generative artificial intelligence platform developed by Chinese maker DeepSeek has seemingly rattled the US-led world of AI, with new president Donald Trump saying that the entry should serve as a “wake-up call” for American AI companies.
DeepSeek’s first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1 are open-source – their codes are available on Github.
DeepSeek’s merits aside, there are multiple reasons why what Trump is saying is not exactly untrue.
What role does it play in the China-US ‘tech’ cold war?
One of the most surprising things about DeepSeek – whose chatbot interface is very similar to erstwhile crowd-favourite, OpenAI’s ChatGPT – is its popularity in the US.
DeepSeek was released on January 20, 2025. By January 27, the AI model powering DeepSeek’s chatbot had began outperforming top US models. Silicon Valley tech funder Marc Andreessen called the release of the model as “AI’s Sputnik moment”.
DeepSeek’s chatbot was the number one product on the Apple App Store in the US. It surpassed OpenAI’s 2023 bot ChatGPT.
This popularity has not only shaken the stock market but has also gone some way towards showing US policymakers that bans on apps like TikTok will not put a dent on Americans’ appetite for Chinese digital services. This is a point for China in what Wired has called the “US-China tech cold war.”
The article notes how in October, 2022, the US government enforced export controls that severely restricted Chinese AI companies from accessing cutting-edge chips like Nvidia’s H100. DeepSeek started out with a stockpile of 10,000 H100s, its founder Liang Wenfeng said in a 2024 interview. But they alone could not fuel a new product. BBC, though, reports that experts believe that this collection could have been up to 50,000.
In response, DeepSeek came up with homegrown methods to train its AI models using a combination of “tricks” by using simpler math (from 32 decimal places down to 8, allegedly), custom communication schemes between chips, and what software engineer Wendy Chang tells Wired is an “innovative use of the mix-of-models approach.”
Combining these successfully is where DeepSeek appears to have scored on over the US momentarily.
What has it meant for the US AI financial world?
Trump has said the shock could spur a “positive” future for US tech companies, as it would force them to innovate more cheaply.
“I’ve been reading about China and some of the companies in China, one in particular coming up with a faster method of AI and much less expensive method, and that’s good because you don’t have to spend as much money. I view that as a positive, as an asset,” Trump said.
DeepSeek’s purported magic lies in the fact that it was reportedly developed for a fraction of the cost of its US rivals.
OpenAI’s Sam Altman wrote on X that the service was “impressive… particularly around what they’re able to deliver for the price”.
Nvidia, the leading supplier of AI chips, was the worst hit and lost close to $ 600 billion in market cap on January 27, in what DW has called the biggest drop for any company on a single day in US history.
Forbes reported that Nvidia, the most valuable company in the world by market capitalisation, fell to third place after Apple and Microsoft on Monday.
In Japan, chip-testing equipment maker Advantest, a supplier to Nvidia, lost 10% on January 28 after diving nearly 9% the day before.
Chip-making equipment maker Tokyo Electron fell 5.3%, while technology start-up investor SoftBank Group was 6% lower.
Over in the US, Broadcom finished down 17.4%, followed by ChatGPT backer Microsoft which fell 2.1% and then Google parent Alphabet which ended down 4.2%.
Dropbox’s AI head Morgan Brown wrote in a series of posts on X how in an “insanely expensive” world of AI model training, DeepSeek is able to do the same training done at a whopping 5% of the cost.
Where the likes of OpenAI and Anthropic spend upwards of $100 million at massive data centers with thousands of $40,000 GPUs, wrote Brown, DeepSeek did it for “$5 million instead.” It also allegedly uses only 2,000 GPUs instead of the usual 100,000 that other AI training companies use.
“Their models match or beat GPT-4 and Claude on many tasks,” Brown wrote.
Veteran analyst Gene Munster told BBC that he was not really convinced of the financials DeepSeek was citing, and wondered if the startup was being subsidised or whether its numbers were correct.