The EU is about to find out what happens when you tell companies to be safe without telling them who pays for it.
The Summary
- The EU AI Act begins enforcement soon, splitting safety obligations between model makers (OpenAI, Anthropic) and application builders (customer service bots, clinical tools), creating unclear incentive structures
- When the US required Anthropic to verify user nationality for national security, the company lacked the infrastructure and temporarily yanked access for everyone
- Meanwhile, China's Moonshot AI released full model weights for Kimi K3, no restrictions asked
- The real question isn't whether to regulate AI, it's who in the supply chain absorbs the cost of compliance
The Signal
AI safety isn't a single checkpoint. It's a relay race where nobody knows who's holding the baton. When you build a medical note summarizer, you start with a foundation model from OpenAI or Anthropic, then adapt it for clinical use. The foundation model maker handles training data choices, general safety guardrails, and documentation. The medical software company handles clinical accuracy, patient privacy, and regulatory compliance for healthcare.
The EU AI Act tries to split these responsibilities cleanly. Foundation model providers get one set of requirements. Application developers get another. But economics doesn't respect regulatory boundaries.
"Regulation can transform the development process by determining who must invest in safety and who waits for someone else to pick up the bill."
Here's the problem: if you require expensive safety testing at the foundation model level, smaller model makers can't compete. If you require it only at the application level, every medical software company reinvents the same safety wheels. Both scenarios waste resources. Both create gaps where nobody thinks safety is their job.
The Anthropic nationality verification incident shows what happens when requirements assume capabilities that don't exist. The US government said "restrict foreign nationals." Anthropic had no real-time way to verify citizenship. Their choice: build expensive identity infrastructure from scratch, or shut down access entirely. They chose shutdown. That's not a safety win. That's a coordination failure dressed up as compliance.
Key tensions emerging:
- Foundation model makers want application developers to handle domain-specific safety
- Application developers want foundation models to arrive pre-safe
- Regulators want accountability but can't map requirements to actual development workflows
China's approach offers a stark contrast. Moonshot AI released Kimi K3's full model weights publicly. No safety theater. No gatekeeping. Just raw capability, distributed. Western policymakers see this as reckless. Chinese policymakers see Western restrictions as self-imposed handicaps. Both are right.
The coming months will reveal whether the EU's layered approach creates clarity or just more expensive confusion. Early signs point to confusion. When multiple parties share responsibility for safety, economic theory predicts underinvestment. Everyone assumes someone else will handle it. Everyone is wrong.
The Implication
Watch how foundation model providers respond to EU enforcement. If they add safety features that application developers actually want and use, the regulation worked. If they add compliance paperwork that gets ignored downstream, it failed. The difference matters for every sector building on these models.
For companies building AI applications: don't wait for foundation models to arrive fully safe. That's not how the economics shake out. Budget for your own safety infrastructure. The regulatory ambiguity means you're liable either way.
For policymakers outside the EU: you're about to get free data on what works. The EU is running the experiment. Learn from their mistakes before copying their homework.