DeepSeek: some trade-related aspects of the breakthrough 

DeepSeek has become a buzzword, and well-deservingly so. The Chinese company released a large language model (LLM) which matches the level of established AI models like ChatGPT (Open AI), Llama (Meta), and Gemini (Google) for a fraction of the cost, using less computing power, and with less energy consumption. The breakthrough will cause ripples across several areas, from infrastructure to data policy. From a  trade perspective, several interesting aspects are worth highlighting. 

Trade and access to technology 

Technology is unevenly distributed across space and time. In each wave of technological development, often only a small number of countries have the necessary resources to be innovators or early adopters, while most developing and least developed countries are placed within the ‘late majority’ or ‘laggards’. Trade is one of the most important mechanisms to allow technology to flow from places that present more significant financial and technological capacity to those that present fewer endowments. 

Rogers’ diffusion of innovations curve. Source: Kurbalija (2024)

When trade flows get disrupted: the example of US export controls on AI semiconductors 

Export controls imposed on technology restrict trade flows, affecting supply chains, production, and global access. One of the pillars of the United States’ strategy to preserve its technological leadership on AI in relation to China has been to impose export restrictions on advanced semiconductors. 

On the one hand, US export controls seem to have been efficient in preventing Chinese companies, such as DeepSeek, from acquiring state-of-the-art technology. DeepSeek’s AI system was developed using NVIDIA H800 GPU and less powerful chips. On the other hand, ‘necessity is the mother of invention’, and technology scarcity has led the engineers of DeepSeek to ‘think outside the box’, fostering an innovative leap forward. If China also develops the ability to produce its own cutting-edge advanced chips, it will display the powerful combination of computing capacity and efficient algorithms, and the main goal of US export control policy will have backfired.

Supporters of export controls argue that companies such as DeepSeek are still benefiting from the stockpiling of US chips, bought before export restrictions were introduced. As technology progresses in the US, the gap between the US companies and Chinese companies – still reliant on NVIDIA – will progressively widen, if export controls remain firmly in place. 

At the same time, the economic and social costs of using export controls to avoid competition and to preserve one country’s technological dominance make them a highly controversial tool. Professor Milton Mueller rightly questions the benefits brought to American citizens from “shrinking or retarding the economic development of 17% of the world’s population”, but it is also important to highlight the negative impact on the global economy as a whole that stems from slowing the progress of one of the global technological powerhouses. 

DeepSeek also shows how export control regulations on dynamic fields such as AI may be quickly outdated. On January 15 the US introduced an interim final rule on new export restrictions, which included, for the first time, limitations on AI Model Weights. Model weights are numerical parameters within an AI model, encoding what has been learned in the pre-training and fine-tuning phases, and shaping AI predictions. The regulation is based on the assumption that the weights for advanced AI models can be produced only by computationally intensive model training. DeepSeek uses a much less intensive computational capacity, and may craete the need for the Trump Administration to reassess the threshold established by the export controls norm in relation to AI model weights.   

DeepSeek and digital services: more cost-effective access to AI

AI models are becoming a layer of the global infrastructure. Companies increasingly use AI as a foundation to build, enhance, and scale their own products and services. Instead of developing AI from scratch, many businesses leverage existing AI models, including through cloud-based solutions (AI as a service or AIaaS), as foundational layers to innovate across industries. An AI model can also be fine-tuned for specific needs, such as medical diagnosis in the context of healthcare, or inventory management within e-commerce, for example. 

DeepSeek has developed models at a fraction of the cost incurred by major American AI companies, and this cost reduction can be transferred to companies developing their  products and services upstream, making AIaaS more affordable. The CEO of Perplexity – an AI-powered search engine, which uses advanced AI models to deliver direct answers to user queries – has already mentioned he intends to use DeepSeek to reduce operational costs, for example. Many developing countries exhibit lower AI adoption rates, often due to limited infrastructure and high implementation costs, and these countries could become the main beneficiaries of cost-effective AI solutions. 

