DeepSeek, a Chinese AI model, has rocked the AI world, cutting hundreds of billions of dollars off market values of US AI companies, and defying the widely-held belief that the US was winning the race for artificial general intelligence (AGI). Its latest model uses a fraction of the power of America’s leading products, and has shot to the top of app download charts.
DeepSeek also raises questions about Washington’s efforts to contain Beijing’s push for tech supremacy, given that one of its key restrictions has been a ban on the export of advanced chips to China.
How DeepSeek Achieved More With Less
AI development has long been constrained by its insatiable demand for computing power. State-of-the-art models like OpenAI’s ChatGPT have required billions of dollars’ worth of GPU clusters and an enormous energy footprint. Writing a 100-word email with ChatGPT consumes around 500ml of water, and uses 140Wh of energy, enough for 7 full charges of an iPhone Pro Max. DeepSeek, however, says it has trained an AI model comparable to the leading models from heavyweights like OpenAI, Meta, and Anthropic, but at an 11x reduction in the amount of GPU computing, and thus cost. DeepSeek’s training may have cost as little as $5.6m (small change in the US AI world). In the last year alone, OpenAI has spent $5bn.
DeepSeek’s key innovation lies in its Mixture of Experts (MoE) architecture, which allows the model to selectively activate only the most relevant subsets of its 671 billion parameters for any given task. Instead of overloading GPUs with full model computations, DeepSeek dynamically allocates resources, requiring fewer active parameters at any given time. This efficiency enabled the company to train DeepSeek on 14.8 trillion tokens using only 2.66 million GPU-hours on Nvidia’s H800 accelerators—a striking contrast to OpenAI’s ChatGPT-4, which reportedly required over 10 million GPU-hours on A100/H100 GPUs for training.
In an interview last year, DeepSeek CEO Liang Wenfeng said: “Money has never been the problem for us; bans on shipments of advanced chips are the problem.”
U.S. export controls are designed to limit China’s access to advanced AI chips, however DeepSeek’s breakthroughs may have demonstrated a flaw in the sanctions. To create R1, DeepSeek had to rework its training process to reduce the strain on its GPUs, a variety released by Nvidia for the Chinese market that have their performance capped at half the speed of its top products. Rather than weakening China’s AI capabilities, the sanctions appear to be driving startups like DeepSeek to innovate in ways that prioritise efficiency, resource-pooling, and collaboration.
The implications of this efficiency are profound. If DeepSeek’s new MoE approach to AI is legitimate and becomes widely adopted, the AI industry may see a decline in demand for the most expensive GPUs, leading to more cost-effective AI development and broader accessibility. This revelation had immediate consequences for the market.
Market Fallout: Nvidia’s $600 Billion Wipeout
Investors were quick to react to the disruptive potential of DeepSeek’s architecture. As news spread that AI development might no longer require massive GPU clusters, Nvidia’s stock plummeted nearly 18% in a single day, erasing a staggering $600 billion in market capitalisation, the biggest drop in the history of the US stock market. This market shock was a direct response to concerns that companies could reduce their dependency on high-performance chips, threatening Nvidia’s dominance in AI hardware.
The panic wasn’t limited to Nvidia. Other AI-related stocks also faced sharp declines, including those of cloud providers and semiconductor manufacturers that supply the AI industry. The energy sector also took a hit, as the predictions that AI could drive a surge in energy usage were knocked by DeepSeek’s efficient model.
The biggest impact of DeepSeek is the questions now being asked of the US AI sector. Many believed that the future of AI would be dominated by the US. Donald Trump announced “the largest AI infrastructure project by far in history” just last week, pledging $500bn of investment in AI infrastructure from US firms like OpenAI, to maintain America’s grip on the sector. The US has been caught off-guard by this leap forward from China, and will now have to respond.
Suffering From Success: DeepSeek Targeted By Cyber Attacks
DeepSeek reached the top of Apple’s Top Free Apps chart in the U.K. and the U.S. this week, dethroning OpenAI’s ChatGPT. Its surge in popularity has not come without challenges. The app has been plagued by cyber attacks, forcing the company to suspend new user registrations temporarily.
New users trying to sign up are met with a banner that reads: “Due to large-scale malicious attacks on DeepSeek’s services, registration may be busy. Please wait and try again. Registered users can log in normally. Thank you for your understanding and support.”
The type of attacks are not clear, but will likely have ranged from distributed denial-of-service (DDoS) attempts to credential stuffing, as cyber criminals test the security resilience of one of the world’s fastest-growing AI platforms.
DeepSeek’s team has been scrambling to fortify its infrastructure, implementing additional layers of authentication and deploying enhanced network monitoring. However, this surge in cyber threats underscores a broader issue: as AI applications become more widely adopted, they become prime targets for exploitation. While DeepSeek has not yet reported any major data breaches, the sheer volume of attempted attacks suggests that adversaries are determined to probe its vulnerabilities.
Security researchers warn that DeepSeek’s immense popularity may make it a lucrative target for state-sponsored actors, particularly as geopolitical tensions between the U.S. and China continue to escalate. The AI industry has seen numerous examples of attackers attempting to manipulate or poison training data to influence model behaviour, and DeepSeek is unlikely to be an exception. The coming weeks will be a critical test of its security posture and ability to withstand high-profile threats.
Is DeepSeek Safe To Use?
While DeepSeek’s rise seems to a technological triumph, it has also raised serious concerns regarding censorship, data privacy, and geopolitical risks. Like many other Chinese AI models, DeepSeek operates under strict Chinese government regulations that mandate AI compliance with state-approved narratives. Asking R1 about topics like the Tiananmen Square massacre, or even just the 3 June 1989, will prompt a blank response that attempts to change the subject, or simply refuses to answer.
Australia’s science minister, Ed Husic, became the first member of a Western government to raise privacy concerns about DeepSeek, saying there were unanswered questions over data and privacy management on the platform. The U.S. Navy has already placed banned staff from using DeepSeek, due to “potential security and ethical concerns associated with the model’s origin and usage”.
DeepSeek’s privacy policy says that is collects large amounts of personal information, including email address, phone number and date of birth, and any user inputs. It says this information will be used to improve DeepSeek and will be kept for as long as necessary. While the amount of data collected should concern anyone careful with their privacy, the terms are similar to those for rival US services like ChatGPT and Gemini.
DeepSeek has demonstrated that computational efficiency can be just as disruptive as hardware scaling. The model’s cost-effective training process, ability to operate on constrained hardware, and MoE-driven efficiency signal a shift in AI development priorities. As a result, the AI industry must brace for a future where strategic innovation—not just raw compute power—determines who leads the next phase of AI advancement.
For U.S. firms, the DeepSeek breakthrough is both a challenge and an opportunity. While it highlights vulnerabilities in existing AI strategies, it also provides a blueprint for how companies can optimize their approach to AI training. Whether or not Nvidia and OpenAI adapt to this new reality will determine their standing in the AI race of the future.
