The DeepSeek Storm

[paserbyp]

Stocks took a dive yesterday, dragged down by AI chipmakers after a Chinese startup’s cheaper high-performing model cast doubt on how much US companies are spending to build AI (more on that in a sec). Nvidia shed $589 billion in market value—the biggest one-day loss ever in the US stock market.

The tech-heavy Nasdaq 100 lost almost $1 trillion in value yesterday after DeepSeek’s release of its low-cost R1 “reasoning” model led to the realization that the US tech giants pouring billions into AI could get left in the dust by a scrappy Chinese competitor.

The free and open-source R1 model is performing on par with OpenAI’s o1 reasoning model, propelling DeepSeek to become the most downloaded free app for iPhone.

But just like the student who claims they got the top test score in the class after cramming for 15 minutes, DeepSeek says the model cost only $6 million to build—as little as 3% of o1’s estimated training price tag.

R1 comes on the heels of DeepSeek’s recent release of its V3 model that matches OpenAI’s GPT-4o. DeepSeek’s products are reportedly powered by Nvidia chips that are less advanced than those used by rivals, due to US semiconductor export restrictions meant to preserve America’s AI edge. DeepSeek’s apparent ability to make do with less had shares of AI chip lord Nvidia nosediving 17% yesterday. Meanwhile, Alphabet and Microsoft stock dipped over 4% and 2% respectively.

Goldman Sachs believes DeepSeek’s success is emblematic of AI upstarts’ potential to challenge Big Tech through the inventive use of resources, though many experts say American companies’ access to the most advanced chips still gives them a major edge.

Meanwhile, some observers are skeptical of DeepSeek’s claims that it achieved a Rocky-esque feat of low-resource AI training. Bernstein analysts dubbed the stock selloff a “Twitterverse panic.” While acknowledging that DeepSeek did use computing power economically, in a reassuring take for Nvidia bulls, they also noted that even if AI development becomes more efficient, demand for it could skyrocket, boosting appetites for computing power.

Tesla (TSLA) CEO Elon Musk seemed to agree with CEO of Scale AI Alexandr Wang suggesting that China's DeepSeek has about 50,000 H100 Nvidia's (NASDAQ:NVDA) chips, which they can't talk about due to U.S. export controls.

"My understanding is that Deepseek has about 50,000 H100s, which they can't talk about obviously because it is against the export controls that the U.S. has put in place. I think it is true that they have more chips that people than other people expect. But on a go forward basis, they are going to be limited by the chip controls and the export controls that we have in place." Wang said in a Jan. 23 interview with CNBC.

Musk replied "Obviously" to a post from a user who had posted Wang's interview on his feed on Jan. 24.

On Monday, DeepSeek restricted registration to China mobile phone numbers.

Moreover, some analysts believe the Trump Administration could further expand restrictions on AI chip exports to China, in an effort to hamper AI progress in the country.

Two years ago, OpenAI’s ChatGPT launched a new wave of AI disruption that left the tech industry reassessing its future. Now within the space of a week a small Chinese startup called DeepSeek appears to have pulled off a similar coup, this time at OpenAI’s expense.

Nevertheless, DeepSeek’s sudden success — the company’s free chatbot mobile app quickly surpassed even ChatGPT for downloads on Apple’s App Store — has prompted questions. Is the DeepSeek story too good to be true? And should businesses in the US and allied countries allow employees to use an app when the company’s Chinese background and operation are so opaque?

The DeepSeek storm hit on January 20 when DeepSeek launched its R1 LLM model to the public, complete with big claims around performance.

Using smaller “distilled” LLM models, which require significantly less processing power while replicating the capability of larger models, DeepSeek’s R1 matched or exceeded OpenAI’s equivalent, o1-mini, in important math and reasoning tests.

That performance generated a surge of interest. By Monday the DeepSeek app had overtaken ChatGPT and Temu to become the iPhone App Store’s top free download — and DeepSeek was reporting delays in new registrations to use the app due to what it described as “large-scale malicious attacks” on its services.

Nobody saw this coming. Somehow, R1 was doing this with less hardware. Moreover, DeepSeek-R1 is available through an open-source MIT license, which allows for unrestricted commercial use, including modification and distribution.

With AI sector share prices unsettled by all of this, the implication is that perhaps usable models don’t need the huge chip clusters deployed by the established players and organizations shouldn’t be paying high prices to access them.

Furthermore, if a tiny startup can get by on more limited hardware while training LLMs for a fraction of the cost, perhaps strenuous US attempts to limit the export of the most powerful AI chips to most of the world including China, are already obsolete before they’ve been fully implemented.

The speed of DeepSeek’s rise is a case of ‘zero-day disruption.’ Organizations have no time to react, and not just because developers across the world have piled in to test DeepSeek-R1 via its API by the thousand. Releasing a free app gives this capability to everyone, including employees who might enter sensitive data into it. By now, DeepSeek is everywhere, which makes it difficult to control.

“The app has raced to the top of the app charts, but I would advise anyone considering installing it and using it to exercise some caution,” warned tech commentator, Graham Cluley.

That said, organizations should already be used to coping with this issue. “Human nature being what it is, there will surely be just as much sensitive data entered into DeepSeek as we’ve seen entered into every other AI out there,” said Cluley. Organizations should probably hold back until it has been more thoroughly audited in the same way they would with any new app.

Or perhaps focusing on the risks is too negative. DeepSeek will ignite more competition in the sector, potentially turning powerful LLMs from an expensive service for the deep pocketed into a cheap utility anyone can access. Rather than dumping existing AI services, organizations should demand a better deal while avoiding becoming too locked into one LLM as new innovations appear.

A lurking possibility is that DeepSeek isn’t as good as it seems, with some skepticism already appearing around its price-performance claims. Stacy Rasgon, a senior analyst at Bernstein Research, questioned DeepSeek’s underlying costs.

“Did DeepSeek really build OpenAI for $5M? Of course not,” he wrote in a client note. “The oft quoted $5M number is calculated by assuming a $2/GPU-hour rental price for this infrastructure, which is fine, but not really what they did, and does not include all the other costs associated with prior research and experiments on architectures, algorithms, or data.”

In use, DeepSeek makes elementary errors, not dissimilar to the ones that afflicted ChatGPT in its early days. Some of its responses also underline that the app imposes guard rails when run from a Chinese host. A good example is this report of its refusal to acknowledge the Tiananmen Square massacre, something the Chinese Government goes to extreme lengths to hide.

In the short term, DeepSeek’s appearance underlines the unstable nature of AI itself. Tech is used to periodic disruptions. AI suggests that these might become more routine, including of its own capabilities. It is unlikely to be the last such breakthrough in a sector that will prove harder to dominate than has been assumed.

Investors and government regulators trying to control AI development won’t like this but if it offers cheaper and earlier AI access across the economy it could still work as a net positive. According to Cluley, DeepSeek should be something for Silicon Valley to worry about.

“If it’s accurate that the Chinese have been able to develop a competitive AI that massively undercuts the US-based giants in terms of development cost and with a fraction of the hardware commitment then that is clearly going to upset the applecart and have a tech billionaire or two crying into their Cheerios this morning,” he said.