The legal skirmish between OpenAI and xAI has entered a new phase. On Monday, OpenAI asked a federal judge in California to rule that the trade secrets lawsuit filed by xAI “should never have been filed” and to force Elon Musk’s company to cover more than one million dollars in legal costs. The filing came just hours after xAI gave notice that it intends to appeal the dismissal of the case, which has already been thrown out twice.

Meanwhile, Apple has begun making its own moves against OpenAI, according to the original source, although precise details of the legal or commercial initiative remain undisclosed. What is clear is that the front of legal disputes around large language models is widening and becoming increasingly acrimonious.

For those watching the industry from an infrastructure perspective, these battles are not mere squabbles among billionaires. They signal a structural problem: protecting trade secrets related to model training and fine-tuning becomes extremely complex when the workflow passes through cloud providers, APIs, and multinational collaborations. Each handover – datasets, model weights, fine-tuning configurations – increases the surface exposed to litigation, with a real risk that the value of the work is diluted or stolen.

The xAI-OpenAI case revolves around alleged trade secret violations, a type of lawsuit that is multiplying in the AI world. Models, training architectures, and data pipelines are critical assets, often more valuable than the code itself. When these resources travel across shared infrastructure, the line between legitimate collaboration and misappropriation can blur. It’s no coincidence that several organizations evaluating on-premise deployment cite data sovereignty and full-stack control as an antidote to this kind of legal vulnerability.

If the court grants OpenAI’s request and forces xAI to pay legal fees, it would set a deterrent precedent for lawsuits deemed frivolous. But the deeper message for AI infrastructure decision-makers is different: uncertainty around intellectual property tilts the balance toward self-hosted architectures, where the boundaries between who develops, who trains, and who controls the data are clearer and more auditable. In air-gapped or bare-metal environments, every access to model weights and datasets is traceable, reducing the ambiguity typical of shared cloud services.

In parallel, Apple’s move – the details of which are still unknown – could trigger a chain reaction among major device and operating system vendors, who are increasingly aware that integrating LLMs into their ecosystems carries legal as well as technical risks. The decision on where to run inference and fine-tuning is no longer just a matter of latency or TCO, but of legal positioning. For those who follow AI-RADAR dynamics, this reinforces the need to carefully evaluate deployment frameworks that allow data and models to remain under direct control, without sacrificing performance.