The release of the freely available and surprisingly capable language model DeepSeek R-1 shocked the world, made it question the growing demand for computer chips and led the mighty NASDAQ to dive on Monday. Here’s a Chinese open-source project matching OpenAI’s capabilities – something we were told wouldn’t happen for years – and at a fraction of the cost. The panic revealed more about our assumptions about AI than about the model itself.
Let’s be clear about what’s actually happening. For years, we assumed that making an AI breakthrough required three things – massive data centres, billions in funding and Silicon Valley zip codes. DeepSeek just flagrantly challenged that narrative, championing efficiency by using seemingly less money and less computing power, whilst staying competitive regardless of existing chip import restrictions.
The timing is particularly interesting. President Trump just announced the USD 500 billion Stargate project to dominate AI infrastructure and then – all of a sudden – this open-source model gains incredible momentum and essentially says ‘hey, we can play this game too – and we’re going to’.
The new US administration now realises that their ‘China problem’ is much bigger than TikTok. To make matters worse, both Bytedance and Alibaba have also released competing models to DeepSeek over the past few days. The irony wouldn’t be lost on those in Team Europe looking up and believing that the AI race was lost long ago.
But let’s not get carried away with the ‘China is winning’ narrative. The reality is more nuanced and, frankly, more interesting. What we’re seeing isn’t so much a shifting of power as a democratisation of AI capabilities. China has developed impressive AI talent pools, with researchers as well-trained as their American counterparts. And R-1 uses 700B parameters and significant computing power – this isn’t exactly AI on a shoestring budget. But it does demonstrate that the barriers to entry are lower than we previously thought.
R-1 is an example of so-called reasoning language models. These are trained to spend more time processing information, which has already led to successful results in maths and coding. As this progress is predicted to generalise to other problem areas, it’s another milestone towards more productivity across the board.
The real story here isn’t about national competition – it’s demonstrating that open-source AI has filled the gap with proprietary models. Just as Linux eventually became the internet’s backbone, open-source AI models are becoming the foundation of our AI future. Not necessarily because they perform better but because they are more accessible and anyone can improve them.
An excellent example of this is the foundation created by Meta’s LLaMa-2 model, which inspired French AI company Mistral to pioneer the algorithmic structure called Mixture-of-Experts, which is exactly the approach DeepSeek just improved. Bear in mind, reactions would have been very different if the same innovation had come from a European company and not a Chinese company.
Even though open-source AI is accessible, that doesn’t mean it gets a free pass on scrutiny. The north stars for practitioners – access to fully-disclosed input data and achieving the most energy-efficient inference – are within reach, though not yet realised. Despite all the admiration piled onto it, DeepSeek hasn’t disclosed the input data for its R-1 model and security researchers have already found sensitive data leaking from it. Although OpenAI also doesn’t usually disclose its input data, they are suspicious that there may have been a breach of their intellectual property.
Also, the jury is still out on the security aspect. Could the open-source nature of these systems transform them into powerful dual use artifacts? And is there a real way for their collective developers to track and – if anything – stop downstream deployment? Again, the timing is right – now it’s up to the EU’s Codes of Practice to ensure that it reflects the peculiarities of open-source AI.
The larger lesson for Europe is one we already knew very well, namely that lacking a stake in the game is caused by lacking skin in the game. After all, if China did it, maybe Europe can do it too. While there are tentative discussions over a collaborative AI Research Council, it becomes clearer every day that it’s time to start building. Initiatives like EuroLLM have the data and Mistral proved that European companies can scale AI models. The puzzle pieces are there, they just haven’t been put together yet.
The DeepSeek furore demonstrates that having a track record of developing prior AI models positions the team to swiftly capitalise on new developments. Although the company started publishing models on Hugging Face only in late 2023, it had already built a range of different AI tools before jumping onto the newest innovation that’s focused on spending more time and effort on fine-tuning models.
So, what does the future hold? Probably more panic, more breathless headlines about AI supremacy, and more and louder calls for protectionist policies. Stargate’s strategy for supremacy directly conflicts with openly sharing innovations with global competitors, leaving the geopolitical balance in limbo.
But the genie is well and truly out of the bottle. The future of AI development is becoming more and more like a global collaborative effort (even Trump has admitted this) – united in diversity, whether the superpowers like it or not.
And that might just be the best news we’ve had in AI development for a long time.