Chinese AI startup DeepSeek sent shockwaves around the world earlier this month when it released R1 — a large-language model (LLM) it says is on par with those developed by ChatGPT maker OpenAI, but was developed at a fraction of the price.
DeepSeek says R1 cost less than $6m to train; the startup offers the model in open source (meaning that the source code is free to download and use) or via an API, at a much lower cost than its competitors.
It’s raising big questions for AI companies building LLMs. With a new high-performance, low-cost player in town, there are reasons to doubt whether competitors will be able to run viable businesses selling the same product — which, in the case of companies like OpenAI, cost billions to build.
In Europe, all eyes are on Mistral AI, the only company developing LLMs in the region — which over the past 18 months has grown to become one of its most high-profile startups.
Launched in 2023 by Meta and DeepMind alumni, Mistral has raised over €1bn to date, reaching a €5.8bn valuation in just a year. It’s become the poster child of European sovereignty, epitomising the region’s dream of securing a spot in the AI race.
With DeepSeek suddenly reshuffling the cards for the sector, Mistral’s next move is likely to be heavily scrutinised — and the French company is coming under mounting pressure to prove that it will still be able to compete.
Commoditising LLMs
DeepSeek has accelerated a trend that many AI experts have been discussing for months: LLMs increasingly becoming a commodity.
“The big takeaway is that the recipe to AI is easy, lots of teams can do it. There is no moat around the skillset required to create high-quality foundational models,” Thomas Wolf, chief science officer and cofounder of French-American AI company Hugging Face, told Sifted.
For Mistral, whose original pitch was based on the promise of building world-leading models thanks to high-quality talent, this could pose a fundamental challenge.
“Mistral’s entire value proposition was: ‘we have one of the best skill sets for building models,’ but today, high-performing models have become commoditised,” says Tariq Krim, a French tech entrepreneur and founder of tech think tank Cybernetica.
This should push the company to differentiate itself thanks to the products it builds on top of LLMs — but Krim highlights that, contrary to DeepSeek or OpenAI, Mistral “doesn’t even have a mobile app.”
If they seize the opportunity right, this is ultimately good for them.
But Antoine Moyroud, partner at Lightspeed Venture Partners, which invested in Mistral, says that Mistral is already facing up to the challenge.
“It’s clear that Mistral has had to integrate engineering and product development into its DNA, evolving from a purely research-focused company to one that now covers those three fields,” Moyroud tells Sifted.
Although it is tempting to compare Mistral to OpenAI or Anthropic and now DeepSeek, says Moyroud, the French startup is playing in a different league; rather than focusing on consumers, the French startup targets enterprise customers.
Although Mistral has developed a consumer-facing chatbot called Le Chat, the company mostly focuses on deploying its models for organisations that need to integrate Gen AI tools in data-sensitive sectors like banking and defence. The startup says that its technology, which is open source, enables better safety and control than closed source providers like OpenAI.
Mistral has signed deals with French bank BNP Paribas, insurer Axa and the French ministry of defence.
That means that the startup isn’t directly competing against DeepSeek. “What DeepSeek has done is impressive, but I don’t see them building the infrastructure and engineering to serve their models commercially for safe and at-scale deployments,” says Moyroud.
Boosting Mistral’s models
For some, DeepSeek could actually serve to boost Mistral’s business. Moyroud says that Mistral’s engineers are already exploring the strengths and weaknesses of DeepSeek’s technology and whether they could leverage the Chinese startup’s innovation.
This could enable Mistral to develop cheaper models that will sell better with cost-aware businesses, says Callum Stewart, associate at investment bank GP Bullhound, which is not an investor in Mistral.
“So many proof of concepts (POCs) never get implemented because the costs are too high,” says Stewart. “But if you have a smart enough model that’s cheap to run, why would you not start bringing this into everything mundane?”
Companies will be wary of directly implementing DeepSeek’s technology due to the constraints and bias of deploying Chinese models, putting Mistral in a favorable position, says Stewart.
“If they seize the opportunity right, this is ultimately good for them,” he says.
Last year, German LLM developer Aleph Alpha announced that it was pivoting away from developing AI models to focus on building a “generative AI operating system”. The company’s cofounder Jonas Andrulis told Sifted that it had been hard to find the right economic model for the company.
So is Mistral headed for the same fate? Lightspeed’s Moyroud thinks not.
“It’s not at all relevant to talk about a similar pivot.”
Mistral declined to comment.
Read the orginal article: https://sifted.eu/articles/deepseek-mistral-what-next/
