August 30, 2023

Open source is no longer a moat for GenAI startups

More companies are building and releasing open source AI models — what does that mean for startups trying to commercialise the tech?

Tim Smith

4 min read

This week, OpenAI, the Silicon Valley-based creator of ChatGPT, made a move that could have some founders and VCs sweating. On Monday, it released “ChatGPT Enterprise” — a new product aimed at businesses that lets them securely input their own company data into the AI system — creating a direct competitor to startups offering GenAI models as a business service.

But for Europe’s two best-funded companies in the space — Stability AI and Mistral — another recent announcement from Silicon Valley is likely to have a much bigger impact. In late July, Meta announced the open source release of its latest large language model (LLM), Llama 2, making it free for commercial use (OpenAI’s models are not open source). 

The move means that other companies can make use of Meta’s cutting edge AI technology to build products without having to train a model themselves (a process that can cost tens of millions of dollars). On top of that, this week, a new open source LLM initiative also launched on the EU’s LUMI supercomputer, adding a publicly funded competitor into the mix.


For startups that have been selling their open source strategy as a USP, developments like these could cause a big headache.

Early open source players

When London-based Stability AI raised a $100m round back in October 2022, the company positioned itself as the open source champion taking on closed-source Big Tech companies in the race to build GenAI technology. 

In a January interview with Sifted, founder and CEO Emad Mostaque waxed lyrical about how the company’s open source strategy would make it the preferred choice over closed-source models like ChatGPT in countries like India and Thailand. 

Then, in June, French startup Mistral raised €105m to build a European competitor to OpenAI, promising that its open source strategy would make it easier for companies to build “better, faster” products than those based on closed models.

Both of these companies are spending millions of dollars in equity fundraising to build their models, and will need to prove substantial commercial value in them to get a return on the investment.

One former Stability employee recently told Sifted that while he appreciated the company’s “pioneering efforts on culturally getting open source foundation models accepted,” it’s no longer a selling point in and of itself.

“Now the problem is that open source is the norm, rather than an exception,” they said. “Stability really needs to find its moats.”

Public money

AI investor Nathan Benaich, founder of Air Street Capital, thinks Llama 2’s open source release means time is running out for companies building large and expensive AI models aimed at enterprises.

“I think it shows there's probably not that many more chips you can play on this board,” he says, adding that the world will likely only need “a small handful” of companies working in the space.

To make matters harder, startups are now also competing with public money. The new European consortium, which launched this week, aims to build “the world’s largest open source LLM."


This initiative is being run on the LUMI supercomputer, which was built with taxpayer money from EU member states, while the data that will train the LLM was collected and curated as part of the EU-funded High-Performance Language Technology project. It’s being led by Finnish AI lab Silo AI and Finland’s University of Turku.

Open source LLMs like these — being built for multiple languages and with the help of public money — will put even more pressure on companies serving the European market.

All this isn’t to say that every open source GenAI company is in existential trouble — but it does mean that startups building big models with VC fundraising will need to start focusing on building useful products with LLMs, rather than just the technology itself. 

Some startups are already taking this route: Paris-based Nabla, for example, is focusing on using GenAI to develop tools for clinicians that speed up making claims to insurance providers. It uses a mix of third party AI models, including open source ones, rather than spending the money to build its own.

And as open source models from the likes of Meta keep improving, the challenge of selling AI models to other enterprises will only get trickier.

Tim Smith

Tim Smith is news editor at Sifted. He covers deeptech and AI, and produces Startup Europe — The Sifted Podcast . Follow him on X and LinkedIn