Digital Economy Dispatch #221 -- DeepSeek and the New AI Strategy Dilemma

The path to digital transformation progress is rarely a straight line. In the field of AI, waves of breakthrough product releases have punctuated long periods of steady research effort, creating an uneven trajectory of innovation. While countless teams have been pushing the boundaries of what's possible, a handful of new research results and product launches have fundamentally reshaped the AI landscape – from GPT-3's original release in 2020 to Claude and GPT-4 in 2023. These inflection points don't just represent technical achievements; they spark new debate about responsible AI adoption, redefine attitudes to AI-driven disruption, reconfigure user expectations for delivering value from AI, and unlock entirely new categories of AI applications.

We are now experiencing another of these key moments. Following the Covid catastrophe and subsequent economic digitally-driven restructuring, establishing an appropriate AI strategy has emerged as central to the plans of institutions and governments across the globe. Furthermore, the rapid evolution of AI is driving intense discussions about the technology's impact on economic competitiveness, national security, and democratic institutions.

Against this backdrop of growing excitement about AI's transformative power, a new player has emerged that perfectly encapsulates both the promise and peril of our AI-driven future. No surprise it has a lot of people’s attention.

The DeepSeek Phenomenon

DeepSeek, a Chinese AI startup that gained prominence in late 2023, has been developing large language models that match many capabilities of Western AI systems while using significantly less computing power and data. Their open-source releases, including specialized coding models, demonstrate that sophisticated AI systems can be built more efficiently than previously thought. This challenges the assumption that only well-resourced tech giants can develop cutting-edge AI technology.

What is getting people’s attention is the pace at which DeepSeek is improving its products. The company claims its R1 reasoning model achieves comparable performance to OpenAI's flagship offering, and its new Janus Pro multimodal AI surpasses Stable Diffusion and DALL-E 3.

DeepSeek's latest model demonstrates remarkable technical achievements. Built with 671 billion parameters and trained on an extensive dataset of 14.8 trillion tokens, the model required approximately 55 days of training at a cost of $5.58 million. In benchmark testing, DeepSeek-V3 has demonstrated superior performance compared to models like Llama 3.1 and Qwen 2.5, while achieving parity with GPT-4o and Claude 3.5 Sonnet.

Given this, for some people, DeepSeek’s latest models demonstrate the arrival of a new approach to train LLMs using much less sophisticated hardware and smaller amounts of data to produce a cost-effective way to create foundational models that will soon power many applications across every domain. Perhaps even pointing the way to a new economics of AI emerging. Stock market reaction has been swift: On 27th January 2025, Nvidia's shares dropped 17%, representing a near $600 billion market cap reduction – the largest single-day drop in US history

Furthermore, the fact that DeepSeek is produced by a Chinese startup company and released as open source for free download, amendment, and use represents one of those ironies that can only drive further innovation and will spark even more teams across the globe to enter the fray.

However, for others the arrival of DeepSeek's latest products is seen in a far more sinister light. Far from a new impetus to innovation, its detractors will point out that what is now being compared is the Chinese ability as a "fast follower" to create a similar solution 10-12 months after those produced in the West. Inevitably, given the time to review available solutions and take advantage of technology improvements, DeepSeek has moved quickly and produced impressive results.

Additionally, detractors are expressing several more fundamental concerns. The first hinges on the legality of how DeepSeek has built on previously released OpneAI products and their rights to do so. Did DeepSeek violate copyright and IP agreements about the use of models from products developed by OpenAI, Google and others? (Some may respond by pointing out that there are some serious concerns about how those companies obtained the data used to train models in the first place! But let's skip that for the moment.)

Second, it is clear that DeepSeek has been impacted by Chinese censorship. Ask it about events that the Chinese government would rather not acknowledge, and DeepSeek avoids answering. Often in not-so-subtle ways ("Let's not discuss that issue"). Additionally, when using online versions of DeepSeek, questions arise about the collection and use of data from users. The complex relationship and influence of the Chinese government on digital technology startups is well documented. What this means in practice for DeepSeek is yet to be clarified.

Third, the rapid advance of DeepSeek opens up questions about the influence (or otherwise) of US technology sanctions aimed at slowing China’s ability to deliver AI impact. Limiting China’s access to the latest technology has been a key element of US industrial and defence strategy in recent years. Do DeepSeek’s latest product releases point to the ineffective nature of these actions? Or have these only served to accelerate Chinese AI efforts toward greater innovation?

How can we understand and relate these divergent perspectives on DeepSeek? What do DeepSeek’s latest announcements tell us about the distinct ways that AI development is viewed around the world? What lessons can digital leaders take from these advances for strengthening their AI strategies?

The Four Internets: Understanding Digital Cultural Contexts

Perhaps we can begin to answer these questions by reconsidering Dame Wendy Hall's concept of the "Four Internets". It offers us an ideal context for understanding these AI developments.

She identifies 4 distinct approaches to digital technology delivery and governance, and examines how key technologies such as AI should be viewed from the resulting distinct cultural perspectives:

1. The Silicon Valley "open" internet, focused on innovation and commercial opportunity.

2. The Brussels "bourgeois" internet, emphasizing regulation and user rights.

3. The Beijing "authoritarian" internet, prioritizing state control and economic development.

4. The Washington DC "commercial" internet, balancing innovation with national security concerns.

Each of these different approaches reflects deeper cultural and political values that influence how societies view and implement AI technologies. DeepSeek's progress and reception illustrates these divisions perfectly. What some see as democratizing AI development, others view as a potential threat to intellectual property rights and privacy norms. While some applaud the role of open source to allow free competition around the world, others voice concern about data privacy and state interference.

Strategic Implications for Digital Leaders

Understanding and reviewing the technical advances embodied in DeepSeek’s latest announcements is important for all those looking to define effective digital strategies in light of AI’s relentless march. However, leaders and decision makers should look beyond these issues. As AI capabilities continue to advance and geopolitical tensions around technology intensify, they must focus on several key questions:

• How should organizations balance the opportunities presented by open-source AI models against potential security and compliance risks – particularly when these technologies are led by organizations with differing cultural perspectives?

• What frameworks should be developed to evaluate AI solutions not just for their technical capabilities, but also their geopolitical implications in the short- and long-term?

• How can organizations maintain competitive advantage while navigating increasingly complex international AI regulations and standards?

• Where should organizations look for new research ideas and products to provide competitive advantage?

From my perspective as someone deeply involved in AI strategy across several different contexts, I believe the DeepSeek case highlights the need for taking a more nuanced approach to AI adoption. Digital leaders cannot ignore these critical questions. They must look beyond technical specifications to consider the broader context of AI development, including questions of data governance, ethical adoption, and geopolitical considerations.

The path forward requires careful consideration of how different AI solutions align with organizational values and regulatory requirements, while remaining open to innovation regardless of its source. As we navigate this complex landscape, the ability to balance these competing demands will increasingly define successful digital leadership.

DeepSeek’s announcements highlight how we live in a world of competing AI visions. How will you define your AI plans to choose a responsible path between speed of adoption of new AI technology and strategic caution to deliver long-term value?