DeepSeek's large language model (LLM) is "making a mockery of the (idea that) we need a trillion dollars to train the next level of AGI" (artificial general intelligence), said Neil Lawrence, machine learning professor at Britain's University of Cambridge.

He was referring to the US announcement last week of the pharaonic "Stargate" programme to build $500 billion worth of AI infrastructure, a plan led by ChatGPT creator OpenAI.

Largely destined to be pumped into data centres packed with the latest AI chips, the scale of the sums underscored the fact that almost no European firm can access the resources to compete at the cutting edge.

But DeepSeek's claim that it succeeded in producing a model with similar capabilities to OpenAI's for just $5.6 million has upended those certainties.

The technology promised "models that are more efficient and less hungry for GPUs (graphic processing unit chips that power many AI applications), for energy and for cash," Laurent Daudet, chief executive of French generative AI company LightOn, told AFP.

"It's interesting for Europe to see that we don't need a Stargate project to do something interesting... to innovate, you don't need $500 billion," he added.

The shock has in recent days battered AI-related tech stocks, including key chipmaker Nvidia.

"It shows that competition is very, very strong and that there'll be a price war too," said Nicolas Gaudemet, AI chief at consultancy Onepoint.

"It could hit the accelerator even more strongly on use of generative AI within companies... an additional provider will bring prices down," he predicted.

- Open source -

Cambridge professor Lawrence said it was a "tragedy" that DeepSeek had not emerged from Europe, pointing to the continent's deep reserves of AI talent in both academia and business.

But he added that the "small change in the recipe" for AI made by DeepSeek was just "a glimpse of the innovation" still to come.

"It is very encouraging for Europe and it's more reflective of what we should expect going forward... there will be more than one DeepSeek," he added.

Lawrence singled out for praise DeepSeek's adoption of open-source methodology, under which the developers exposed the guts of their project to the wider AI community who can then build further on it.

That was good news for European contenders such as France's Mistral, Gaudemet said.

"They can reuse (DeepSeek's models) to train up theirs and stay in the race," he pointed out -- although "they'll have to be very good, because the competition isn't just between the US and Europe, China is showing that it's capable".

- 'Card to play' -

Cheaper AI could mean more AI tools adapted for local markets and individual businesses rather than a massively-resourced, centralised model.

"There is a future for more frugal models that can perform just as well, especially for business needs," LightOn's Daudet said.

He described his company's role as building the "chassis" of a usable vehicle around the "motor" of an AI model -- "how a business can use it in a totally secure, personalisable way, with all the guarantees they want".

"There's obviously a card to be played for (European) companies, which is security, the aspect of 'your data will stay in Europe' and be handled by people whose interests are... in Europe," Onepoint's Gaudemet said.

In the future, "we may be less dependent on American model providers," he added.

Europe has "some of the strongest researchers in the world and bottom-up, we understand where those strengths are," Lawrence said.

"We don't need massive amounts of investment, but we do need people who are listening to people in their own countries and their own continent... not having their heads turned by whatever (OpenIA chief) Sam Altman's latest narrative is," he added.

tgb-kf/rlp

LightOn

Nvidia