Different angles, same battlefield: trust, safety, and control over information.
Over the weekend, Satya Nadella dropped what might be the most important AI essay of the year, and it’s not about model quality or benchmarks. It’s about the reverse information paradox:
the idea that in the AI era, buyers now have to give away their most valuable knowledge just to use the product.
In the classic Arrow “information paradox,” the seller risks giving away the knowledge they’re trying to sell. In Nadella’s version, AI just flips this -
you can’t get value from a model without feeding it your prompts, your workflows, your evals, your internal tools, and your corrections. That “exhaust” is the stuff that actually makes your company different – the playbooks your competitor could never buy, even with a blank cheque.
The uncomfortable truth is that today, most enterprises are paying for AI twice.
Once with money and
once with their crown‑jewel know‑how.
Nadella’s answer is clear:
put a hard boundary around that exhaust.
Own your own AI memory.
Keep your orchestration layer loose enough that you can swap models in and out without sacrificing the institutional knowledge you’ve built up on top.
In other words, the moat should live with you, not with whatever lab happens to have the hottest model this quarter.
Of course, this isn’t a purely philosophical move. It’s also a positioning move.
While Nadella talks about “AI” in the abstract, the subtext is quite obvious: the default “send everything to the frontier lab via an API and trust the terms of service” model is no longer acceptable for serious enterprises.
And, somehow very conveniently –
the proposed cure looks a lot like Microsoft’s own stack?
tenant‑bound copilots,
governed agents, and
an orchestration layer (Azure AI Foundry) that sits between you and the labs.
Zoom out a bit, and you can see how neatly this slots into the week’s other storylines.
Apple is suing OpenAI, accusing it of stealing trade secrets and using ex‑employees and candidate interviews to siphon off sensitive hardware and product knowledge.
Whether or not the claims stand up in court, the narrative is devastating: the cool AI partner is suddenly painted as the firm that can’t be trusted with your secrets.
Anthropic, for its part, has spent years telling the world that advanced AI is dangerous and that it is the lab that takes risk seriously.
The message is simple:
others are racing recklessly;
we are the grown‑ups.
Even as their own policies evolve under competitive pressure, that “we’re the safe ones” branding is their core differentiator.
Put all of this together and a pattern emerges:
Apple is fighting over who controls product and hardware secrets.
Anthropic is fighting over who can be trusted not to blow up the world.
Nadella is fighting over who gets to own the knowledge exhaust that makes your company valuable.
Different angles, same battlefield.
For AI‑native companies and infra operators, this isn’t just gossip about big tech. It’s a warning about architecture.
If you centralize everything – prompts, traces, evals, tool calls inside someone else’s black box, you’re effectively letting them slowly reconstruct the one thing you can’t afford to give away: your domain intuition.
The alternative is to treat AI like you’d treat a critical piece of industrial equipment.
You don’t just plug your rail network or your port operations into a random vendor’s opaque system and hope for the best. You design around clear boundaries:
what runs where,
who sees what,
which traces stay in your environment, and
how easily you can rip and replace a component without losing your institutional memory.
That’s the deeper conversation Nadella’s essay opens up, and it’s why the timing matters.
As enterprises move from pilots to real deployment, we’re about to lock in the default pattern for “how AI is consumed.” The big players are all scrambling to claim the mantle of “trusted.” But the real question isn’t “who is safest?”
it’s “who ends up owning the knowledge you can’t afford to lose?”
In my last piece, “The Great AI Heist,” I argued that the real game isn’t just about training data; it’s about everything that happens after deployment –
the traces,
the playbooks,
the corrections,
the silent human expertise that accumulates around the model.
The events of this week just made that thesis painfully visible.
The next phase of AI isn’t just a race for better models.
It’s a race to decide where trust, safety, and control over information actually live – with the labs, with the platforms, or with you.
Latest on the Intelligent Founder AI -
