DeepSeek diplomacy: Disruption, dominance and data

DeepSeek’s emergence doesn’t solve AI hallucinations. The problems with the reliability of information persist. But the sudden splash caused by the Chinese company’s large language model is disruptive, challenging as it is to US tech dominance along with data access concerns.

To recap, Chinese company DeepSeek released an open-source generative AI model in December and free consumer chat bot in January. It hit the top of Apple's App Store chart and is comparable to OpenAI’s offering GPT-4. But the impact is much wider. The value of stock for chipmaker NVIDIA plummeted. It also threw into doubt many of the assumptions behind the strategies adopted by the burgeoning AI industry to date – as well as diplomatic competition.

First, it challenges US dominance in AI technologies and the idea that Chinese development could be curbed through export controls. It is possible – although widely debated – that export controls fuelled innovation and forced efficiency gains creating a comparable model, more cheaply. Questions are also being asked about the wisdom of rolling out huge, resource-intensive data centres, for example, and that the business case for scaling might not be scientifically based. The prevailing notion that leading-edge large language models will continue to require significant technical and financial resources has been challenged. One of the key insights from DeepSeek is that there is still opportunity for vast improvements in efficiency.

DeepSeek and the social media behemoth TikTok both offer genuine disruptions, having acquired large Western user bases. DeepSeek’s emergence has kicked off a wave of Chinese releases, which claim to be comparable to those of US companies and orders of magnitude more efficient (thus cheaper). These include Alibaba’s Qwen 2.5, ByteDance’s Doubao-1.5-pro and Moonshot AI’s latest model Kimi k1.5. Global, incremental improvements will continue.

DeepSeek V3 is open source, which means it can run locally, is more accessible and replicable. Thus, it is possible to establish independent versions of the model to mitigate some data concerns. This will continue to stimulate new models and broaden development opportunities to new countries, who have innovative talent, including Australia.

Open-source software creating new forms of value is a step change and could be a huge opportunity for the region.

Indonesia and India announced a partnership (through AIonOS and Indosat) establishing an AI centre of excellence, leveraging DeepSeek’s models. India’s upcoming AI Compute Facility will use DeepSeek models on local servers to address data concerns.Local instances may overcome reported issues with censorship and inability to answer questions critical of Chinese Communist Party policy.

Consumers and companies want and need products that add value and seize the potential of AI as quickly as possible. And as CFR’s Kat Duffy suggests, the existing US approach is not aligned with the needs of most countries around the world.

Open models go a long way to achieving that. Ideally, this could cause a reset in the AI market, focusing development on technologies that add real value, improve people’s lives and contribute to, rather than increase, global challenges such as climate change and stable governance.

Second, it highlights another quite unexpected concern. Separation between corporations and the state is a critical component of functioning governments. It’s one of the many reasons for legitimate national security concerns about the collection and sharing of data by Chinese companies, including TikTok and DeepSeek.

In an astonishing and unprecedented scenario, the world is now asking if such separation remains in the United States under the Trump administration, given immense tech company involvement. In particular, Elon Musk’s role, having become officially involved in the administration, being heavily invested in AI, and leveraged in China.

In recent days, Musk and team, in a somewhat unclear legal and organisational position, gained access to the federal payments system of treasury data. Allegations also include that Musk pressured senior officials and forced US aid agency members “on leave”. Musk aides locked workers out of the US Office of Personnel Management, accessed their computer systems and data, and installed former Musk employees. This leaves conflict of interest concerns as well as privacy and national security risks.

Finally, the DeepSeek emergence has reinforced the focus on large language models at the expense of other forms of AI, which may well be worth trillions. There is a growing sense that generative AI in its current form is heading towards the end of its shelf life, with some predicting valuations to drop precipitously. At Davos in January, Meta’s Yann LeCun suggested that within three to five years large language models would make way for other AI, that may not have the current limitations.

DeepSeek has a different organisational model to much of Silicon Valley. It is fully funded by High-Flyer (a leading Chinese quantitative hedge fund last valued at $8 billion) and has no plans to fundraise. CEO Liang Wenfeng said “even OpenAI’s closed source approach can’t prevent others from catching up”. The emergence of DeepSeek and subsequent models shifts the existing paradigm of AI. They don’t solve the reliability of information or hallucination problems, however we can expect a trajectory of continual improvement on existing models, especially when they are open source.

DeepSeek diplomacy? Expect more disruption, more data concerns and more scrambling for tech dominance, leaving a wake of market and geopolitical volatility.