DeepSeek also decided to provide an open source solution – making the Code of its AI system available for scrutiny, re-use and adaptation. This also empowers countries with less economic resources to technically learn from breakthroughs and to seek to leapfrog their path toward AI development. 

Although it holds great potential, the provision of cross-border AI services may not be smooth in practice. In order for companies to operate in other countries, they need to comply with local law. Recently, the Italian data protection authority decided to block DeepSeek on the basis of privacy concerns. This is an example of the obstacles created by lack of regulatory harmonisation in areas that are essential to nurture the trust that must underpin digital trade.

Algorithms and intellectual property: open source solutions as a way to foster accountability and bottom-up AI 

AI algorithms are a specific type of source code. The legal protection of the source code takes place through intellectual property law. In many countries, copyright can be invoked, but trade secrets – which can be understood as confidential business information that provides a company with a competitive advantage and is not publicly known – will often be the main form of protection. Some companies, such as Open AI and Google, keep their AI algorithms proprietary. Others, such as Meta and DeepSeek, have chosen to make the source code of their AI models available. 

Although so far proprietary models have predominated in the market, open source has been gaining traction, as noted by Yann LeCun, Meta’s chief AI scientist, who wrote “To people who see the performance of DeepSeek and think: ‘China is surpassing the U.S. in AI.’ You are reading this wrong. The correct reading is: ‘Open-source models are surpassing proprietary ones.’” This is very important, for at least two reasons: 

1. It reinforces transparency and accountability. Access to the source code of AI systems may be needed in many practical situations, such as ensuring compliance with safety or health regulations, in the context of financial regulation, in order to ensure fair competition, and to ascertain legal responsibility. In spite of that, legitimate requests for access to the source code of foreign companies have been met with a legal barrier created by source code provisions enshrined in trade agreements. These provisions restrict the cases in which an actor may request access to the source code or algorithm embedded in a product or service. These limitations are increasingly at odds with domestic laws that aim to align AI with safety standards and with fundamental human rights. Source code provisions create such an important barrier for AI transparency and accountability, that some experts from the Oxford University suggested removing source code provisions from trade agreements altogether, while proposing alternative measures to protect companies’ intellectual assets. 

2. It lowers the barriers for the development of bottom-up AI, mitigating the current transfer of knowledge and data to big AI players. Open source AI gives users the freedom not only to scrutinize, but also to modify and build upon AI models. Once AI models become accessible and customisable, actors will increasingly be able to develop their AI solutions in-house, without the need to hand-in their data to a major AI company, fostering bottom-up AI development. This will allow smaller players to develop AI systems attuned to the specific needs of their target audience. At Diplo, the notion of bottom-up AI has been integrated into the life-cycle of the organization, and data labeling has become part of a daily practice in teaching and research. This gives the engineers working at Diplo’s AI Lab access to high-quality data to train our AI models. The development of concrete applications, such as the AI-powered reporting of major international events, such as the UNCTAD e-commerce week, shows the value of bottom-up AI in serving our target audience and fulfilling Diplo’s mission, in a way that strengthens agency and autonomy at an organizational level. 

An AI race? Where is it heading? 

The idea of an AI race is a poor metaphor. A race is a zero sum: someone wins; someone loses. From a technical standpoint, the zero-sum mentality is misleading. Innovation does not come from thin air, but from standing on the shoulders of those that preceded us. DeepSeek’s technical note mentions the importance of Llama (Meta) and Qwen (Alibaba) for model training. In parallel, the CEO of Perplexity mentioned that DeepSeek’s open source nature will likely help US innovators, who will be able to introduce elements inspired by DeepSeek’s breakthroughs in their own AI models. 

From a global perspective, trade flows have facilitated the diffusion of technology across geographies, improving people’s lives and the global economy. The path for the development of AI has no finishing line, but it is a constant ‘work-in-progress’ based on the efforts of scientists and engineers that learn from each others’ breakthroughs, and on the proactivity of companies that convert these advancements into product development. Together, openness and collaboration make a positive impact on trade in technology products, services, and global GDP.