The capital game in the field of large models continues, and the financing record is constantly being refreshed.
Recently, OpenAI's potential European rival, French artificial intelligence startup Mistral AI, announced the completion of a 600 million euros (about 4.6 billion yuan) Series B financing, which consists of 468 million euros of equity and 132 million euros of debt. The latest valuation stands at $6.2 billion.
Compared to other large model players, Mistral AI is a completely young company, only a year old, but that doesn't stop its ability to attract money, and to date, Mistral AI has raised a total of more than $1 billion from investors.
An unverified rumor is that Mistral AI's valuation soared from $5 billion to $6.2 billion, measured in "weeks". Regardless of whether this news is true or not, the valuation and amount of this financing show that even if the wave of star large model companies last year ended up with the founder leaving or being acquired at a low price, the team of investors who want to enter the big model around the world is still very long, which is nothing more than the length of the team.
Six-month valuation tripled
It is reported that Mistral AI is composed of a group of young scientists in the field of artificial intelligence, and the three co-founders, Timothée Lacroix, Guillaume Lample and Arthur Mensch (for ease of reading, let's call them Little T, Little G, and Little A), are from Meta and DeepMind respectively.
Among them, Xiao A is the CEO of Mistral AI, previously educated at a top university in France, with a PhD in machine learning and functional magnetic resonance imaging, and completed two years of postdoctoral research in mathematics. After that, he joined Google DeepMind, where he worked as a researcher for two and a half years, focusing on large language models (LLMs), and quit his job in 2023 to start his own business.
T and G are from Meta, serving as CTO and chief scientist respectively, and they have played important roles in the core engineering of Llama & Llama2, and have in-depth research and contributions to open source models.
The founding team also includes other experts from Meta (formerly Facebook AI), Hugging Face, and DeepMind, who have previously led the development of important models including Llama 1&2, Flamingo, Chinchilla, RETRO, and BLOOM, and the founding team's background has enabled Mistral AI to quickly make its mark in the AI space.
In September 2023, the development team launched the first open-source language model, Mistral 7B, which is distributed under a free Apache 2.0 license. Soon after, a more powerful version of the open-source neural network Mixtral 8x7B was released, and in April of this year, the same team released Mixtral 8x22B.
At the same time, Mistral AI has also released proprietary models (such as Mistral Large and Codestral) that are designed to be repackaged as API-first products through which users can use Mistral Large but pay based on usage. The company also offers a free chat assistant called Le Chat.
A record-breaking seed round
Back in April last year, the global AI arms race was heating up. It's not a good time for most entrepreneurs to enter the game, but just a few weeks after its launch, Mistral AI raised €105 million in seed funding, making it one of the largest seed rounds in European history.
Then, at the end of 2023, Mistral AI announced the closing of a $415 million Series A funding round led by a16z and Lightspeed Ventures at a valuation of $2 billion. In the Series A financing, Microsoft, Nvidia, and Salesforce are also on its list of shareholders.
It was also reported that Microsoft made a small investment of about $16 million in Mistral AI in February this year as part of a commercial partnership to offer Mistral's products through its Azure cloud computing platform, and does not hold any equity in the company.
It should be noted that Mistral AI is the second company after OpenAI to offer a commercial language model on Azure. Much like the OpenAI partnership, Microsoft's partnership with Mistral will also focus on the development and deployment of next-generation large language models.
An interesting fact is that although Mistral AI is a French company, it has many American investors, and the amount of the Series B round also set a new record for large-scale startups outside of Silicon Valley.
除了规模十分可观外，这次投资者名单很长，包括 Lightspeed Venture Partners、Andreessen Horowitz、Nvidia、三星风险投资公司和 Salesforce Ventures。
其他投资者还包括Belfius、贝塔斯曼投资、法国巴黎银行、Bpifrance、思科、Eurazeo、Headline、韩华资产管理公司的风险基金、IBM、Korelya Capital、Latitude、Millennium New Horizons、Sanabil Investments、ServiceNow 和 SV Angel。
That's why Arthur Mensch believes it's still a capital-intensive track, and he will use the funds to expand its commercialization and buy more computing resources.
"Differentiation" wins
In just one year, from a small team that is unknown to a super unicorn that can be called OpenAI, in my opinion, the core reason for the rapid rise of Mistral AI is "differentiation".
First of all, unlike most company locations, Mistral AI is a French company based in Paris. Even though Paris is an emerging center for AI R&D, the presence of Mistral AI has pulled Europe into the big model competition between China and the United States.
From the outset, Mistral positioned itself as a European alternative to the big American tech giants, and claimed that its products, such as the chatbot Le Chat, were available in multiple languages, including English.
At a time when AI is gaining geopolitical significance, French President Emmanuel Macron and other French government officials are fully supporting the company. Macron called the Mistral AI a sign of "French ingenuity" and invited Mensch to dinner at the presidential palace.
Secondly, in terms of product and technology route selection, Mistral AI is also different. Unlike companies like OpenAI and Anthropic, Mistral AI believes that AI software should be open-source, meaning that programming code should be available for anyone to download, copy, tweak, and repurpose. In addition, the release of the proprietary model also heralds the balance of open source and commercialization for Mistral AI.
Third, and most crucially, in the face of soaring computing costs, Mistral used only the "more than 1,000" high-performance GPU chips needed to train AI systems, and spent only "tens of millions" of euros to build models that rivaled those built by companies like OpenAI, Google, and Meta on larger budgets.
"The company's capital efficiency is outstanding, and Mistral is spending only a fraction of the capital its competitors need to build competitive AI models," General Catalyst investors said. ”
Admittedly, less cost, better performance, which is extremely attractive in any scenario.
At present, the open source manufacturers led by Mistral AI and the closed-source manufacturers represented by OpenAI have burned the war from the technical route to the commercialization and ecological field, and the industry is also paying more and more attention to profitability and landing.