In the last few weeks, the sovereignty conversation has moved from academic to urgent. Palantir’s Alex Karp went on CNBC to explain the logic behind his company’s partnership with NVIDIA and spent 20 minutes arguing that enterprises are handing frontier labs something far more valuable than tokens: the “alpha” of their business.
Days later, Microsoft’s Satya Nadella told the Wall Street Journal something similar from the other side of the table. He said he doesn’t want “a few models and companies doing all of the learning for the world,” and he’s built Microsoft’s Copilot strategy around cheaper, swappable models that run, as he put it, “hill-climbing inside of a machine you control.” And Accenture just published research showing that 61% of business and government leaders are now more likely to seek out sovereign technology solutions amid rising geopolitical and competitive pressures.
Everyone is suddenly talking about AI sovereignty. Almost all of that talk is about infrastructure: whose GPUs you run on, whose models you call, and whether your data ever leaves your infrastructure. That’s a real concern, but it’s not just about the data. Your business semantics is the real “alpha.”
Sovereignty Is Guarding the Wrong Door
Karp’s argument: if you let a frontier model train on how your business actually runs, you’re not just paying for tokens. You’re sharing your proprietary know-how with your competition. His question to enterprises is blunt: why would you hand over control of your compute, your models, your data stack, and your alpha, when you could own the means of production yourself?
Nadella is making a version of the same argument from the vendor side. He’s said the character of a company going forward will be defined by “the tacit knowledge that they contain,” a blend of human judgment and AI, and that keeping that intellectual property intact is what keeps a business from being commoditized. Karp wants enterprises to leverage the Palantir platform, powered by local models, so a frontier lab never sees their alpha. Nadella wants enterprises building on Microsoft’s stack to protect their alpha.
Accenture’s data backs up the underlying worry, but it also shows how narrowly enterprises are applying it. In their survey of nearly 2,000 leaders, 60% of organizations apply sovereignty oversight to their data, and 46% apply it to infrastructure. Only 22% extend that oversight to their AI models, and just 32% apply it to the applications built on top of them. Most of the industry has locked the front door and left the model layer wide open.
Data governance and infrastructure control matter. But the thing that actually differentiates one enterprise from another was never sitting in the raw tables to begin with.
Your Data Warehouse Was Never the Moat
What actually makes your company different from a competitor? Not the data warehouse technology. Not the customer records. Not the transaction log. It’s how you calculate revenue, compute customer lifetime value, set your pricing rules, and make your forecasting assumptions. All of that is accumulated business knowledge, built up over years of decisions about what matters and how to measure it.
The semantic layer is where that knowledge already lives. It’s the enterprise’s codified operating model. If you can’t articulate right now who owns your definition of “active customer,” you’ve already lost the fight you think you’re having.
The Layer Nobody’s Governing
The alternative to letting an LLM query raw enterprise data directly is to place governance, consistent metrics, authorized access, lineage, and security between the model and the business. The AI never has to infer what “profitable customer” or “active account” means from scanning millions of rows of raw data. It reasons against a definition that your business has already agreed on.
Sovereignty in this context isn’t about keeping AI from touching your data. It’s about controlling how AI understands it. That matters more than where the model happens to run.
Make Meaning Portable, Not the Model
We’re in the middle of benchmarking open source models against frontier models, and the gap is closing faster than the marketing from Anthropic or OpenAI would suggest. As open models improve and inference costs decline, switching between models becomes easier. The model becomes the commoditized, replaceable part.
Nadella described his own target architecture in almost exactly those terms: a spectrum of models, priced and capable differently, “hill-climbing inside of a machine you control.” Swap out the vendor framing, and that’s a description of a governed semantic layer with interchangeable models sitting underneath it.
Right now, most enterprises are asking which LLM to standardize on. Wrong question. Here’s the one that matters: how quickly can you move between models without rebuilding your business logic each time? A governed semantic layer makes that possible because the definitions and metrics stay put even when the model underneath them changes.
Context Is the New Scarcity
A lot of energy goes into optimizing token pricing, inference cost, and latency. Within a few years, those will become commodity decisions, just as compute pricing became one after the first wave of cloud adoption.
What doesn’t commoditize is trusted context: the reusable business logic and governance that make any model’s output usable inside a regulated or high-stakes process. As model intelligence gets cheap, context gets scarce.
This is what AtScale’s semantic layer engine actually does. It’s not documentation sitting next to your warehouse. It’s a live query engine that compiles your business definitions into optimized queries, on demand, against any source, for BI tools, AI agents, or a SQL client. Ask a model for churn rate, and it isn’t guessing from raw tables. It’s hitting a definition your business already locked in, computed the same way every time.
Alpha Was Never Just Data
Karp keeps returning to that word “alpha,” the thing enterprises are giving away for free. He’s right that it’s worth protecting, but he’s wrong about the fix. Alpha isn’t only proprietary data. It’s how an organization defines and measures success in the first place, the logic that turns raw numbers into a decision. That logic lives in the semantic layer.
Whoever Owns the Meaning Wins
Models will keep changing. Token prices will keep falling. The organizations that own and govern their own business logic will still have something worth having when the dust settles: AI systems that reason using their own accumulated knowledge, not someone else’s approximation of it.
Go check who at your company can change what “revenue” means without anyone else finding out. That’s your real sovereignty problem.
SHARE
How to Evaluate Context Platforms