May 2025 OpenAI order leaves flagged user data on legal hold
A 2025 court order preserved deleted OpenAI output logs, exposing data-control risks behind a $600 billion sovereign AI market built around open models and concentrated computing.
By Amara Osei
5 min read

What's News
- OpenAI’s May 13, 2025 order covered deleted output logs; flagged data remained on legal hold after the broad order ended in late September.
- McKinsey estimates sovereign AI could reach $600 billion by 2030, with up to 40% of AI workloads moving into sovereign environments.
- CNAS finds about 90% of frontier AI compute in the United States and China; McKinsey counts about 30 countries capable of advanced workloads.
- Nvidia agreed in August to pay Poolside $6 billion to license its model-building platform and offered jobs to more than 100 engineers.
- CNAS tracks 184 government-backed sovereign AI projects; more than three-fifths name at least one foreign partner.
By Amara Osei
On May 13, 2025, a federal magistrate judge in Manhattan ordered OpenAI to preserve output logs that users had told the company to delete, according to Osei’s account of its copyright dispute with The New York Times.
The order applied to ChatGPT Free, Plus, Pro and Team accounts, plus API customers without zero-data-retention agreements. Some enterprise customers were exempt, but many teams on Team plans and standard API keys did not join the case. Their deleted logs still fell under a preservation order arising from someone else’s litigation.
The broad order ended in late September, but data from accounts The New York Times flagged remains on legal hold. OpenAI challenged the measure. COO Brad Lightcap called it “an overreach.”
What does the order reveal about sovereign AI?
Osei defines sovereign AI as control over data, contracts, legal exposure and operating rules, not as a claim rooted in national borders or flags. When AI workloads run on infrastructure the customer does not own, the provider’s litigation, contracts and policies determine whether customer data survives a deletion request.
The court case supplies the mechanism: a company can agree that data should disappear while a judge orders it retained. Osei cites that outcome as a warning for claims files, source code and customer health records stored in externally operated AI systems.
How large is the sovereign AI market?
McKinsey surveyed 300 executives, investors and government officials in late 2024. It found 71% regarded sovereign AI as either an existential concern or a strategic imperative.
McKinsey estimates the market could reach $600 billion by 2030. It also sees up to 40% of AI workloads moving into sovereign environments, according to Osei’s analysis of the survey.
Osei also cites three surveys showing the close connection between sovereignty and open-source software:
- Nearly four in five organizations called sovereign AI a strategic priority in a Linux Foundation survey. About 90% said open source was essential or very important.
- Of more than 900 IT leaders across Europe, the Middle East and Africa, 74% told Red Hat that AI sovereignty was a priority for the next 18 months. Ninety-two percent called enterprise open source key.
- A McKinsey and Mozilla Foundation survey of more than 700 technology leaders in 41 countries found 63% of organizations already run open-source AI models. Seventy-six percent expect to use more.
Can enterprises control models without frontier scale?
Osei argues that open models now handle common enterprise work, including summarization, research, instruction following and document processing, well enough to make frontier benchmark gains less decisive. An Olympiad-level math result, for example, does not accelerate a company’s claims process.
Running open models such as Qwen on local hardware, Osei says, allows fine-tuning without sending a record outside. The author describes local execution as an unserious enterprise option two years ago and a viable choice today.
Companies can run a strong open-weight model inside their own walls or in a sovereign cloud, tune it with internal data and retain the weights, according to Osei. The aim is control over deployment rather than ownership of every hardware component.
In August, Nvidia agreed to pay Poolside $6 billion to license its model-building platform. Nvidia also offered jobs to more than 100 Poolside engineers, linking the transaction to development of its Nemotron family of open-weight models.
Osei uses the transaction to argue that supplier investment will keep improving open models. Confidential inference provides another layer: a trusted execution environment inside a CPU or GPU keeps data encrypted until it reaches a walled-off section of the chip that even the cloud provider cannot see.
According to Osei, this architecture can lower cost per workload, avoid surprise price increases and prevent a model from being deprecated, while giving boards a precise account of where data goes.
Where does sovereign AI break down?
The Center for a New American Security’s Sovereign AI Index finds about 90% of frontier AI compute in the United States and China. McKinsey, according to Osei, counts only about 30 countries with compute capable of advanced AI workloads.
Owning a stack requires GPUs, power, cooling and staff who can operate them, Osei says. Budgeting only for chips leaves the rest of that infrastructure unfunded.
CNAS tracks 184 government-backed sovereign AI projects. More than three-fifths disclose at least one foreign partner, and most of those partners are U.S. companies, Osei notes.
The World Economic Forum and Bain describe that reality as “strategic interdependence.” Osei agrees, recommending that companies decide which layers they must control and which they can safely rent.
Open models also lag on the hardest tasks, with McKinsey warning that localized models can fall short of frontier performance. Osei’s compromise is a split architecture: sensitive, high-volume work runs on models the company owns, while a governed path to frontier APIs handles workloads that require the strongest systems.
Why do enterprise programs fail?
Operational sovereignty is the part Osei says enterprises miss most: who can access the system, change its configuration, audit its actions and shut it down. The stakes rise when agents act on a company’s behalf. If an outside vendor holds the keys, Osei argues, the result is a hosted service rather than sovereign AI.
Model portability is the second requirement. Enterprises should be able to swap models without rebuilding surrounding systems, because a sovereign stack tied to one vendor merely replaces one dependency with another.
Osei treats sovereignty as an engineering specification, not a procurement checkbox. Each layer needs an owner and a metric: where inference occurs, where prompts and outputs are logged, how long those logs remain, which model sees which data, and who can turn the system off.
That discipline matches Osei’s experience at Marlabs, where the model is rarely the main problem; implementation is. Humans, architecture and governance determine whether the system works, and sovereign AI raises the cost of getting those elements wrong.
What will prove the model works?
The affected Team-plan users did nothing wrong beyond trusting their vendor agreement, Osei writes. A Manhattan courtroom they never entered decided their data fate.
For boards, regulators and major customers now asking who controls enterprise data, Osei says a credible sovereign AI program must identify where data resides and who holds access. The forward test is whether companies can meet those demands without recreating vendor dependence through hardware, skills or model lock-in.
Source: unite.ai
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,503 articles