Our data. Our knowledge. Our control.

September 19, 2026 · OSASI Briefs

The case for shared AI infrastructure starts with privacy: keeping confidential information and professional knowledge under the control of those entrusted with it.

What prompted this brief

The closing section of The AI Daily Brief’s September 18 episode raises a question relevant to any profession handling sensitive information: how much control should an organisation retain over the AI it uses? Its written companion discusses enterprise interest in privately controlled models and concerns about dependence on outside providers. Episode notes.

Latham & Watkins offers a concrete example. The firm has invested in Nvidia servers and is experimenting with open-weight models in a controlled data centre. Bloomberg Law’s reporting.

For particularly sensitive client information, Latham’s chief information officer, Rene Mendoza, explained: “we don’t want to put it to any cloud vendor.” His comments to the Financial Times are reproduced by Legal IT Insider.

Cost and continued access to AI services are also part of the discussion. Jaya Gupta writes: “Pharma and banks are already picking up open weight models partly for margins, partly because a revocable lab API is a dependency they increasingly don’t want.” The concern is both financial and practical: dependence on AI access that an outside provider can withdraw. Jaya Gupta, September 16, 2026.

Mendoza also told Bloomberg Law: “Nobody can predict where any of this is going, so we have the best optionality.” That supports keeping options open as the market develops. Bloomberg Law.

Privacy and professional independence

For OSASI, the issue is protecting confidential information and the knowledge practitioners bring to their work. Client material, working notes, methods and accumulated experience should not become resources for someone else’s AI development simply because a practitioner uses a tool.

That is a principle for the system we would want to build, not an allegation that every provider trains on customer data. The questions need specific answers: what leaves the practice, who can access it, how long is it retained, and what other uses are permitted?

Shared infrastructure could give smaller practices a way to fund and govern a protected environment together. Shared operation should not mean a shared pool of client files. Each practice would need control over its material, with clear boundaries between users and no automatic contribution of their work to a common training dataset.

The proposed rules should be explicit: no scraping or harvesting of private work, no training on submissions, and no reuse of professional knowledge to develop another party’s products without specific authorisation from those entitled to give it. Those rules would need to cover the operator and any outside services it uses.

Owning servers would only be one part of that arrangement. Access controls, retention and deletion practices, security maintenance, and the ability to check compliance with the rules would matter just as much. A cooperative proposal would need to demonstrate those protections and a sustainable operating model.

There may also be value in choosing reliable reference material and excluding irrelevant or poor-quality sources. That is a separate design choice; it need not depend on collecting practitioners’ confidential work, and it would not by itself establish privacy or security.

The purpose is to preserve professional independence: access to useful technology while keeping our individual members in control of the information entrusted to them. The sources referenced in this post highlight organisations getting ahead of these issues. There is also value in preserving options as the cost, availability and terms of access to AI services evolve. OSASI prefers a proactive rather than reactive approach.

Listening reference

The AI Daily Brief — The AI Challenges Businesses Are Actually Focused On Right Now, September 18, 2026. Suggested listening: approximately 25:30 to the end, chapter “The Rise of Open Source and Owned AI Models in Enterprise.” The supplied screenshot shows 26:02; the start time is the contributor’s estimate.