The ongoing evolution of large language models has spurred competition between open-source and proprietary models. Which will win? Excel V. Dyquiangco explains. 
 

The development of large language models continues to advance, with new models introducing notable capabilities. DeepSeek, a company founded by Wenfeng Liang (with a background in tech and finance, including co-founding the hedge fund High-Flyer), released DeepSeek-R1 in 2025, focusing on open-source development. This model led to Liang being named on multiple year-end as a key instigator of AI in 2025, including by British journal Nature (which called the 40-year-old Chinese entrepreneur a “Chinese finance whizz”); additionally, it has garnered attention for its reasoning performance and operational efficiency, demonstrating the ongoing evolution of LLMs. 

DeepSeek’s open-source philosophy, in contrast to OpenAI’s proprietary approach, is expanding access to powerful AI, prompting important discussions about the future direction of AI development. As Sean Brogan, a principal at AJ Park in Auckland, said, “The speed at which AI integrated and supported products and services have been entering the B2B and B2C markets has been rapid and continues to grow exponentially. The impact of this from a competition and business and consumer protection perspective is not fully clear and as is almost always the case, regulation is lagging.”  

As everyday consumers and businesses are exposed to AI, the companies behind them need to ensure that they are complying with business and consumer laws, for example: 

• By not exaggerating, or being misleading in relation to, the effectiveness or accuracy of these products and services. 

• Where they offer subscription models to consumers and small businesses, they must ensure that their terms are clear, transparent and reasonable. 

• Ensuring that their products and services do not inadvertently infringe intellectual property and data protection regimes, which includes being transparent with customers about who owns the output from these products and services, the confidentiality of any information inputted through their products and how they might use this information. 

“The legislative regimes requiring compliance with these requirements are aimed at protecting consumers and small businesses and also seek to, in some way, level the playing field when dealing with multinational organizations that wield substantial (and in some cases excessive) market power,” said Brogan. “We see compliance with consumer, business and data protection regimes when rolling out an AI backed product as an attempt to do so fairly and in a competitive manner.”  

Other ways that the large players can ensure they are not breaching competition laws include:  

• Engaging in fair pricing practices, not using their market power to undercut or keep competitors out of the market 

• Not entering exclusive contracts, prohibiting their suppliers from supplying competitors can severely impact competition. Relevantly, there has been significant interest in the infrastructure, including computer chips, that are used by these companies.  

The secret to trade secrets 

And still, there are trade secrets, which are an effective way for a company to maintain a unique competitive advantage. However as soon as the trade secret becomes publicly available – whether through a leak or reverse engineering – the trade secret loses its power.   

“Some AI platforms such as ChatGPT utilize information that has been provided by users,” said Dominic Scott-Jones, an associate at AJ Park in Auckland. “This information is used to further train and improve the platform and there are valid concerns that once you have shared a trade secret with one of these platforms, it has compromised the secret’s confidentiality.” 

He said that if an organization plans on integrating their trade secrets with an AI product, then it is essential that they utilize a product that does not share any information outside of their organization. Products such as Microsoft Copilot are marketed as having this level of security. 

“If aspects of their core offering are not open source, then it is essential that organizations like OpenAI and DeepSeek closely guard and keep secret the source code underpinning these offerings,” said Scott-Jones. “There are two broad categories which organizations can focus on to protect their trade secrets: implementation of digital and physical security measures preventing access to their systems, infrastructure and premises; and implementing internal policies and procedures which focus on and support a culture of confidentiality (such as educating employees, utilization of appropriate contractual restrictions when dealing with third parties and taking action against infringers or third parties that obtain unauthorized access to your IP).” 

He added that there are also other methods that they can take to protect core innovations, depending on what these are and the jurisdiction, this includes registering designs, copyright and patents. OpenAI has already taken serious steps in this direction, having filed more than 150 patent applications covering various aspects of its AI technology. 

According to Deanna Wong, the founder and CEO at DeLab Consulting in Hong Kong, while open-source policies and the retention of certain key trade secrets are not necessarily paradoxical, opting for a policy of open-source AI development does risk complicating trade secret protection by creating a clear tension between general transparency and secrecy.  

