The French Treasury has just published a depressing but realistic note on the ability of France, and more broadly Europe, to carve out a place for itself in the artificial intelligence sector, dominated by the GAFAMS. The note has been criticized by the cultural world for suggesting, with little argument, that the transparency requirement for generative AI training data, which has just been adopted at European level, should be downplayed.
The French Ministry of the Economy’s Treasury Department has published a note entitled « the artificial intelligence value chain: economic issues and France’s place« . Far from the grand declarations and rhetoric of France 2030, the note presents a cold, and to put it bluntly, almost despairing snapshot of France’s, and more broadly Europe’s, place in this sector of the future. To put it in a nutshell, the conclusion we can draw from the note is that France and Europe will be users of AI, but probably not players in the sector, and even less beneficiaries of the economic spin-offs of AI, except perhaps in specialized models. Of course, companies can hope to increase their productivity as users, and that’s something. In short, we’re here at the Treasury Department, a serious body, and realism is the order of the day, not political communication. And this realism, which could be likened to the findings of the Draghi report, is welcome, because how can we build without having made an honest study of the terrain? The problem is that the report makes very few recommendations, and that its only major recommendation – to limit the IA Act’s transparency obligation – shows that, once again, culture is Bercy’s adjustment variable. Yet it’s hard to see how lowering authors’ and performers’ rights would enable France to carve out a place for itself in AI.
Investment train
On market findings, the three authors of the note note that the AI value chain is divided into three parts, and that France can only claim a substantial share of the latter today. The first is made up of the inputs to AI systems, i.e. computing power, data and skilled labor. The authors point out that regarding these inputs « France does not have players of comparable scale to the world leaders in the chip manufacturing and computing capacity rental markets » but that it « does, however, benefit from a skilled workforce and a dynamic innovation system. » The second part of the sector is model development. Here, the note states that « some French players are emerging, but this segment is dominated by the big digital players, who pre-existed the arrival of this technology« . This is reminiscent of the finding of the European Court of Auditors, which we reported on last week, that Europe’s failure to jump on the platform investment bandwagon is resulting in the absence on the old continent of companies able to finance the following innovations, including AI. Here, the Treasury note speaks of « a few players », but goes on to note that « the development of AI foundation models is directly or indirectly dominated by large Silicon Valley digital companies, even though France has one start-up (Mistral AI)« . Of note, Mistral develops quality models, and has managed to secure a commendable €6 billion valuation, and an agreement with Microsoft, but remains an economic dwarf next to its congeners. According to Sifted, Mistral’s 2024 revenues will be around 30 million euros. That’s a far cry from Open AI’s 3.7 billion.
82% financed outside Europe
The note goes on to point out that, as investment capacity in Europe is more than limited, the development capabilities of the French models that would emerge would not be sufficient. « While the French and European venture capital ecosystem makes a significant contribution to these fundraisings, and appears to be performing well for seed-stage startups, the availability of French and European funds remains limited for fundraisings in excess of €50 or €100m, which are generally necessary to accelerate the move to international scale, » write the authors, who add that 82% of European scale ups between 2013 and 2023 had a main investor located outside the EU. So Europe has the individual talents, but the economic benefits that flow from these talents’ innovations often go abroad for lack of support.
« Not a major player in computing »
Similarly, for computing capacity: France has the talent – computer scientists, engineers, mathematicians – but « is not today a major player in computing« , according to the note. The reason is well known: the GPU market is dominated by Nvidia (85%) and AMD (10%). The rental market for computing infrastructures is also the playground of the big American players, « against which the existing French players play a marginal role. » The note points out, however, that « Scaleway has announced investments in computing infrastructure for AI and a partnership with Nvidia to office services for training AI models. » For the rest « other French players are positioning themselves on inference, i.e. the deployment and use of a model once it has already been trained. »
Data
After this observation that the market is the GAFAMs’ playground, we then come to the subject of the data that feeds AI. And, here too, the GAFAMs are in a strong position, since, as the note says, « the players with their own data are at an advantage in the development of foundation models or applications. This is the case for large companies, such as Google, Apple, Facebook, Amazon or Microsoft (the GAFAMs), which have both privileged access to large volumes of data via the content they host (YouTube, for example, offers Google a major source of training data for AI models) or via the data associated with the use of their services. « There’s nothing wrong with this observation, which is realistic, even if we could point out that there are currently disputes over the legality of the use of some of this data – such as copyrighted works – by these companies to train their AI tools. The note doesn’t mention this.
Copyright: an obstacle
To enable – virtually non-existent, therefore – European players to have access to quality data, just like the GAFAMs, the note suggests the deployment of public databases. But it also talks about access to what it calls « cultural data« , by which we understand it to mean data protected by copyright. The solution proposed in the memo has raised the hackles of many a cultural representative. The note’s authors start by asserting that « negotiating individually with each data owner or right holder can be a significant cost for AI producers« . Then it notes that « the European AI Act passed in February 2024 provides for an obligation of transparency with regard to the training sources of AI systems. As data is an important differentiator for AI system providers, this transparency is likely to hinder their development. » One might retort that the transparency obligation is not yet applicable and yet no European AI model of commercial scale has reached the market… So there doesn’t seem to be a substantial brake there. The note concludes that the implementation of the transparency obligation « will have to find a balance. » The Minister of Culture herself did not appreciate this suggestion to downplay copyright. Speaking about the Treasury note on her X account, she wrote: « those who oppose transparency and the development of AI have understood nothing. On the contrary, it is by guaranteeing creators’ rights that we will offer AI access to authentic content, which is the key to differentiation.«