China once again surprised the IT industry with the unveiling of a new company. This time it wasn’t a new social media app. Instead, it was an AI startup. The company, DeepSeek, came on to the scene seemingly out of the blue and caused a lot of emotions in the industry.
Most of them were in the area of policy, user experience, and overall geopolitics. But DeepSeek also highlighted a potential major weakness in the current top AI algorithms and their ways of handling and consuming resources. DeepSeek now has the potential to change even the industry’s approach to data centers thus causing a major shift in current investments and goals. Can that really happen? If DeepSeek’s claims end up being sustained over a longer period of time – then yes, we could expect substantial changes in the way AI is developed and deployed.
In order to see what could happen, let’s take a journey to explore what DeepSeek is, why it caused all these emotions and could it really be that major change to the entire AI sector. And of course – how will all of that reflect on data centers.
What is DeepSeek?
DeepSeek is a Chinese AI startup. In January 2025 it released its first public AI model – R1. The company stated that R1 rivals and even beats OpenAI’s ChatGPT in most of its capabilities. Not only that – it is able to achieve those results with substantially less resources for training and overall computing power. Consuming massive resources has been one of the main issues with most AI models – and it’s the main reason for the massive boom in data center demand over the past two or three years.
DeepSeek’s R1though shattered the concept of AI being resource-intensive. The algorithm is aimed at OpenAI’s o1 – its current best reasoning model. But while OpenAI says it spent “over 100 million dollars” to train its GPT-4 model (an earlier and not so advanced model compared to o1), R1 needed just 6 million dollars to be trained to o1’s level. Such a massive difference in costs caused a massive, “earthquake” in the AI industry. Has the Western IT world been on the wrong track the whole time? Did it start developing AI in a way that’s not sustainable? Or has DeepSeek used some trickery to achieve these results?
The Chinese company as a simple answer: Unlike o1 and other proprietary models, R1 is fully open-sourced. It’s based on some of Meta’s AI models, which have been mixed with other methods and has been fined-tuned with GRPO (Group Relative Policy Optimization). The question is what data has it been using? OpenAI initially claimed DeepSeek had “stolen” their data they used to train their models. DeepSeek’s supporters though retaliated R1 is trained with publicly accessible data and OpenAI shouldn’t be the one holding other companies accountable for what data their use, considering their own controversial history.
The DeepSeek fallout
The spat about what data is being used was just a small part of the massive fallout DeepSeek had over the IT industry. Once R1 was available in a chatbot form via an app and thus accessible to users and businesses, it became seemingly an overnight success. DeepSeek’s app rose to the top of the Apple App Store free apps chart and overtook ChatGPT. The app raked in millions of downloads.
Once the news that its R1 model is so cheap to train and deploy hit the mainstream media, that caused a huge stock market slide for a lot of major Western IT publicly traded companies. Nvidia took the heaviest hit, losing about 600 billion dollars of its market value – the largest single drop ever recorded. It pulled the company from the top as the world’s most expensive organization to third place behind Apple and Microsoft. Even the U.S. President, Donald Trump, commented on the event saying DeepSeek is a “wake-up call” for U.S. companies.
Why did Nvidia take the hit? Because it rose to the top thanks to its AI processors which became the industry’s benchmark and are the most desirable hardware for AI data center operators. And DeepSeek can’t use the latest generation of those chips due to the US sanctions over China. As a result, the Chinese startup decided to use whatever hardware it had (it was also accused of circumventing the sanctions by stockpiling Nvidia’s H100 chips, although that can’t be proven). Instead, the company focused on optimizing its software and algorithms and partnered with local Chinese firms to further streamline R1.
As a result, DeepSeek managed to rattle the world’s most powerful companies with a fraction of their budget. Even after the stock market slump, Nvidia’s value is still a massive 2.9 trillion dollars. DeepSeeks main competitor, OpenAI, is valued at about 157 billion dollars and has over 13 billion dollars investments from Microsoft alone. By comparison, DeepSeek has reportedly raised just under 1 billion dollars in total and its estimated valuation is less than 10 billion dollars.
