EU AI Act Goes Live August 2: What the New Enforcement Powers Mean for OpenAI, Google, and Every AI Company
Editorial note: This article is journalism, not legal advice. Regulations evolve. Confirm current obligations with qualified EU counsel before making compliance decisions.
On August 2, 2026, the European Union flips a switch that AI companies have been dreading for two years. The EU AI Act's enforcement powers go live. The AI Office and national regulators can begin investigating and fining the makers of general-purpose AI models.
The headline number: up to €15 million or 3% of global annual turnover, whichever is higher.
For OpenAI, Google, Anthropic, Meta, and Mistral, that is not a rounding error. For Alphabet, 3% of turnover is more than $10 billion.
What Actually Changes on August 2
Three things activate at once.
- The AI Office gets teeth. The Commission's 36-person AI Office can now open formal investigations, request documents, and issue fines against providers of general-purpose AI models (GPAI).
- Article 50 transparency rules apply. Any AI system that interacts with humans, generates synthetic content, or produces deepfakes must disclose that fact. Chatbots must say they are chatbots. AI-generated images and video must carry machine-readable markers.
- National authorities take over local enforcement. Each member state must have designated an AI regulator by this date. They handle everything outside GPAI.
The GPAI rules themselves have been on the books since August 2025. What was missing was the power to punish anyone for ignoring them. That is what changes.
What Does NOT Change
This is where most headlines get it wrong.
The full high-risk AI system rules — the ones covering hiring tools, credit scoring, medical devices, and critical infrastructure — do not become enforceable in August 2026. In May 2026, the Digital Omnibus agreement pushed that deadline out by roughly 16 months.
So the picture is split. Foundation model makers face live enforcement now. Companies deploying AI in regulated verticals get more time.
Who Is in the Crosshairs
The AI Act defines GPAI as models trained on broad data at scale that can perform a wide range of tasks. That is a short list in practice.
- OpenAI — GPT-5 and successors
- Anthropic — Claude family
- Google — Gemini
- Meta — Llama
- Mistral — the EU's own flagship
- xAI — Grok
Any of these providers offering their models in the EU market falls under the rules. That includes API access to EU customers, not just physical presence.
Models trained with more than 10^25 FLOPs of compute face an additional tier — 'systemic risk' obligations. That covers most frontier models shipped in the last 18 months.
The Four Core Obligations
Every GPAI provider must now maintain:
- Technical documentation describing the model's training data, capabilities, and limitations.
- A summary of copyrighted content used in training, made publicly available.
- A copyright compliance policy respecting EU text-and-data-mining opt-outs.
- Downstream information for developers building on top of the model.
Systemic risk models add four more: model evaluations, adversarial testing, incident reporting, and cybersecurity protections.
The Fine Math
Penalties scale with the violation.
- Up to €35M or 7% of global turnover — for prohibited AI practices (social scoring, real-time biometric ID).
- Up to €15M or 3% of global turnover — for GPAI obligation breaches.
- Up to €7.5M or 1% of global turnover — for supplying incorrect information to regulators.
The 3% turnover figure is the one that matters for foundation model providers. Alphabet's 2025 revenue was roughly $350B. Microsoft's was near $270B. A 3% ceiling puts theoretical maximums in the $8B–$10B range per violation.
What This Means for Investors
The market impact will not be a single-day shock. It plays out over quarters.
Compliance costs rise. Legal, safety, and policy teams grow at every major AI lab. Expect this to show up in operating expense lines through 2027.
The EU becomes a bottleneck for launches. New model releases may hit the EU weeks or months after the US. Meta already delayed Llama Multimodal in Europe. Apple delayed Apple Intelligence there for the same reason.
Smaller players get a moat problem. Mistral, France's flagship, argued for lighter rules for exactly this reason. The compliance burden is easier to absorb at OpenAI scale than at startup scale.
AI governance software becomes a real category. Vendors like OneTrust, Credo AI, and Holistic AI benefit as every enterprise using EU customer data now needs documentation workflows.
The Bigger Picture
The EU AI Act is the first comprehensive AI law from any major economy. What happens next matters globally.
The Brussels Effect is real. GDPR reshaped data privacy worldwide because global companies chose one compliance regime, not fifty. The same pattern is likely with AI. Whatever documentation OpenAI produces for Brussels will end up being the global standard.
Watch two things over the next 12 months. First, whether the AI Office actually opens an investigation — enforcement without a headline case is toothless. Second, whether the US responds with a federal AI law or continues the state-by-state patchwork.
When does the EU AI Act enforcement start?
August 2, 2026. That is when the AI Office and national regulators can begin fining providers of general-purpose AI models for violations of the rules that took effect in August 2025.
How large are the fines?
Up to €15 million or 3% of global annual turnover for GPAI violations, whichever is higher. Prohibited practices carry higher penalties of up to €35M or 7% of turnover.
Does the EU AI Act apply to US companies?
Yes. Any provider offering an AI model or system in the EU market is covered, regardless of where the company is based. Selling API access to EU customers is enough.
What is Article 50?
Article 50 covers transparency. AI systems that interact with people, generate synthetic media, or produce deepfakes must disclose that fact. This applies to chatbots, image generators, and voice clones.
Are high-risk AI system rules also live on August 2?
No. The May 2026 Digital Omnibus agreement pushed most high-risk AI system obligations out by roughly 16 months. Only the GPAI and transparency rules become enforceable in August 2026.
