“Simplification” of the AI Act and Mistral Compute
The playbook for hollowing out the AI Act, and Mistral's pivot to the cloud market in a push for sovereign AI
Hey all,
Today we cover the nuance behind the proposed loosening of the AI Act rules and recent developments on Mistral compute. It's a long one, so if you feel you want to steer clear of the intricacies of regulatory tussles, scroll down for our analysis on Mistral’s pivot to a cloud company.
Before starting - we recently published a new landscape report titled Artificial Power. It puts forward a strategy for the public to reclaim agency over the future of AI. In the aftermath of the “AI boom,” the report examines how the push to integrate AI products everywhere grants AI companies – and the tech oligarchs that run them – power that goes far beyond their deep pockets. See our report here!
Also, for your situational awareness: Europe has a new AI champion. Military-focused AI start-up Helsing took the throne of the most valuable AI startup of the continent, after a funding round that valued it at 12 billion euros (for comparison, Mistral is currently nominally valued at 6 billion euros). While the next funding round of Mistral is likely to put them back to the top, for a moment, the drone makers of Munich have their moment in the sun.
How does one “simplify” an AI Act? Scope, standards and downward harmonisation
The shift in focus from AI regulation to AI industrial policy (see chapter AI Arms Race 2.0 in our report) sweeps the globe. While not as radical as the proposed US state moratorium on AI (still facing significant hurdles to pass the Senate, namely the Byrd rule), the winds in the European Union are also blowing against the enforcement of existing regulation. While the EU’s regulatory pride seems to prevent it from walking back on the existing rulebook tout court at least for now, we’re still seeing more subtle forms of hollowing out the digital rulebook. Instead of the full frontal dismantling of the Act, the strategy relies more on exploiting potential tensions and weak points in the architecture of the Act.
In this edition of the newsletter, we look at concrete recent proposals from industry players to better understand the nuts and bolts of the deregulation that’s taking shape We look at three recent documents: position papers from the main lobby organization DIGITALEUROPE, engineering industry organization Orgalim, and the summary paper compiled of industry feedback, gathered by the rotating Presidency of the Council of the European Union for basis for discussions on simplification.
Gnawing away at the scope of the Act
The first core objective is to give precedence of sectoral rules above the horizontal rules agreed in the AI Act, alongside expanding exemptions to exempt AI systems already on the market. Orgalim and DIGITALEUROPE both propose streamlining Annex I of the AI Act to integrate its requirements with laws for machinery, medical devices, and radio equipment. The Presidency note echoes this, citing the challenging interplay with sectoral legislation as a major issue raised by stakeholders. Moreover, both Orgalim and DIGITALEUROPE recommend expanding the legacy clause to exempt AI systems already on the market, including GPAI models, from new obligations unless they undergo significant changes. In combination with the recent review of the high-risk use cases, these work to significantly pare down the scope of the Act.
Industry has long pushed for a more sector-specific approach to relieve administrative burden and blunt the law's impact on existing operations This approach, typically associated with the United States lawmaking tradition, which the European Union explicitly contrasted with creating a universal ruleset to harmonize markets in all AI products. As a sum of various loopholes and exceptions, the applicability of the Act is in danger of narrowing down substantially .
Industry capture of the implementation layer
Increased prominence for standardisation and the relaxing public oversight feature prominently. To recap - due to the product legislation origin of the AI Act, the Act merely sets high-level guidelines that companies need to follow. In practice, where the rubber hits the road is in the implementing standards and guidelines. One especially prominent site is the standards produced by the European standardisation organizations CEN-CENELEC. These arcane and opaque processes, largely driven by the private sector, are crucial for the actual impact of the AI Act as by following these standards companies are assumed to be compliant with the legislation. This helps companies to avoid compliance costs that would otherwise follow from trying to interpret the law.