“Open-sourcing parts of AI models undeniably fosters collaboration and the progress of the industry as a whole; however, it also risks exposing proprietary elements like training methodologies or unique architectural components that competitors could replicate,” she said. 

She gave an example. While open-source models still claim to employ access controls and licensing restrictions to protect trade secrets related to model optimization and data curation, this balancing act becomes even more complex when employees inadvertently input sensitive data into public AI tools or when AI-generated outputs blur the lines between original trade secrets and derived content.  

“AI companies can address this tension by employing multi-pronged strategies to safeguard their innovations,” Wong said. “First, hybrid IP approaches combine selective patenting with trade secrets – some companies’ shift toward patents protects core algorithms while retaining proprietary training data as trade secrets. In fact, an open-source policy doesn’t require the company to disclose all its developments and technology, the crown jewels can remain protected as trade secrets or patented technology.”   

She added: “Second, technical safeguards such as encryption, strict access controls and watermarking AI outputs help track misuse, as with some of the ones with more AI/data use restricted environments. Third, legal housekeeping and policy frameworks become critical: robust NDAs for employees, consultants, partners and related parties, training, buy in, real implementation and impactful discussions on limitation to AI platform access for employees from work devices and compliance with emerging regulations like the EU AI Act and China’s emerging AI legislation.” 

Differing international IP laws 

Given the international nature of both OpenAI and DeepSeek, generally, the international scope of all these companies means that they must navigate various national IP laws, which are largely harmonized through the various global treaties, but still maintain certain variations between regions such as the United States, European Union and Asia.  

“These differences affect everything from patent eligibility and duration to the validity and enforcement of trade secret protections, compelling AI companies to tailor their legal strategies to meet diverse local requirements,” said Wong. 

In some jurisdictions, the lack of efficient trade secret enforcement may force AI companies to look for stronger IP protection such as patents (which, on the downside, require disclosure of the technology to the public). In contrast, other regions may offer stronger trade secret protections, prompting firms to lean more on internal controls and non-disclosure agreements.  

Wong said that balancing these approaches is key to safeguarding proprietary technology while complying with local legal frameworks.  

Another essential element to consider is that the law concerning AI (whether AI platforms risk copyright infringement based on their training datasets) is still in its initial stages of development, with various major cases making their way through the national courts, and which may lead to different outcomes. 

“Further, there are differing standards and applicable legislation in data privacy legislation and cybersecurity, which further complicate the legal landscape. Companies must adopt region-specific compliance measures to protect sensitive data and algorithms and must collaborate with local legal experts to draft enforceable agreements and establish robust internal protocols that mitigate the risk of inadvertent disclosures or loss of IP,” said Wong.  

For export control laws, and the limitations of access to high end computer chips, factoring into the IP legal battles between companies, export control laws and restrictions on high-end computer chips may risk creating an additional layer of complexity in IP legal battles between companies generally, said Wong.  

“For some time now, countries have tried to restrict access by using export controls, pricing, infrastructure, FDI regulations, among others, on access to related technology and raw materials including advanced semiconductor technology, which can limit the transfer of key hardware and software components essential for developing cutting-edge AI systems for others,” she said. “It also means that companies will find alternatives to develop its platform and technology despite those restrictions, which were already in place at that time. It therefore seems doubtful whether these restrictions will be the main factor in settling the IP and commercial battles between these companies.” 

Who may win? 

With the development within the AI Industry already going far beyond legislative IP frameworks in many jurisdictions, policy and decision makers are taking steps to update the law to reflect these advancements.  

“The attitude towards artistic works and copyright in China is quite different from the U.S. and rest of the western world,” said Brogan. “If DeepSeek wins the battle, it may have implications for how copyright works used in the creation or training of these generative products are treated. It may get to a point of no return, where such works have been used to such a degree that legal action is futile.” 

He said that due to external factors, there are significant differences in the output of the generative models created by DeepSeek and OpenAI, including where it comes to the censorship of the output.  

“Because of this, we see consumers outside of China having more faith in the OpenAI outputs,” he said. “However, it is difficult to predict who will win this battle as there are many unknowns that will factor heavily into the direction of AI generally. It is also worth noting that superior infrastructure and funding does not always equate to superiority in innovation. Innovation and disruption will often come from unexpected sources despite perceived adversity and uneven playing fields.”