DeepSeek stirs intrigue and doubt across the tech world

China's AI disruptor rattles industry watchers with unproven claims

In a busy week for GenAI, the tech industry is weighing the impact of the latest interloper on the LLM scene. China's DeepSeek shocked stock markets on Monday, slashing $600 billion off the value of erstwhile AI golden child Nvidia.

As the dust settles and the markets recover, the industry is questioning the implications and whether China's apparent AI renaissance is quite what it seems.

A measure of the interest comes from an investment analyst quizzing ServiceNow CEO Bill McDermott, not about his company's performance, but about how he views DeepSeek as ServiceNow bakes LLMs into its workflow solutions.

He said: "these models are being commoditized at a rapid rate and probably more rapid than anyone could have dreamed of," the Big Mc told analysts following the company's Q4 results.

"It's exciting for platform and app vendors like us since the competitive differentiation will happen at our level in terms of the applications, the business processes ... and then ultimately get a business outcome."

He was not only thinking of the benefit to customers but also eyeing the potential for "the gross margin benefit" of significantly cheaper LLMs for ServiceNow, McDermott told the crowd.

Howver the CEO also added a note of caution. "As it relates to DeepSeek, I would just say we have a deep belief [in] measuring twice if not three times and cutting once. We believe very strongly in understanding what's really going on, and we're highly committed to responsible AI." He said ServiceNow teams were looking into the implications of DeepSeek's arrival.

The reason for McDermott's enthusiasm centers on DeepSeek's claims that it has built its V3 and latest R1 models efficiently and relatively cheaply. While Western competitors have poured tens of billions of dollars into training their LLMs, DeepSeek is said to have achieved comparable performance with V3 – an LLM R1 is based on – that might have cost less than $6 million to train in the cloud.

But that dollar estimate was calculated by DeepSeek multiplying the number of Nvidia H800 GPU hours it spent training V3 by a typical cloud rental rate of $2 an hour (see page five of its paper); in reality, DeepSeek built its models using thousands of GPUs it owns and spent many millions on. One independent analysis suggests DeepSeek blew $1.6 billion on AI hardware. Then there's all the research and development costs that need to be factored in.

Yes, V3 and R1 didn't cost zillions of dollars to build – the sort of money Silicon Valley likes to drum up – but it wasn't a trivial outlay, either.

As for cloud API pricing, the DeepSeek R1 reasoning model seems cheaper to use than OpenAI's o1. Still, some doubt the Chinese lab's claims.

Firstly, there is the idea that it is a decent open source alternative to other vendor-provided LLMs. While some US vendors make claims about their open source credentials, academic research has found they are not really so open.

Similarly, DeepSeek's open source claims merit scrutiny. Systems integrator and IT services vendor Version1 has pointed out that while DeepSeek has released model weights, but it has not revealed its training data nor training code, so it cannot be reproduced.

In a Reddit discussion group dedicated to OpenAI, a netizen shared evidence suggesting DeepSeek R1 could have trained its models on benchmark answers, thereby potentially gaming the market without producing a general-purpose model.

Users claimed some answers show a 50-90 percent similarity to benchmark answers. Others in the group cast doubt on whether it is possible to get anything like a general-purpose LLM using the approach.

Whether or not its performance and open source claims stack up, DeepSeek has also taken a hit on its security credentials after it left a backend database wide open, and had to limited new signups for the web-based interface to its models due to what's said to be an ongoing cyberattack.

"Due to large-scale malicious attacks on DeepSeek's services, we are temporarily limiting registrations to ensure continued service," the biz said in a note on its status page.

Regardless of the impact on stock valuations, tech users are bound to have more questions as China continues its push into Western software markets. ®