The EU's Digital Omnibus could either deepen regulatory fragmentation or unlock harmonized rules for platform AI — much depends on design and institutional coordination; unclear boundaries with the DSA risk higher compliance costs and uneven enforcement that favor large incumbents.

Main Finding

The Digital Omnibus aims to simplify and harmonise the EU’s dense digital rulebook, but evidence in the policy brief indicates that the central problems for the EU digital economy are not aggregate regulatory burden per se. Instead, the dominant governance constraints are structural: high concentration in AI-related markets (compute, hardware, cloud), growing data–privacy tensions driven by large-scale model training, and limited institutional enforcement and coordination capacity. Some proposed Omnibus changes would be substantive (data reuse, AI Act timelines) and could reshape governance outcomes rather than merely reduce administrative costs.

Key Points

• Scale of regulatory burden

  • Aggregate compliance costs for high‑risk AI systems are estimated at €100–500 million across the EU (EC impact assessment).
  • Typical firm-level costs for high‑risk AI: €193k–€330k initial; ~€70k/year monitoring/auditing.
  • Only ~5–15% of AI systems are expected to be classified as high‑risk; businesses devote ~3.9% of labour input to regulatory compliance (OECD).
  • Conclusion: compliance costs are real and can constrain specific SMEs, but are moderate in macroeconomic terms and not the primary systemic constraint.

• Data demand and privacy tensions

  • Frontier models require enormous datasets assembled via large‑scale scraping; models can memorize and reproduce training data, including personal data (Carlini et al. 2021).
  • Proposals in the Omnibus to broaden dataset access would interact with EU data‑protection rules, creating trade‑offs between innovation and privacy/compliance.

• Market and infrastructure concentration

  • Frontier AI development concentrated among a few firms (e.g., OpenAI, Google DeepMind, Meta, Anthropic); in 2023 industry orgs produced 51/66 notable models (Stanford AI Index).
  • Cloud providers (AWS, Azure, Google Cloud) control >60% of global cloud infrastructure; NVIDIA ~80% of market for AI GPUs.
  • Training advanced models is capital‑ and compute‑intensive (GPT‑4 training >$70M estimated); private investment in generative AI reached ~$25B in 2023.
  • These structural barriers (compute, data, capital) are the primary sources of market power and entry costs, not EU regulatory compliance alone.

• Political economy and scope of the Omnibus

  • Although framed as administrative simplification, some Omnibus proposals imply substantive regulatory reinterpretation (data reuse for training, AI Act implementation timelines).
  • Risk that simplification measures could weaken recently introduced governance mechanisms if not carefully designed.

Data & Methods

• Evidence base: synthesis of EU policy impact assessments (AI Act IA), OECD regulatory burden indicators, Stanford AI Index, academic papers on model training and memorisation (Brown et al. 2020; Carlini et al. 2021; Chowdhery et al. 2022), industry estimates/reports (CEPS, Korinek & Vipra), and advocacy/monitoring reports (Corporate Europe Observatory).
• Metrics used: aggregate and firm‑level compliance cost estimates; percent of labour devoted to compliance (OECD); counts of notable model releases; market shares in cloud and GPU hardware; investment flows to generative AI; estimated compute costs for training frontier models.
• Methodology: literature and policy‑document review triangulating impact assessments, sectoral statistics and academic results to evaluate (a) magnitude/distribution of compliance costs, (b) data requirements and privacy risks of model training, and (c) concentration in infrastructure and capital.
• Limitations noted in the brief:
  • Many quantitative estimates derive from ex ante impact assessments and industry reports rather than representative firm‑level panel data.
  • Rapidly evolving technology and market dynamics mean cost and concentration estimates are time‑sensitive.
  • Enforcement capacity and coordination effects are assessed qualitatively; empirical measurement of enforcement gaps is limited.

Implications for AI Economics

• Policy priorities should shift from blanket simplification toward targeted interventions that address structural constraints:
  • Competition & industrial policy: address bottlenecks in compute and hardware markets (e.g., procurement policy, support for diversified supplier ecosystems, conditional access to public compute).
  • Data governance design: craft mechanisms that balance lawful data reuse for safe AI innovation with privacy protections and accountability (e.g., controlled access pools, certified datasets, privacy‑preserving training techniques).
  • Support for SMEs and entrants: targeted subsidies, compliance assistance, or shared testing/assessment resources to lower fixed costs of meeting AI Act obligations for smaller developers.
• Enforcement and institutional capacity:
  • Strengthen EU enforcement infrastructure and cross‑agency coordination to make existing rules effective (invest in supervisory bodies, streamlined cross‑regulation procedures, resourcing for conformity assessments).
  • Monitoring and evaluation: collect firm‑level data on compliance costs, enforcement outcomes and market access to inform iterative policy adjustments.
• Regulatory design and dynamic effects:
  • Beware reinterpreting rules under the guise of simplification—changes to data‑reuse rules or AI Act timelines can have long‑run competition and privacy implications.
  • Consider complementary instruments (competition interventions, public‑interest compute/data commons, standards and certification regimes) rather than only deregulation to foster an innovation‑friendly yet safe ecosystem.
• Research agenda for AI economists:
  • Microeconometric studies quantifying firm‑level compliance burdens and investment responses to AI regulation.
  • Modeling the interaction between compute costs, data access, and market concentration on innovation and welfare.
  • Evaluations of governance interventions (e.g., shared compute pools, certified dataset schemes) on entry, competition and privacy outcomes.

Overall, the brief suggests that the Digital Omnibus should be assessed not only for administrative efficiency gains but for its likely effects on data governance, market structure and enforcement capacity—areas that matter most for the economics of AI deployment and competition in Europe.