When Washington Pulls the Plug: Time for Private AI

The debate over digital sovereignty and operational dependence on U.S. hyperscalers is becoming noticeably more intense. Recent incidents at Google, Anthropic, and OpenAI have revealed just how quickly «AI as a Public API» can become an operational risk. Steffen Märkl, Director of Solutions Engineering at Cloudera, explains what this means for companies—and how private AI approaches enable greater resilience, control, and predictability.

Availability can no longer be taken for granted

The past few months have been a stress test for everyone who uses AI as an externally hosted service. First, Google Gemini went offline temporarily—and suddenly, a production process came to a standstill without any action on our part. Shortly thereafter, access to Anthropic’s latest models was blocked for all non-U.S. citizens due to an export restriction imposed at short notice by the U.S. government. Something similar happened with OpenAI’s latest model series: At the direction of the U.S. government, it was initially made available only to select customers in the U.S. and later withdrawn again.

These examples reveal an unpleasant reality: availability has become a variable that, due to external influences, is neither controllable nor negotiable.

The problem isn't with the technology

In fact, the real problem lies not in the technology behind public large language models (LLMs)—it lies in business resilience. By using public models, companies are at the mercy of geopolitical decisions, global outages, and opaque product and price changes. These factors primarily affect availability and day-to-day operations—precisely where AI now creates the greatest value.

Added to this is a business management aspect that is often taken seriously too late in many AI programs. Variable pricing models can lead to a situation where, as adoption grows, the benefits of AI increase, but the costs—especially when using agent-based AI systems—grow exponentially. GitHub Copilot provides a recent example: Following the switch to a token-based billing model, numerous developers have reported drastic budget cuts even with normal usage. Uber, too, found that the company had already exhausted its entire budget for AI programming tools and tokens for generative AI tools like Claude Code for the year 2026 after just four months. These developments show how quickly opaque cost structures for public AI services can become an operational burden.

Focus on Data Protection and Intellectual Property

In the context of publicly hosted models, data protection and the disclosure of sensitive corporate information and intellectual property must also be discussed. According to the Stanford AI Index Report 2025, the number of documented incidents related to AI privacy and security rose by 56.4 percent within a year. At the same time, it is becoming more difficult to ensure—through the use of public APIs—the seamless traceability of training and validation data, as well as infrastructure provenance, required by the EU AI Act for AI systems categorized as high-risk.

This is where the premise «we do have a contract, after all» is often overestimated: Contracts can define obligations, but they prevent neither disruptions nor geopolitical restrictions on access.

Private AI: Control, Resilience, and Predictability

Private AI comes into play exactly here. The key difference lies in the interaction between AI applications, models, and data. Unlike when using external AI models, all information remains within the organization’s own sovereign domain—regardless of whether the setup is in a public cloud, a sovereign cloud, on-premises, or a private cloud in the organization’s own data center. This enhances security, protects intellectual property, and provides the necessary traceability and auditability, which are becoming increasingly important in an era of digital sovereignty efforts. Above all, however, companies retain control over availability, governance, and costs.

When it comes to private AI, the focus shouldn’t be solely on defense. Those who operate AI in their own controlled environment can develop customized solutions specifically tailored to individual business goals and use cases. By automating repetitive tasks, decision-making processes can be accelerated and productivity increased—without compromising security or compliance. Although private AI often requires a higher initial investment than public services, the approach pays off in the long run: dependence on third-party providers decreases, and ongoing costs for storage, processing, and licenses drop—and, above all, become more predictable.

Public models as an option—not as a foundation

Anyone who wants to seriously deploy AI on a large scale should consider public models as an option—but not as a foundation. The foundation is digital sovereignty. The examples mentioned at the beginning illustrate this point: An outage at Gemini or a politically enforced access restriction, as seen with Anthropic and OpenAI, are not isolated incidents but recurring risks that companies will have to contend with in the future as well.

Private AI offers a solution here: AI is no longer an external risk factor that creates further dependencies; rather, it becomes a resilient, compliant platform on which companies can drive and scale innovation—without being hindered or even rendered unable to act in the event of outages, export controls, or price spikes.

Source: www.cloudera.com

This article originally appeared on m-q.ch - https://www.m-q.ch/de/wenn-washington-den-stecker-zieht-zeit-fuer-private-ai/