DeepSeek Means Companies Need To Consider AI Investment More Carefully

Analysis The shockwave following the release of competitive AI models from Chinese startup DeepSeek has led many to question the assumption that throwing ever more money at costly large-scale GPU-based infrastructure delivers the best results.

As The Register reported earlier, shares of some of the largest American tech brands in the AI boom tumbled following the debut of the DeepSeek R1 model, which is said to perform favorably against those from OpenAI and Meta and was trained using fewer Nvidia GPUs.

The China-based company's claims that DeepSeek's performance is on par with the best existing models and that it cost less than $6 million to train are also unverified

The move called into question the assumption that spending billions on datacenter infrastructure in a race to build larger and more complex models is the way forward if China can do this with limited supplies of older hardware. Nvidia, which has enjoyed record profits from its GPU accelerators for AI, lost almost $600 billion off its market valuation in a day.

The hysteria comes on top of a growing unease that more investment is being funneled into AI development and the infrastructure to support it, with little return to be seen so far.

However the inital panic may have been misplaced as the freefall in US tech shares soon halted, and experts pointed out that DeepSeek appears to have used output from existing models developed by Anthropic and OpenAI in its training. The China-based company's claims that DeepSeek's performance is on par with the best existing models and that it cost less than $6 million to train are also unverified.

"I believe the concerns regarding DeepSeek's innovations are highly overblown," Omdia's Principal Analyst for Datacenter IT, Manoj Sukumaran, told The Register.

"There is no doubt that there are some ingenious innovations in DeepSeek's model pre-training, like the use of reinforcement learning as a core training methodology, moving away from the reliance on large labeled datasets, sparse activation of model parameters, and adaptive routing to select the expert models to work," he said.

But these innovations are essential to make GenAI accessible to more users, Sukumaran added, and will instead hasten user adoption of this technology.

As far as the infrastructure to power all this goes, Sukumaran says that massive AI buildouts are likely to continue.

"The AI inference market is just unfolding and it will grow significantly over the next several years. Omdia has estimated that the number of servers shipped each year for AI inference will increase at a 17 percent CAGR out to 2028," he added.

Nevertheless, Taiwan-based research operation TrendForce says that it expects to see organizations conduct more rigorous evaluations of AI infrastructure investments in future, and focus on adopting more efficient models to reduce reliance on hardware such as GPUs.

The analyst also envisages growth in the adoption of infrastructure using custom ASICs (application-specific integrated circuits) to lower deployment costs, and that demand for GPU-based products could see "notable changes" from 2025 onward.

"Historically, the AI industry has relied on scaling models, increasing data volume, and enhancing hardware performance for growth. However, escalating costs and efficiency challenges have prompted a shift in strategy," TrendForce says. "DeepSeek has adopted model distillation techniques to compress large models, improve inference speed, and reduce hardware dependencies."

• Microsoft catapults DeepSeek R1 into Azure AI Foundry, GitHub
• DeepSeek stirs intrigue and doubt across the tech world
• Microsoft talks up 'significant capital investments' in AI as sector reacts to DeepSeek
• Guess who left a database wide open, exposing chat logs, API keys, and more? Yup, DeepSeek

Earlier this week, IBM CEO Arvind Krishna said he saw in DeepSeek some validation for his own company's approach to AI.

"We have been very vocal for about a year that smaller models and more reasonable training times are going to be essential for enterprise deployment of large language models. We have been down that journey ourselves for more than a year," Krishna claimed during the company's recent earnings call.

"We see as much as 30 times reduction in inference costs using these approaches. As other people begin to follow that route, we think that this is incredibly good for our enterprise clients. And we will certainly take advantage of that in our business, but I believe that others will also follow that route."

In a note on the implications of DeepSeek issued by Gartner, the analyst said efficient scaling of AI will in future be more important than how much compute can be assembled to build it.

"DeepSeek-engineered systems combine models, frameworks, and underlying infrastructures to more effectively utilize infrastructure resources. This results in lower costs while delivering efficiencies," it observed.

However, said that the Chinese AI doesn't set a new state of the art for model performance as it often matches but doesn't surpass existing models.

As far as infrastructure goes, Gartner says that "it's not proof that scaling models via additional compute and data doesn't matter, but that it pays off to scale a more efficient model."

The takeaway is that DeepSeek isn't going to suddenly lead to a drastic drop in demand for AI infrastructure, so Nvidia investors and those pumping money into datacenters can perhaps rest a bit easier. Neither is it the harbinger of the AI bubble bursting that some anticipate.

Instead, it serves as a reminder that things can always be done better, and that just throwing money and resources at a problem is not always the best way to solve it.

"DeepSeek's superior price-to-performance ratio serves as a reality check for the AI industry, particularly US companies and their venture capital backers," said Neil Roseman, CEO of security firm Invicti. "While companies make massive bets on AI, current results don't justify these investments. Success will come from efficient, focused development addressing genuine needs." ®