Anthropic finds itself at the center of an increasingly visible contradiction in the AI race. The company, founded to build safe and aligned systems, now responds to accusations of rapid power concentration by claiming that its own success is the very condition for harmless AI. The statement, reported from circles close to the leadership, adds no technical details but crystallizes a strategic dilemma: safety depends on control, and control breeds imbalances.

The Anthropic doctrine: setting the pace to avoid being overrun

For Anthropic, being a protagonist in the large language model market is the only way to impose safety standards before less cautious players do. It’s not just about alignment research: it means having the infrastructure, data, and talent to set the agenda. In this view, a company that develops LLMs without stringent commercial constraints can afford to slow down when needed, test more deeply, and embed ethical constitutions and interpretability mechanisms. But all this requires capital, GPUs, and influence that few can match.

When the guardian becomes the gatekeeper

Critics see in this narrative a classic regulatory oxymoron: whoever holds the keys to safety becomes the sole enabler. In a scenario where the most capable models are served via APIs by two or three providers, true decision-making sovereignty fades. The issue is not just technological but one of governance: if safety depends on a single vendor’s commercial success, every decision — from content filtering to API deprecation — influences the entire ecosystem without contestation. This is where the on-premise model introduces a disruptive variable.

DIY safety: on-premise as a counterbalance

For organizations handling sensitive data or operating in regulated sectors, running LLMs on their own hardware is technically feasible today, albeit complex. Quantization, efficient fine-tuning, and architectures like Mixture of Experts reduce VRAM consumption, while frameworks such as vLLM optimize inference on clusters of consumer or enterprise GPUs. The advantage is not just TCO or latency: it’s the guarantee that the alignment pipeline remains under one’s own control. Safety decisions — from prompt systems to blacklists — do not depend on an external provider. This overturns Anthropic’s assumption: safety does not stem from a single company’s success, but from distributing the capacity to run and inspect models.

The invisible trade-off: centralized alignment vs autonomy

The real stake is the compromise between alignment consistency and operational autonomy. A centralized model benefits from uniform protection updates, but creates a single point of observation over data and user behavior. A fleet of self-hosted instances avoids this bottleneck, yet fragments the enforcement of safety policies and requires in-house skills for continuous auditing. There is no one-size-fits-all solution. However, recent IT history suggests that ecosystems overly dependent on one vendor end up absorbing its rigidity, while federated architectures, though costlier to manage, better withstand shifts in someone else’s strategy.

Beyond rhetoric: what really matters

Anthropic’s statement is a symptom, not a turning point. It reminds us that the AI safety debate risks becoming an argument to justify dominant positions. For those designing LLM deployments today, the implicit message is clear: trust in a vendor does not replace the ability to run models locally, verify their weights, and adapt them to one’s own regulatory context. AI-RADAR’s frameworks offer analytical tools to weigh these trade-offs, but the final choice belongs to those who decide how much delegation to accept in exchange for a promise of safety.