Cloud Hosting

Artificial Intelligence is the next frontier of market concentration in the internet economy, but experts who spoke to Euractiv feel that even the EUs shiny new regulatory tools might be ill-suited to prevent abuses of market dominance.

In the coming weeks, EU policymakers are expected to finalise the AI Act, a landmark legislation to regulate Artificial Intelligence (AI) based on its capacity to cause harm. Since the draft law was first proposed, the discussion has been disrupted by the meteoric rise of ChatGPT and similar models.

The key to ChatGPTs success was not its use of generative AI, which has been around for some time, but rather the unprecedented scale and performance of its model, OpenAIs GPT-3.5, which has already been surpassed by GPT-4.

As a result, the discussions on the AI Act have been departing from the original horizontal nature of the law in favour of introducing stricter obligations for high impact foundation models like GPT-4.

This more targeted approach focusing on the most impactful actors, which incidentally happen to be primarily non-European companies, has become increasingly recurrent in EU digital policy, from the very large online platforms of the Digital Services Act (DSA) to the gatekeepers of the Digital Markets Act (DMA).

References to these categories are increasingly common in legislative provisions targeting Big Tech companies. However, no such cross-link is available for the EUs AI rulebook due to the DMAs most spectacular failure to date: not managing to designate any cloud service.

Big Tech is leveraging its market power in the cloud sector to gain a dominant position in the AI market. This process has been ongoing for a long time, Kris Shrishak, a senior fellow at the Irish Council for Civil Liberties, told Euractiv.

The question of which foundation models should be considered high impact is still a moving target, with policymakers oriented toward a combination of different criteria. However, one of the criteria initially floated has been the amount of computing power used to train the model.

Computing power is a critical component of AI. It is concentrated mainly in the hands of companies that have reached massive economies of scale for their commercial cloud services, hyperscalers like Amazons AWS, Microsofts Azure and Google Cloud.

There is no direct relation between being a hyperscaler and being a leading company in the field of AI. In addition, using the computing power used to train a model as a criterion to designate a high impact foundation model might also have a perverse effect, as investing more initially usually means the model is more robust.

However, training a model is only one part of the equation, as constant computing power is needed to fine-tune the model and its day-to-day operations.

Moreover, the impact of a foundation model is, to a large extent, proportionate to its user base. At the same time, only a few companies worldwide can run an AI model with hundreds of millions of users, such as ChatGPT.

Nobody can build a cutting-edge foundation model without having some kind of partnership with a Big Tech company, Max von Thun, Europes Director for the Open Markets Institute, told Euractiv.

In this context, leading AI companies are partnering up with tech giants without any intervention from competition authorities, as was the case for OpenAI with Microsoft and Anthropic with Amazon. These investments are often accompanied by more or less exclusive arrangements on the underlying cloud infrastructure.

Considering these partnerships as mergers is tricky because it depends on whether the cloud provider has a stake and influence on the generative AI provider and the type of relationship, like whether its an exclusivity or only strategypartnership, Christophe Carugati, an affiliate fellow at Bruegel, told Euractiv.

Behind great Artificial Intelligence, there is great computing power. Computing capacity is a much under-discussed aspect of the AI race, on which we tried to shed some light with Vili Lehdonvirta, professor at the Oxford Internet Institute.

The idea of a foundation model is that it can be adapted to various purposes, as new AI applications can be built on top of them. Since ChatGPTs public launch, the hype around AI has led to the blossoming of thousands of AI-driven companies.

However, the expensive infrastructural costs related to powerful AI models are already pushing this market to concentrate on fewer hands.

Many of the current players are suffering huge losses, largely because of how expensive the models are to run, said Zach Meyers, a Centre for European Reform research fellow.

It seems inevitable that many of the current players will either be left behind or acquired by bigger companies.

According to Andrea Renda, one of the experts who has contributed the most to shaping the AI Act behind the scenes, we are going toward a platformisation of the AI market, whereby most new AI models will be built upon a handful of foundation models.

This market concentration could lead to several ways dominant players could further entrench their position. For instance, when an AI solution is built on a foundation model, the downstream economic operator might be forced to run its AI application on the same cloud infrastructure, in a process known as bundling.

That is already the case when an AI solution is built as an Application Programming Interface (API) to a foundation model, which provides a sort of filter adapting the models response to the needs of the AI solution. As the query is being run directly to the foundation model, the API is supported by its underlying cloud infrastructure.

Conversely, hyperscalers would be incentivised to self-preference or bundle their foundation models with their cloud offers.

What we are witnessing is some of the Big Tech giants occupying the territory by making large investments in a handful of Gen AI companies, without anyone looking into it. Its like we learned nothing from the recent past, antitrust economist Cristina Caffarra told Euractiv.

The usual suspects are grandfathering market power into the future, and there is a lot of hand-wringing, but its already happened, she said.

One way to unbundle the foundation model and the cloud service underneath is by using a fully open-source foundation model. However, these are rather rare since many AI models that claim to be open-source tend to retain critical information.

Andrea Renda, a senior research fellow at the think tank CEPS, has worked on the EUs AI Act since its conception, advised EU policymakers during the negotiations and is currently part of the discussions on the AI Code of Conduct

Self-preferencing and bundling are critical elements that enabled the formation of mono- and oligopolies in critical parts of the internet economy, precisely what the DMA promised to prevent with its ex-ante obligations, as antitrust probes in the online sphere tend to conclude when the damage is already done.

One of the aims of the DMA is to move faster to prevent monopolisation before its too late. Ironically, the platforms designated so far are in markets that are already highly concentrated. With the AI and cloud, there is the possibility to be more proactive,vonThunadded.

The DMA failed to designate any hyperscaler as a gatekeeper because its quantitative thresholds did not fit the cloud sector.

Euractiv understands that France and Germany are pushing the European Commission to launch a market investigation following the qualitative criterion. Still, this process could take years and might take years of litigation to conclude.

Meanwhile, the AI market is moving at break-necking speed, with new generations of foundation models released every few months.

According to Jonathan Sage, a senior policy advisor at Portland, without the DMAs cloud designation, there is little the EU can do to prevent them from creating dependencies between their cloud infrastructure and the foundation models.

Still, the DMAmight be unable to prevent the entrenching of market power in AI since it does not directly cover foundation models.

A more effective solution would be replicating the DMAs systemic approach specifically for foundation models, as it is still unclear what consequences market dominance in this sector will have for downstream operators, Sebastiano Toffaletti, secretary general of the Digital SME Alliance, told Euractiv.

However, putting in place new rules or amending existing ones takes years, which is precisely what the AI market might not have. Anti-trust economist Caffarra stressed it was a matter of timing.

The DMA is looking at old problems but does not have the means to pre-empt a tight oligopoly forming at the foundation level in AI. Its just not the right tool. Before anything moves, it will be far too late, she concluded.

[Edited by Zoran Radosavljevic/Alice Taylor]

Original post:

Are EU regulators ready for concentration in the AI market? - EURACTIV