DeepSeek and the myth that AI is only for the giants

On 20 January, DeepSeek released R1, a reasoning model with open weights under an MIT licence. According to its technical paper, its performance in maths and coding is comparable to OpenAI’s o1. It also released six smaller, “distilled” versions that can run on far more modest hardware.

On Monday the 27th, the markets reacted. Nvidia fell by around 17% and lost almost $600 billion in market value, the largest one-day loss of market value for a single company in US history. The DeepSeek app even overtook ChatGPT in downloads on the App Store.

Beware the magic number

It has been repeated endlessly that DeepSeek trained its model for less than $6 million. That needs some nuance: the figure refers to the final training run of V3, the base model, not to the full cost of the research, staff and infrastructure behind R1. It is still a remarkable feat of efficiency, but it does not mean just anyone can do it on an SME budget.

What does change for a business

That said, the underlying news is good for anyone who is not a tech giant:

• Costs will keep falling. If similar performance can be achieved with fewer resources, the price per query of these models will tend to drop. Projects that did not add up a year ago are starting to.
• Open models are a serious option. For many business tasks — classifying documents, extracting data, drafting text — a smaller open model, run on your own or European infrastructure, can be enough.
• Don’t marry a single provider. Today’s best model may not be the best in six months’ time. Designing your tools so that you can switch models without rebuilding everything is an architectural decision you will be grateful for very soon.

And a word of warning

A model being open does not mean that any way of using it is appropriate. Using the DeepSeek app with company data means sending that data to servers outside the EU, with all the data protection implications that entails. Downloading the weights and running the model on your own infrastructure, where the data never leaves the building, is a very different matter.

The distinction between “model” and “service” is going to matter more and more. And it is one of the first things worth explaining to any team that is starting to use AI.