Brussels to label first “systemic” AI models this week, setting tougher EU obligations for OpenAI, Google and rivals
Enzo Reno ·
The European Commission’s AI Office plans to name the first general purpose AI systems deemed “systemic” on Wednesday in Brussels, according to officials involved in implementing the EU AI Act. The designations will likely cover the largest models from OpenAI, Google, Anthropic, and Meta, which would face tighter evaluation, cybersecurity, and incident reporting requirements than baseline providers.
This matters because the list will lock in which vendors must meet the act’s highest tier of obligations, with noncompliance subject to fines of up to 7 percent of global turnover, as set out in the law published in the Official Journal in 2024. CIOs and chief compliance officers will need to confirm supplier attestations and road maps before renewals in the second half.
What changes this week
A systemic label under the AI Act applies to general purpose models that meet capability and compute thresholds that the Commission has signaled around 10 to the power of 25 floating point operations, based on drafts discussed with member states in 2024. External estimates from Stanford’s Center for Research on Foundation Models and Epoch AI indicate that today’s frontier systems already exceed that level, which would sweep in OpenAI’s flagship family, Google’s Gemini Ultra tier, Anthropic’s Claude flagship, and Meta’s top Llama releases.
Once designated, providers must participate in standardized evaluations coordinated by the AI Office, publish more detailed system cards, and report serious incidents to EU authorities. Thierry Breton, the commissioner for the internal market, said in June that systemic providers should expect “tough supervision, robust testing, and clear accountability,” adding that the AI Office will coordinate with national market surveillance bodies on enforcement.
What enterprise buyers should do
Enterprises with European users should ask vendors this week for written AI Act readiness, including evaluation protocols, red teaming cadence, content provenance, and security incident playbooks. Legal teams are already pushing for copyright and safety indemnities that map to EU enforcement, a shift that mirrors the warranty language Microsoft and Adobe added to generative AI contracts in 2023 and 2024.
The United Kingdom and United States are moving in parallel on testing, though without binding model tiers. The UK AI Safety Institute has expanded public model evaluations, and the US NIST AI Safety Institute has issued profiling guidance aligned to the AI Risk Management Framework. Vendors that can show harmonized testing across EU, UK, and US regimes will have an advantage in procurement this quarter.
For the broader technology landscape, a formal systemic tier in Europe accelerates the shift from model hype to auditable performance, which will push capital toward evaluation tooling, safety operations, and compliance automation, and it will pressure foundation model providers to differentiate on documented reliability rather than raw benchmarks alone.