The UK can emulate DeepSeek, and change the game of AI
An AI truism was shaken this week. And that might be good news for the UK.
It had long been thought that building advanced AI models required billions of dollars of investment and immense amounts of power consumption.
Wall Street investors certainly believed it: they’ve lent 11bn dollars to ‘neo-cloud’ companies with their hands on ‘the world’s hottest commodity’: Nvidia AI chips. And the reason Microsoft just signed a deal to recommission and purchase power from Three Mile Island nuclear plant is to feed the growing power consumption demands of training AI models.
But what if it wasn’t necessary? What if AI didn’t have to cost the world?
Enter: DeepSeek.
The Chinese AI startup rocked world news this week, releasing their ground-breaking R1 model. Perhaps the most advanced AI model we have seen and far cheaper than we believed possible. All built and trained with unprecedented efficiency.
What makes DeepSeek R1 so special?
The exact model used here is very technical. It combines Generative AI (makes new things and does new work), Reinforcement Learning (allows AI to learn with super human ability), with something called ‘Chain of Thought’ reasoning.
In simple terms, the latter allows their model to undertake a more complex form of reasoning. One which thinks deliberately about a problem, reflects on how its solution could be improved, explains and justifies its answers. I expect an arms race in the coming months, as other big firms release similar, improved, versions of this new type of model. But DeepSeek got there first.
But what really matters is the level of performance that DeepSeek achieved at such a low cost. The energy consumption and infrastructure (things like computing power, data centres) was far less than in other comparable models like ChatGPT. For comparison’s sake, it’s like buying an $1000 iPhone 16 for $50, with better functionality and higher quality.
Is it too good to be true?
China’s ‘Grand Plan’ for the nation’s future is to lead in several key technologies: AI, renewables, synthetic biology. The country is investing in technology education. There are some schools that only teach data science.
But one thing they lack is compute power. US controls ban the export of the powerful Nvidia chips to China – partly with the aim of keeping the US one step ahead. But that might have had the perverse effect of forcing them to think differently, and more efficiently.
The reason Nvidia shares slid dramatically on the news of DeepSeek’s breakthrough is because they did all this without the company’s latest AI chips.
But there’s a major caveat to all this. Some analysts are suspicious about whether that was possible – including Elon Musk. Microsoft, meanwhile, is currently investigating whether DeepSeek siphoned off OpenAI training data. We’ll doubtless know more soon.
But if they are telling the truth, the methodology could be game changing. Including for countries like the UK.
What does it mean for the UK?
Much of the media coverage has framed this as an AI arms race between China and the US – and DeepSeek is a definite win for China. But it’s more important than that.
The amazing recent advances in AI have mostly relied on massive training data sets and eye-watering computing power. OpenAI’s ChatGPT4, for example, cost more than $100 million to train. In other words: it’s very expensive, and so only a handful of companies have been able to do it.
DeepSeek claims their model cost just $6 million. If true, this opens the market to new, less well-resourced firms. Importantly, DeepSeek uses an open-source licence, which means other users can see, use, and build on the model. As we speak, other companies will be trying to replicate what DeepSeek achieved. And if they can, it will mean the development of AI models can take place almost anywhere. You don’t need $500 billion dollar funds (as the US announced last week) in order to become a ‘world leader’ in AI. You don’t need giant power stations. (The energy consumed to train GPT-4 could have powered 500 US homes for a decade). Sustainable, low energy AI could even be in the picture. We could see a genuine democratisation of AI innovation, as it becomes available in countries without the infrastructure and national grids of giants like the US or China.
Which brings me to the UK. Prime Minister Keir Starmer said this month he wants the UK to be a world leader in AI. There was some sniggering at that: how could we ever compete with the big players? That doesn’t seem quite so crazy now.
I’m not suggesting that everyone starts using DeepSeek. There are still plenty of unanswered questions – some relating to privacy and security. But I expect our AI industry is studying DeepSeek’s model right now, seeing if it’s possible to replicate what it’s done. How it can be improved. Mastering the same techniques could address some of the blockers to AI progress we face here. (One of which is power consumption. We have high energy costs compared to the US, which could really stifle the sector. But DeepSeek suggests power consumption to build AI models could be reduced by anything up to 90%).
We’ve always believed at QA that becoming an AI superpower is not just about infrastructure and money. If those things are less important than we thought, the other things will become more so: the people, the skills, the labs, the culture. And the UK has all those things in abundance.
Discover how our AI training programmes can help your organisation realise the power of AI in pursuit of its goals.
