A four-week-old lab just raised $113M to take on OpenAI

A four-week-old French startup with zero public products just pulled in a massive $113 million seed round at a $260 million valuation.

The Paris-based company is called Mistral AI, and it is stepping directly into the ring against OpenAI. But instead of building another proprietary black box, it is betting everything on open source.

The LLaMA pedigree

The funding amount is staggering for a seed stage, but the roster explains the hype. Mistral was founded by three French researchers in their early thirties who met back in university.

CEO Arthur Mensch came from Google DeepMind in Paris. CTO Timothée Lacroix and Chief Science Officer Guillaume Lample both hailed from Meta's Parisian AI lab. Lample, most notably, led the development of LLaMA, the foundational model that basically supercharged the open-source community earlier this year.

Lightspeed Venture Partners led the round. Antoine Moyroud, the partner who led the deal, put it plainly: there are only about 70 to 100 people on the planet with the technical chops to build and optimize frontier language models, and Mistral has three of them.

The cap table reads like a European tech summit. Backers include Redpoint, Index Ventures, French telecom billionaire Xavier Niel, former Google CEO Eric Schmidt, and the French state investment bank Bpifrance.

The enterprise open-source play

Mistral's premise is simple: OpenAI stopped being open, and enterprises do not want to be entirely dependent on a single American black-box vendor.

Mensch believes the biggest hurdle in AI today is simply making it genuinely useful for companies. Mistral plans to open-source its models and datasets, targeting its first text-based generative AI release sometime in 2024.

Crucially, the team intends to train models strictly on publicly available data to steer clear of the mounting legal and copyright headaches plaguing other generative AI outfits. Customers will also be able to bring their own internal datasets to tune the weights.

Lightspeed views foundation models like early cloud infrastructure: an expensive commodity race where only five or six global players will ultimately dominate.

Why it matters

This is Europe planting its flag in the frontier AI race, and doing it with a very specific strategic wedge.

For developers and enterprises, Mistral represents a potential escape hatch from proprietary API lock-in. If you are a European corporation worried about data sovereignty and looming EU regulations, a credible Paris-based open-source model is infinitely more appealing than handing your proprietary data over to a California-based cloud platform.

For the broader industry, it shows that the battle between open and closed AI models is just getting started. Meta proved there is insatiable appetite for open weights, and having the actual architects of LLaMA launch a venture-backed startup means open source will not just be an academic hobby.

That said, $113 million sounds like a fortune until you factor in current compute costs. Training frontier-scale models burns cash at a terrifying clip, and restricting the training set to strictly public data could limit performance. Mistral has bought itself a seat at the table, but the clock to deliver a working 2024 model is already ticking.