Paying for an AI service doesn’t mean you own what happens to your work once it’s inside.
A live dispute making news right now shows exactly what that gap can cost someone: a researcher’s own paid AI usage may have helped a rival effort catch up to his breakthrough before he could publish it. Owning your AI infrastructure outright is the only way to make that risk structurally impossible instead of just unlikely.
What happened when a paying customer’s breakthrough ran through someone else’s AI?
NYU mathematics professor Tristan Buckmaster was closing in on proofs toward one of math’s seven Millennium Prize Problems, working with Anthropic mathematician Levent Alpöge and using AI models including OpenAI’s Codex, a service he was a paying customer of.
Buckmaster says a parallel OpenAI research effort, chasing a similar approach almost nobody else in the field was pursuing, caught up right as his own results were about to go public.
Buckmaster’s own account, reported by TechCrunch, describes an OpenAI researcher pushing him to drop Alpöge’s credit because Alpöge works at a rival lab, and warning him, “Why would you ruin your career?” OpenAI disputes that characterization and says it never accessed the pair’s unpublished work through any channel.
Days later, OpenAI announced its own full proof. Nobody outside OpenAI has independently verified it yet, and neither account of what happened to Buckmaster’s underlying usage can be settled from the outside.
That unresolved question is the whole problem.
Why doesn’t a stronger contract or a stricter policy fix this?
Because the issue isn’t bad intent. It’s structure.
When your most valuable work runs on infrastructure owned by a company that also builds its own competing products, there’s no boundary a policy can enforce that you can independently verify from the outside.
That’s true whether the work is:
-
- A math proof
- A product design
- A pricing model
- Or a document workflow your organization spent years refining
If someone else’s hardware sits between you and your own work, you’re trusting them to police a line you can’t see.
A stronger terms-of-service agreement doesn’t remove that trust requirement. It just writes it down.
What actually removes the risk: owning the infrastructure underneath your AI
The only way to make this kind of dispute structurally impossible, not just less likely, is to own the hardware your AI runs on.
GPU servers, clusters, storage, and networking sized for your own workload mean there’s no vendor positioned to see your work and simultaneously compete with it. There’s no boundary to trust, because there’s no third party standing at the boundary.
This is a different claim than “own your model.” A model you technically own but run on someone else’s rented, shared infrastructure still puts a vendor in the middle of everything that model touches.
Sovereignty starts at the hardware layer, not above it.
Why do most organizations get AI infrastructure sizing wrong?
Once an organization decides to own its infrastructure instead of renting access, sizing it correctly is the next place things go wrong, in two directions.
Overbuilding. Capacity gets bought for a workload the organization doesn’t have yet, sized off a vendor’s standard package instead of its own document volume or growth curve. The capital sits idle.
Underbuilding. The minimum gets bought to start, real production volume shows up within a year, and the organization falls back to paying a cloud provider by usage to cover the gap, right back into the dynamic this whole piece is about.
Both failure modes trace to the same root cause: sizing off a spec sheet instead of the actual workload.
Where does this fit in your organization’s tech evolution?
Most organizations move through stages on the way to owning their AI outright. Digital Transformation turns paper and legacy systems into digital, AI-ready data. IT Modernization refreshes the infrastructure that has to carry that data reliably. AI Acceleration is where infrastructure gets sized for real production use instead of a pilot.
AI sovereignty, owning the full stack outright so no vendor ever sits between your organization and its own work, is the destination past that. It only holds up if the infrastructure underneath it was sized and owned correctly from the start.
How does BIS size AI infrastructure, and remove the middleman, for a real organization?
We spec infrastructure against the workload actually in front of us:
-
- Real volumes of content, data, documents, etc
- Real user count
- Real growth curve
We don’t hand over a GPU server off a catalog and call it done.
We are uniquely positioned to help you because we hold ourselves to the same standard we recommend.
Aristotle, our internal AI server, runs the software we write for Grooper today. Because we own and run that infrastructure ourselves, there’s no vendor positioned to see our own development work and compete with it. The exact boundary problem at the center of the OpenAI dispute doesn’t exist for us, because there’s no third party at the boundary.
That discipline extends to the software layer too. Grooper exposes its own capabilities as MCP (Model Context Protocol) tools, so any AI agent, not one vendor’s proprietary assistant, can operate against infrastructure an organization already owns.
We move our own data in and out of Microsoft Dataverse using tools we built ourselves, so data stays portable on the organization’s terms, not a platform’s release schedule.
What should your next step be?
Ask a direct question about every AI tool your organization pays for right now:
-
- Who else can see what runs through it?
- What proof do you actually have either way?
If the honest answer is “I’m trusting the vendor,” that’s not a compliance detail to sort out later. It’s the same gap Buckmaster’s dispute is making public right now, just before it becomes your organization’s story instead of a research lab’s.
Closing that gap starts with sizing infrastructure for what you actually need, not guessing from a catalog.
That’s the conversation we’ve been having with organizations moving from AI ambition to AI reality for years. Contact us today and let’s talk about your situation to become AI sovereign.