The strong emphasis on standardisation in the industry demands suggests the importance for standardisation as a site of increased flexibility. The recent delays in the development of harmonised standards is a critical issue for all parties. DIGITALEUROPE and Orgalim both strongly advocate for postponing the application of high-risk AI requirements until at least 12 months after the relevant standards are published. The Presidency's report endorses this solution, listing the "Stop-the-clock" mechanism as a key idea to tackle this challenge, which would postpone application dates for laws if compliance tools like standards are unavailable. This idea has been gaining purchase even at the highest levels of EU policy-making.
The reliance on standardisation bodies by industry also signals confidence from industry bodies that they can control the process. The capture of the standardisation processes by industry bodies has been highlighted by civil society organisations as one of the suspicious components in the architecture of the EU legislation. Behind the scenes, there has been ferocious fights over the definition of core concepts and the possibility to transpose more industry-friendly, international ISO/IEC standards to the European context.
So far, the European Commission has been overseeing the standardisation, disciplining the process with critical comments over the inadequacy of the standards under development. To what extent the Commission resistance to the adoption of international standards and other industry-aligned proposals will continue under the current political pressure, remains to be seen The Commission left itself a backdoor to propose its own implementing rules, known as 'common specifications', in case the private standardization process fails. Unsurprisingly, this move faces criticism. The backdoor faces criticism from both Orgalim and DIGITALEUROPE, latter noting how it “discourages investment in the standardisation process”.
Conflicts between laws and downward harmonisation
Conflicts between existing laws also provide opportunities for reneging on some of the agreements made in the Act. For instance, the relationship between the AI Act and the GDPR is a significant concern for industry actors. DIGITALEUROPE and Orgalim make similar calls, with DIGITALEUROPE suggesting the replacement of Fundamental Rights Impact Assessments (FRIAs) with Data Protection Impact Assessments (DPIAs) to reduce redundancy—a concern reflected in the Presidency's summary of duplicative requirements as well. Moreover, both Orgalim and DIGITALEUROPE call for harmonizing some of the provisions between the GDPR and the AI Act where the AI Act goes beyond the existing data privacy regime.
The shift from FRIA to DPIA holds strong political importance. The expansive fundamental rights impact assessment, designed to gauge the societal impacts of AI, was a key ask from left-of-center parties during the Act's negotiations. While FRIAs were ultimately limited to the public sector and a select number of private actors, equating FRIA with DPIA would further narrow the scope by focusing primarily on data-centric harms, rather than the more expansive focus on the use of AI systems envisioned in the FRIA. Businesses are eager to stick to the narrower DPIA, which is already incorporated into their business operations.
Next steps
This is just a glimpse of the industry's wish list at this juncture. Additional proposals include deleting transparency provisions, elevating the role of flexible measures such as regulatory sandboxes as pathways to legal compliance, generally decreasing the administrative burden of documentation, and integrating business actors more explicitly into the governance infrastructure through new forms of "regulatory dialogue."
Forces within and outside the EU prevent a crude undoing of the legislative process. The European bureaucracy creates path-dependency and stability. Moreover, the effort to maintain the appearance of regulatory sovereignty in the face of US pressure creates constraints for EU policymakers. Hence, instead of brazenly walking back legislation, the regulatory rollback in the EU occurs at a more sophisticated level through reducing the scope, capturing the implementation phase and promoting downward harmonization in the case of conflict of laws. Exploiting the real weaknesses resulting from the centrifugal tendencies of EU policymaking, core parts of the law can be rewritten after the fact, bypassing the legislative process.
Discussions on simplification will evolve over the summer and mature by the autumn. While the legislative text of the AI Act is set, its real-world power and ability to protect fundamental rights are now being decided in the technical and bureaucratic trenches. We will continue monitoring this space as the situation evolves.
Mistral becomes a cloud company, Nvidia as the merchant of sovereignty
The Annual Vivatech conference took place in Paris last week. AI Now was represented by Amba, who took part in two panels on societal impacts on AI, one focused on democracy and other on privacy. The conference, one of the more prolific annual gatherings of the luminaries of European tech and funded by the French media-tycoon Maurice Lévy, is also often a site for new announcements for the European tech sector. This time, perhaps the most crucial new announcement was the launch of Mistral Compute.
