Two stories from this week’s roundup — one from Silicon Valley, the other from Chinese courtrooms — converge on a point that anyone investing in AI should heed: the growing unpredictability of the external environments that models and hardware rely on.

Anthropic, the company behind Claude, announced it is tightening access controls to its models. No details were given about which restrictions will take effect, but the mere decision to stiffen policies has an immediate consequence: it introduces a risk factor for all organizations that have embedded those services into critical workflows. When you consume an LLM via API, you delegate to the provider not just inference, but also the continuity of security policies, usage limits, and ultimately the predictability of the service. When a vendor unilaterally changes the terms of engagement, the real Total Cost of Ownership of a cloud integration can drift significantly from initial estimates — not for technical reasons, but because of administrative decisions.

This is not a new insight: moving inference on-premise, with self-hosted models, gives the organization full control over operating conditions and data sovereignty. But Anthropic’s move brings it back into focus with a concrete case involving a top-tier provider. For those evaluating local deployment, the point is not to evade a specific rule, but to eliminate reliance on decisions made elsewhere that can shift without warning.

The second story comes from the semiconductor industry. The legal dispute between Innoscience (a Chinese GaN specialist) and Infineon (the German giant) has highlighted how China’s judicial system can become a pressure tool in industrial conflicts. Beyond the merits of the case, what matters is that the availability of power-electronics components — also essential for servers running AI workloads — can be influenced by geopolitical maneuvers. For anyone designing on-premise infrastructure, this adds another layer of supply-chain fragility: relying on hardware suppliers concentrated in an area of high trade friction exposes you to delays, unexpected costs, or selective blockages.

These two episodes, though unrelated, paint the same picture: external dependencies, whether in software (a model only accessible via API) or in hardware (a strategic component subject to judicial leverage), multiply the points of failure. On-premise AI does not eliminate all risks, but it shifts the center of gravity of control from the outside to the inside. As organizations continue to weigh the trade-off between the upfront cost of local infrastructure and the flexibility of the cloud, such news encourages them to give greater weight to “predictability” in the overall balance sheet.