Governments, Regulators Increase Scrutiny of DeepSeek

Chinese AI startup DeepSeek has published a new method for training large language models, aiming to address scalability issues in AI development. The method, known as 'Manifold-Constrained Hyper-Connections' (mHC), was co-authored by DeepSeek founder Liang Wenfeng. Analysts have described the approach as a significant breakthrough in AI training efficiency.

The release comes amid reports that DeepSeek is working on its next flagship model, R2, following a previous delay in its launch. The company has faced challenges including chip shortages and internal dissatisfaction with the model's performance.

DeepSeek's latest research has been welcomed by some experts in the field. Principal analyst Wei Sun from Counterpoint Research called the method 'a striking breakthrough,' noting its potential to improve model performance while minimizing costs.

How Did Markets React?

The AI industry is closely watching DeepSeek's developments. The company's R1 model, introduced in early 2025, triggered a short-lived sell-off in Nvidia shares and was credited by Nvidia CEO Jensen Huang with accelerating the shift toward open-source AI.

DeepSeek's new mHC architecture may enable the company to bypass computing bottlenecks and achieve significant performance gains, according to some analysts. This has led to speculation that the architecture could be integrated into future models.

What Are Analysts Watching Next?

Experts are cautious about DeepSeek's roadmap. While the company's research is promising, there are questions about whether the mHC approach will be implemented in a standalone R2 model or integrated into future iterations like V4. Analyst Lian Jye Su from Omdia noted that DeepSeek's openness could serve as a strategic advantage in the competitive AI space.

DeepSeek has not yet announced a timeline for the next model release. However, its past pattern of publishing foundational research before major product launches suggests that the mHC method is likely to be used in future models.

What Are the Global Regulatory Responses?

Governments and regulators around the world are increasing scrutiny of DeepSeek. In Australia, the government banned DeepSeek from all government devices in early February over security concerns. The Czech Republic followed suit in July, restricting the use of DeepSeek in public administration.

France's privacy watchdog has also taken an interest in DeepSeek, planning to question the company about its AI systems and potential privacy risks. Germany has asked Apple and Google to remove DeepSeek from app stores, citing data safety concerns.

Italy recently closed an antitrust investigation into DeepSeek after the company committed to improved disclosures about AI 'hallucinations'. These commitments were deemed sufficient to close the case by the Italian regulator AGCM.

South Korea suspended new downloads of the DeepSeek app in mid-February due to compliance issues with data protection rules. Taiwan has also banned government departments from using DeepSeek's service.

Despite the regulatory focus, Russia has taken a different approach. President Vladimir Putin instructed Sberbank to collaborate with Chinese AI researchers, including DeepSeek.

What Are the Implications for Investors?

DeepSeek's research and regulatory responses highlight the growing complexity of the AI landscape. The company's ability to innovate with limited resources could position it as a significant player in the global AI market. However, the regulatory environment poses a challenge to its international expansion.

Investors should monitor DeepSeek's upcoming model releases and regulatory developments. The company's strategic positioning in open-source AI and its collaborations with international partners could influence market dynamics.

The AI industry as a whole may benefit from DeepSeek's innovations, particularly as rival labs adopt similar approaches. However, the pace of adoption and regulatory responses will play a critical role in shaping the competitive landscape.