In a significant move to capitalize on Europe's critical juncture on technological sovereignty, French AI darling Mistral AI announced the launch of "Mistral Compute". The new initiative marks a strategic partnership with the chipmaker Nvidia to build and operate a large-scale AI cloud platform. This development signals Mistral's ambition to expand beyond model development and become a key player in the AI infrastructure landscape.
The core of Mistral Compute will be powered by 18,000 of Nvidia's GB200 chips, with full deployment expected by 2026 in Mistral’s upcoming datacenter in the city of Essone, south of Paris. The platform is tailored towards enterprise and public sector clients with the hardware and software to develop and run their own AI applications, offering a European alternative to the US-based cloud providers.
Mistral's co-founder and CEO, Arthur Mensch, emphasized the strategic shift, stating, "We're expanding from an AI company doing software to a cloud company." The initiative seeks to offer wide variety of solutions, from raw computing power in the form of ‘bare metal’ GPUs to more fully managed product offering.
The announcement was met with strong support, including from French President Emmanuel Macron, who hailed the partnership as a crucial step for the nation's technological independence. The initiative has already attracted launch partners such as German generative AI startup Black Forest Labs, French bank BNP Paribas, and defense corporation Thales. Additionally, energy giant TotalEnergies has announced a collaboration with Mistral AI to develop new compute capabilities for developing AI in the energy sector.
The development of a European alternative to the big tech AI compute ecosystems is a response to the current geopolitical moment. Mistral’s effort to penetrate more deeply into the infrastructure layer and dislodge the usual suspects by providing access to GPUs is part of the recent broader focus on compute sovereignty in the European Union. By targeting the inference layer specifically, it complements the analysis on the gaps we have identified in public-private initiatives such as the gigafactories. Mistral’s expanding web of strategic partnerships in different industries will also provide a tractable distribution network, offering promise of defensible avenues for expansion.
However, these claims to autonomy are still on shaky ground.
First, Mistral’s offering is far from replacing the wholesale cloud offerings of the hyperscalers, as it focuses instead specifically on AI compute. As hyperscalers bundle their AI offerings into their existing wider cloud product offering with an entrenched customer base, the inertia benefits the incumbents.
Second, even if not attempting to replicate the scale of big tech, the capital expenses required for infrastructure necessitate colossal amounts of financing to be handed to Nvidia and other chip manufacturers. This kind of heavy capital expenditure does not fit the balance sheets of traditional venture capital, which turns the gaze towards institutional investors, such as large US private equity firms, such as Blackstone and KKR, and (Gulf) sovereign wealth funds. While some of the neoclouds, such as Coreweave and Fluidstack, have experimented with using their stock of existing GPUs as collateral for loan financing, the sustainability of such financial engineering has been questioned.
Third, how this initiative fits to the broader emerging ecosystem of alternative computing providers in France, remains an open question.
Behind all this, there’s one clear winner of the global AI sovereignty push - Nvidia. Jensen Huang has managed to position himself as the node through which dreams of tech sovereignty flow. From cutting ribbons in the supercomputer in Denmark, to speaking in panels alongside the Prime Minister of the United Kingdom and to the stage of Vivatech next to the French elite, his message remains the same: in order to keep up in the AI race, the best bet a sovereign country can make is to submit a purchase order for Nvidia chips. Huang’s approach is a masterclass in navigating the current geopolitical climate: locking-in sales across markets, while simultaneously mitigating key business risks by diversifying customer base away from the traditional big tech giants, whose business decisions directly reverberate to Nvidia's financials.
This is in line with our recurrent analyses last autumn and elsewhere. While there is some potential turbulence in the cloud level in the European Union, on the level of chips Nvidia is entrenching its position as the ultimate beneficiary of our contemporary tech moment. Competitors such as AMD and Cerebras are making small inroads to capture corners of the lucrative chokehold of sovereign compute.
However, as things stand, Nvidia reigns as the king of the AI jungle.