It’s difficult to know the exact figures as DeepSeek is currently a private company. It was founded in December 2023 by Liang Wengfeng who has degrees in electronic information engineering and computer science, but also has experience in finance and is even a CEO of a hedge fund.
Naturally, he’s now in big demand from the media. Asked why DeepSeek seems to have surprised the Silicon Valley, Liang says as quoted by the BBC: “Their surprise stems from seeing a Chinese company join their game as an innovator, not just a follower – which is what most Chinese firms are accustomed to.”
Can we trust DeepSeek?
Liang’s confidence wasn’t well received in the West. Several countries announced restrictions for the usage of the company’s algorithms. Australia banned it on government devices and systems, saying it’s a national security risk. Italy blocked it in full and even ordered the company to stop processing the personal information of its citizens. The U.S. is also exploring limitations.
Then Cisco released a study, claiming R1 has failed all safety tests. The research was done along with the University of Pennsylvania and checked if DeepSeek’s algorithm can handle itself against abuse attempts, harmful prompts, misinformation, cybercrime, etc. The researchers say DeepSeek responded to all their prompt attacks and didn’t deny any requests. For comparison, OpenAI’s o1 responded to just 26% of the harmful prompts, while Claude 3.5 Sonnet from Anthropic scored 36%. The security platform Enkrypt AI also claimed DeepSeeks is on average 11 times more likely to generate harmful outputs compared to o1.
Despite that, Amazon, Nvidia, and Microsoft all announced support for DeepSeek’s algorithms on their hardware and infrastructure.
The impact on data centers
According to Newmark, by 2030, the data center industry will double in size to accommodate the rise of AI. But that estimate didn’t factor a disruptive change like DeepSeek. Whether it’s a perfect algorithm or not, R1 showed the data center world a lot can be done with much less.
R1’s investments seem modest compared to what happens in the West. According to CNBC, OpenAI is negotiating a new investment round of around $40 billion. Data center operators are also rushing to get investments in the hundreds of millions to the single digit billion figures. And Apollo Global Management is forecasting that in order to cover all of the demand for data centers from now up to 2030, an additional $2 trillion of investments will be needed.
This should be great news for investors and companies as it means massive opportunities ahead. But DeepSeek made the industry take a step back and rethink its approach. Yes, even if R1 is indeed so great with resources, it will still need massive data centers. Especially to cover all of the gigantic demand expected for AI – from users, businesses, organizations, and institutions all over the world.
“The market is still exploring the impact of the new AI models, but it could be a healthy correction, Even if it has an impact on demand for data centers for AI, there is still growth expected, it just may be slightly lower growth,” says Mr Timo Buijs, senior director of project and infrastructure finance at ABN Amro Bank to The Straits Times.
Others are also looking on the positive side. “If the reports on DeepSeek are accurate, this will only push AI innovation forward. Innovation in AI doesn’t reduce demand — it fuels it. As AI becomes more accessible and cost-effective, the industry will see continued expansion, maintaining the need for high-performance data center infrastructure,” says Mitch Lenzi, vice president of sales and operations at Baxtel, to DataCenterKnowledge.
One possible opportunity is that “lighter” AI models will open the door for modular and Edge data centers. They are now seen as a helper to the main facilities, but if AI models can run at full capacity on smaller data center footprints and with less resources, this will make it much easier to make Edge data centers and deploy them closer to the data source – ideal for the low latency demands of AI.
So, while initially it may seem that DeepSeek disrupts the data center boom, it can be beneficial to it. The company can inspire the industry to correct its course and use the investments even better, focusing on actual solutions, instead of just going the “more is more” approach and hoping for the best while risking to eventually end up with massive facilities that aren’t being utilized properly.