You have 1 article left to read this month before you need to register a free LeadDev.com account.
Estimated reading time: 10 minutes
Don’t miss out
Get weekly engineering leadership content in your inbox.
The arrival of a Chinese AI model has torched perceptions of what’s possible – but what does it mean for you?
Monday, January 27 was a historic day on the stock market. Following a social media snowball effect that propelled a Chinese AI chatbot called DeepSeek into the public consciousness, investors took fright and pulled nearly $600 billion of market value out of the chipmaker Nvidia, knocking its share price back by 17%.
It was the single worst one-day loss for a company in the history of the stock market. And tech stocks more generally fared little better: the Nasdaq 100 lost nearly $1 trillion in value.
What caused such a seismic shock? DeepSeek, developed by a quantitative hedge fund founder in China, and it’s underlying R1 model, is open source and uses a chain of thought technique to eke out performance equal to the best funded models in the west at a fraction of the price, using second-class GPUs that Chinese companies can still access despite trade embargoes. (The latter is up for debate, as some claim China has still managed to import the latest Nvidia GPUs.)
Bigger isn’t always better
DeepSeek’s R1 model calls into question the broad industry assumption that bigger is always better when it comes to AI models.
“The team that developed it focused not only on achieving high performance in existing tasks and benchmarks, but also did it in a computationally efficient way,” says Jerry Spanakis, assistant professor in computing at Maastricht University.
The resulting panic may have been overblown, however. Spanakis believes that, despite the eye-popping claims made in its accompanying paper, continued peak performance is not guaranteed.
“We don’t have access to the data the model was trained on, or the code,” he says. “Therefore, while we can fine-tune the model, we can’t really assess the full rationale of its performance. That being said, the team behind it is just really good and that also shows in the papers they release.”
More like this
Unlocking tailored models
The biggest shock DeepSeek has brought on the industry is how it potentially brings a wrecking ball down on the sizable barriers to entry for anyone wanting to develop their own large-language model (LLM).
“This is rather like when bureau computing was replaced by in-house computers – only to flip back again to cloud computing when storage was the issue,” says Alan Woodward, professor of cybersecurity at the University of Surrey. “The reservation some organizations have had is relying on external services for AI, especially if they let valuable data outside the corporate boundary. If DeepSeek delivers on its promise, it’s suddenly feasible that organizations might be able to bring it in-house.”
It’s suddenly possible to imagine competing with off-the-shelf commercial AI models like OpenAI’s GPT-4 and Anthropic’s Claude using an open-sourced model that organizations fully control themselves.
The idea of tailored models controlled within a company has been growing pace since Meta open-sourced its Llama models, but has always seemed slightly beyond the mainstream.
“The fascinating part is watching organizations building their own models to capture their unique corporate knowledge,” says Woodward. “Couple that with the DeepSeek approach, and targeted LLMs with a DeepSeek-style reasoning engine start to look very plausible.”
London • June 16 & 17, 2025
Speakers Gergely Orosz, Camille Fournier and Lara Hogan confirmed.
What now?
Rank-and-file tech workers may well be left wondering what this means for them. Big tech firms in the US have completely reorganized their businesses around building generative AI systems that are bigger in terms of computational might, rather than smarter in how they work.
That all changes with the performance capabilities unlocked by DeepSeek. “It might raise the competition bar higher as to how to build efficient LLMs and not just bigger ones,” says Spanakis. Workers on those teams will have to work smarter, not necessarily harder, to improve performance – and can’t guarantee a neverending supply of brand-new GPUs to fuel growth.
That could well be bad news for chipmakers, who saw their stock price battered earlier this week. With lower valuations, it’s possible that yet-more layoffs could follow. It could also be a similar story for cloud providers, who had banked on the AI revolution turning a few more times.
