Doing More with Less: DeepSeek Challenges AI Giants

Chinese startup, DeepSeek, brought the two-year-long AI euphoria on Wall Street to an abrupt halt by launching its free app on January 27, 2025. Nvidia lost over half a trillion dollars ($589 billion) in market cap as its stock plummeted 17%. Broadcom, another AI major, was down 17.4%. But that wasn’t the end of it. Stocks of leaders in AI models, such as Microsoft, Alphabet, and Meta shed between 2% and 4% each. The low-power DeepSeek weighed on energy utility firms too, with Vistra losing lost 28.3% in share value and Constellation Energy’s stock declining 20.8%. Benchmark indices, the Nasdaq Composite and S&P 500 Tech, also reflected the market shock, losing 3% and 5.6%, respectively, and marking one of the worst days for the tech sector in four years.

The burning question is – can the Chinese DeepSeek replicate the success in the AI industry that the disruptor TikTok saw in the social media sector?

What Distinguishes DeepSeek?

Chinese innovators have done a commendable job of launching an AI model that performs at par with Open AI’s ChatGPT but at a fraction of the cost. This is despite the sanctions and restrictions on access to top-of-the-line AI chips in China. Let’s dive in.

The total cost of building an AI model is threefold. The cost of creating the algorithms, the cost of training it, and the cost of processing power when people begin using it. What has shaken the world is the astonishingly low cost of creating and running DeepSeek.

An AI model is trained with billions of parameters to process a wide variety of queries and respond to all kinds of prompts. Surprisingly, DeepSeek is trained with fewer parameters than most of the leading AI models but is able to offer comparable results in terms of relevance and accuracy.

The cost of processing includes the cost of energy expended to perform an operation. DeepSeek’s official WeChat account claimed that the DeepSeek-R1 operations are between 20 to 50 times cheaper than those of OpenAI’s o1, depending on the task. Here’s a comparison:

Here, a token is the smallest unit of data a query is broken down into for processing. AI models work with these tokens. The resulting tokens after processing are converted to a response suitable for the user to understand.

Reasoning Prowess

DeepSeek’s R1’s reasoning capabilities are at par with those of OpenAI’s o1, the reasoning model that is accessible only via paid subscription. More importantly, the latter is considered suitable for only “more complex” problems as the free GPT 4 are still the “best option for most prompts.” DeepSeek on the other hand, seems to do it all. So how does it cost less?

DeepSeek uses an 8-bit float instead of the standard 32-bit float, requiring significantly less memory and reads multiple tokens simultaneously. DeepSeek employs the older and less-powerful 50,000 H100 Nvidia GPUs, instead of H800s, which lowers the energy expenditure. Plus, the model is so trained to activate 37 billion parameters per token for efficiency purposes, while GPT 4 and Llama 3.3 activate up to 70 billion of them.

Is DeepSeek a Threat or a Promise for the AI Industry?

DeepSeek has not only intensified the race to AI supremacy but also awakened the world to a new possibility of more affordable, accessible, and profitable AI models. Thanks to DeepSeek’s claims made on WeChat, it became the most downloaded app on Apple’s App Store, challenging leading AI tools, such as Gemini and ChatGPT. There is a high possibility that the download volume was initially driven by the intriguing claims of the app maker. As a marketing strategy, it worked! The initial evaluations support these claims. However, a few concerns remain.

How Sustainable are Free Operations?

Undoubtedly, DeepSeek has created an opportunity for small-scale and low-budget businesses to leverage cutting-edge AI technology. This may democratize the use of artificial intelligence and pave the way for widespread adoption. However, the concern here is – how long can DeepSeek sustain on the free-weight model. If the model can be sustained for the long-term, it is bound to create pressure on AI giants to justify their pricing and investments. Further, there is a risk that lower investments from AI frontrunners could decelerate innovation.

Threat to Data Privacy

A major concern among adopters will be whether the price of the technology would be paid with the loss of data privacy and rights to intellectual property. China is the third most cybercrime-prone country, following Ukraine and Russia. Plus, Microsoft and a few others have accused the Chinese government of aiding cybercrime in the past.

Here’s how you check if Your Brokerage is  Ahead of Cyber Criminals

The word on the street is that R1 alters responses to align with the Chinese Communist Party’s values. This puts DeepSeek’s credibility and trustworthiness under the lens. As of January 29, 2025, the company has received its first complaint letter from the Italian Data Protection Authority (DPA), demanding clarification about server locations and exposure of the data of millions of users.

Should AI Majors Worry?

The current stock market panic may just be a knee-jerk reaction to the unexpected feat pulled by the Chinese startup. Leaders of several tech giants claim that the model could not have been built at a low expenditure of $5.6 million. Notably, DeepSeek has only proven its mettle against benchmark test suites. Its performance, trainability, and accuracy across business-specific use cases remain to be examined.

While establishing a reputation and forging relationships with global companies will remain a challenge for DeepSeek, it has brought the world’s attention to AI models being overpriced. The launch does spur innovation and may encourage developers worldwide to pioneer cheaper AI solutions. Tech majors may have to rethink ROI on AI projects. Microsoft and Meta, both of whom had released their 2025 capex at $80 billion and $60-$65 billion, respectively, may have to accelerate AI monetization.

Such disruptions bring huge volatility to the stock markets. Investors worldwide will be keeping their eye on the latest developments in the AI sector to make investment decisions, especially given the returns this sector has been providing over the past few years. Brokers and fintechs need to be equally disruptive in attracting and retaining clients. Using AI for your content needs might be tempting but can it really serve your needs the best?

Our blog details Which is Better for Your Content Needs – an AI model or a human.

For a detailed analysis of your content needs and a winning strategy, partner with an experienced content writing team. Veda Informatics has over two decades of experience creating impactful content for the finance sector. Speak to our experts to leverage our expertise in creating finance and fintech content.