I was struck reading two texts published days apart—struck because I agreed with the diagnosis, and just as surprised by where it was coming from.
On July 1st, Alex Karp published on Palantir's X account a nine-point manifesto arguing that companies handing their data and models to frontier labs are quietly transferring their own 'alpha' to someone else. On July 12th, Satya Nadella published an essay making close to the same argument, even citing Karp by name.
The interesting reading isn't that they suddenly agree. It's that neither is really responding to the other—both are describing the same movement from two different standpoints. Frontier AI labs climbing the value chain: no longer selling a model, increasingly occupying the interaction layer of knowledge work itself. Claude Code, Claude Science, Codex, ChatGPT Enterprise: less a series of launches than one continuous climb. Less disruption than predation, and less predation than annexation: a new layer of the software stack absorbing the ones around it. It isn't new, either—reminiscent of how Microsoft won the spreadsheet wars by the mid-1990s, not by beating Lotus on features, but by owning the layer above it.
Nadella calls it the 'Reverse Information Paradox', after economist Kenneth Arrow's classic problem: a seller of information can't prove its value without giving it away. AI, he argues, inverts that—now it's the buyer who leaks. To get real value from a model, a company feeds it documents, workflows, judgment calls, and corrections—all of it what he calls "exhaust."
"You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful."
— Satya Nadella, The Reverse Information Paradox
Taken on its own, the diagnosis holds up. But the real question is: what follows from it, in practice?
Nadella's prescription is to own your evals, build your learning environment inside your own boundary, keep orchestration decoupled from any model, so value compounds toward you. He calls this step 'Compound'. But dependency compounds by the same logic: every quarter spent inside a comfortable, tightly integrated stack accrues sovereignty debt, compounding on the same curve his argument describes for alpha. And neither recommendation, as it happens, points away from their home—less a way out of dependency than a different address for it.
To me, sovereignty isn't something a vendor sells you—it's rather the room they leave you to compose independently of them. For a handful of genuinely open players, that's the entire pitch: value that grows because you can leave, not despite it. Not an idealistic bet: for a decade, it's been my engineers' smart choices, assembling those stacks, that built our most durable partnerships and grew our greatest capital—skills, and the judgment to know what to build with it.
